WHITE PAPER FUJITSU PRIMERGY SERVERS PERFORMANCE REPORT PRIMERGY RX350 S7

Transcription

1 WHITE PAPER PERFORMANCE REPORT PRIMERGY RX350 S7 WHITE PAPER FUJITSU PRIMERGY SERVERS PERFORMANCE REPORT PRIMERGY RX350 S7 This document contains a summary of the benchmarks executed for the PRIMERGY RX350 S7. The PRIMERGY RX350 S7 performance data are compared with the data of other PRIMERGY models and discussed. In addition to the benchmark results, an explanation has been included for each benchmark and for the benchmark environment. Version Fujitsu Technology Solutions Page 1 (54)

2 Contents Document history... 3 Technical data... 4 SPECcpu SPECjbb SPECpower_ssj Disk I/O OLTP vservcon VMmark V STREAM LINPACK Literature Contact Page 2 (54) Fujitsu Technology Solutions

3 Document history Version 1.0 New: Technical data SPECcpu2006 Measurements with processors of Xeon series E SPECjbb2005 Measurement with Xeon E OLTP-2 Results for Xeon E processor series vservcon Results for Xeon E processor series VMmark V2 Measurement with Xeon E STREAM Measurements with Xeon E processor series LINPACK Measurements with Xeon E processor series Version 1.0a Minor corrections Version 1.1 New: SPECpower_ssj2008 Measurement with Oracle Java HotSpot VM Version 1.2 New: Disk I/O Measurements with LSI SW RAID on Intel C600 (Onboard SATA), LSI SW RAID on Intel C600 (Onboard SAS), RAID Ctrl SAS 6G 0/1, RAID Ctrl SAS 5/6 512MB (D2616) and RAID Ctrl SAS 6G 5/6 1GB (D3116) controllers Updated: SPECpower_ssj2008 Measurement with IBM J9 VM LINPACK Measurements with 2 Xeon E and up to 2 PY NVIDIA Tesla C2075 GPGPU Version 1.3 Updated: Technical data GPGPUs/coprocessors added LINPACK Measurements with 2 Xeon E and up to 2 PY NVIDIA Tesla K20 GPGPU Measurements with 2 Xeon E and up to 2 PY NVIDIA Tesla K20X GPGPU Version 1.4 Updated: LINPACK Measurements with 2 Xeon E and up to 2 Intel Xeon Phi Co-Processor 5110P Fujitsu Technology Solutions Page 3 (54)

4 Technical data PRIMERGY RX350 S7 LFF base unit (3.5" HDDs) PRIMERGY RX350 S7 SFF base unit (2.5" HDDs) Decimal prefixes according to the SI standard are used for measurement units in this white paper (e.g. 1 GB = 10 9 bytes). In contrast, these prefixes should be interpreted as binary prefixes (e.g. 1 GB = 2 30 bytes) for the capacities of caches and storage modules. Separate reference will be made to any further exceptions where applicable. Model Model versions Form factor Chipset Number of sockets 2 Number of processors orderable 1 or 2 PRIMERGY RX350 S7 LFF base unit: 1) 4-port SAS backplane for 4 3.5" hot-plug SAS or SATA HDDs including cabling for the connection to a modular 8-port SAS controller SFF base unit: 2) 8-port SAS backplane for 8 2.5" hot-plug SAS or SATA HDDs or SSDs including cabling for the connection to a modular 8-port SAS controller Rack server Intel C600 series Processor type Intel Xeon series E Number of memory slots Maximum memory configuration Onboard LAN controller Onboard HDD controller PCI slots Max. number of internal hard disks 24 (12 per processor) 768 GB 2 1 Gbit/s Controller with RAID 0, RAID 1 or RAID 10 for up to 4 3.5" SATA HDDs, optional: SAS Enabling Key for Onboard Ports for up to 4 3.5" SAS HDDs 2 PCI-Express 3.0 x4 (mech. x8) 4 PCI-Express 3.0 x8 (1 reserved for RAID card) 1 PCI-Express 3.0 x8 (mech. x16) 2 PCI-Express 3.0 x16 1 PCI-Express 2.0 x4 (mech. x8) LFF base unit: 12 SFF base unit: 24 1) LFF = Large Form Factor 2) SFF = Small Form Factor Page 4 (54) Fujitsu Technology Solutions

5 Cores Threads WHITE PAPER PERFORMANCE REPORT PRIMERGY RX350 S7 VERSION: The processor frequency specified in the following table is always at least achieved given full utilization. Processors with Turbo Boost Technology 2.0 additionally permit automatically regulated, dynamic overclocking. The overclocking rate depends on the utilization of the processor and its ambient conditions. As far as utilization is concerned, the number of cores subject to utilization as well as the type and strength of core utilization play a role. Added to these as influencing factors are the strength of the heating, the level of the ambient temperature and the heat dissipation options. As a result of overclocking it is even possible to exceed the thermal design power of the processor for short periods of time. How much a processor benefits from the Turbo mode in an individual case depends on the respective application and can in some application scenarios even differ from processor example to processor example. Processors (since system release) Processor Cache [MB] QPI Speed [GT/s] Processor Frequency [Ghz] Max. Turbo Frequency at full load [Ghz] Max. Turbo Frequency [Ghz] Max. Memory Frequency [MHz] TDP [Watt] Xeon E Xeon E n/a n/a Xeon E n/a n/a Xeon E Xeon E5-2630L Xeon E Xeon E Xeon E Xeon E Xeon E5-2650L Xeon E Xeon E Xeon E Xeon E Xeon E Xeon E Fujitsu Technology Solutions Page 5 (54)

6 Capacity [GB] Ranks Bit width of the memory chips Frequency [MHz] Low voltage Load reduced Registered ECC WHITE PAPER PERFORMANCE REPORT PRIMERGY RX350 S7 VERSION: Memory modules (since system release) Memory module 2GB (1x2GB) 1Rx8 L DDR U ECC (2 GB 1Rx8 PC3L-12800E) 4GB (1x4GB) 2Rx8 L DDR U ECC (4 GB 2Rx8 PC3L-12800E) 4GB (1x4GB) 1Rx4 L DDR R ECC (4 GB 1Rx4 PC3L-10600R) 4GB (1x4GB) 1Rx4 L DDR R ECC (4 GB 1Rx4 PC3L-12800R) 4GB (1x4GB) 2Rx8 L DDR R ECC (4 GB 2Rx8 PC3L-12800R) 8GB (1x8GB) 2Rx4 L DDR R ECC (8 GB 2Rx4 PC3L-10600R) 8GB (1x8GB) 2Rx4 L DDR R ECC (8 GB 2Rx4 PC3L-12800R) 16GB (1x16GB) 4Rx4 L DDR LR ECC (16 GB 4Rx4 PC3L-10600L) 16GB (1x16GB) 2Rx4 L DDR R ECC (16 GB 2Rx4 PC3L-12800R) 32GB (1x32GB) 4Rx4 L DDR LR ECC (32 GB 4Rx4 PC3L-10600L) GPGPUs/coprocessors (since system release) GPGPU/coprocessor Cores Peak double precision floating point performance [GFlops] Max. number of GPGPUs PY NVIDIA Tesla C2075 GPGPU PY NVIDIA Tesla K20 GPGPU PY NVIDIA Tesla K20X GPGPU Intel Xeon Phi Co-Processor 5110P Power supplies (since system release) Max. number Power supply 450W (hot-plug) 4 Power supply 800W (hot-plug) 4 Some components may not be available in all countries or sales regions. Detailed technical information is available in the data sheet PRIMERGY RX350 S7. Page 6 (54) Fujitsu Technology Solutions

7 SPECcpu2006 Benchmark description SPECcpu2006 is a benchmark which measures the system efficiency with integer and floating-point operations. It consists of an integer test suite (SPECint2006) containing 12 applications and a floating-point test suite (SPECfp2006) containing 17 applications. Both test suites are extremely computing-intensive and concentrate on the CPU and the memory. Other components, such as Disk I/O and network, are not measured by this benchmark. SPECcpu2006 is not tied to a special operating system. The benchmark is available as source code and is compiled before the actual measurement. The used compiler version and their optimization settings also affect the measurement result. SPECcpu2006 contains two different performance measurement methods: the first method (SPECint2006 or SPECfp2006) determines the time which is required to process single task. The second method (SPECint_rate2006 or SPECfp_rate2006) determines the throughput, i.e. the number of tasks that can be handled in parallel. Both methods are also divided into two measurement runs, base and peak which differ in the use of compiler optimization. When publishing the results the base values are always used; the peak values are optional. Benchmark Arithmetics Type Compiler optimization SPECint2006 integer peak aggressive SPECint_base2006 integer base conservative SPECint_rate2006 integer peak aggressive SPECint_rate_base2006 integer base conservative SPECfp2006 floating point peak aggressive SPECfp_base2006 floating point base conservative SPECfp_rate2006 floating point peak aggressive SPECfp_rate_base2006 floating point base conservative Measurement result Speed Throughput Speed Throughput Application single-threaded multi-threaded single-threaded multi-threaded The measurement results are the geometric average from normalized ratio values which have been determined for individual benchmarks. The geometric average - in contrast to the arithmetic average - means that there is a weighting in favour of the lower individual results. Normalized means that the measurement is how fast is the test system compared to a reference system. Value 1 was defined for the SPECint_base2006-, SPECint_rate_base2006, SPECfp_base2006 and SPECfp_rate_base2006 results of the reference system. For example, a SPECint_base2006 value of 2 means that the measuring system has handled this benchmark twice as fast as the reference system. A SPECfp_rate_base2006 value of 4 means that the measuring system has handled this benchmark some 4/[# base copies] times faster than the reference system. # base copies specify how many parallel instances of the benchmark have been executed. Not every SPECcpu2006 measurement is submitted by us for publication at SPEC. This is why the SPEC web pages do not have every result. As we archive the log files for all measurements, we can prove the correct implementation of the measurements at any time. Fujitsu Technology Solutions Page 7 (54)

8 Benchmark environment System Under Test (SUT) Hardware Model Processor Memory Power Supply Unit Software BIOS settings PRIMERGY RX350 S7 Xeon E processor series 1 processor: 8 8GB (1x8GB) 2Rx4 L DDR R ECC 2 processors: 16 8GB (1x8GB) 2Rx4 L DDR R ECC 2 Power supply 450W (hot-plug) SPECint_base2006, SPECint2006, SPECfp_base2006, SPECfp2006: Processors other than Xeon E5-2603, E5-2609: Hyper-Threading = Disabled Operating system Red Hat Enterprise Linux Server release 6.2 Operating system settings echo always > /sys/kernel/mm/redhat_transparent_hugepage/enabled Compiler Intel C++/Fortran Compiler 12.1 Some components may not be available in all countries or sales regions. Page 8 (54) Fujitsu Technology Solutions

9 Number of processors SPECint_base2006 SPECint2006 Number of processors SPECint_rate_base2006 SPECint_rate2006 Number of processors SPECint_rate_base2006 SPECint_rate2006 WHITE PAPER PERFORMANCE REPORT PRIMERGY RX350 S7 VERSION: Benchmark results In terms of processors the benchmark result depends primarily on the size of the processor cache, the support for Hyper-Threading, the number of processor cores and on the processor frequency. In the case of processors with Turbo mode the number of cores, which are loaded by the benchmark, determines the maximum processor frequency that can be achieved. In the case of single-threaded benchmarks, which largely load one core only, the maximum processor frequency that can be achieved is higher than with multithreaded benchmarks (see the processor table in the section "Technical Data"). Processor Xeon E Xeon E Xeon E Xeon E Xeon E5-2630L Xeon E Xeon E Xeon E Xeon E Xeon E5-2650L Xeon E Xeon E Xeon E Xeon E Xeon E Xeon E Fujitsu Technology Solutions Page 9 (54)

10 Number of processors SPECfp_base2006 SPECfp2006 Number of processors SPECfp_rate_base2006 SPECfp_rate2006 Number of processors SPECfp_rate_base2006 SPECfp_rate2006 WHITE PAPER PERFORMANCE REPORT PRIMERGY RX350 S7 VERSION: Processor Xeon E Xeon E Xeon E Xeon E Xeon E5-2630L Xeon E Xeon E Xeon E Xeon E Xeon E5-2650L Xeon E Xeon E Xeon E Xeon E Xeon E Xeon E Page 10 (54) Fujitsu Technology Solutions

11 The following four diagrams illustrate the throughput of the PRIMERGY RX350 S7 in comparison to its predecessor PRIMERGY TX300 S6, in their respective most performant configuration. SPECcpu2006: integer performance PRIMERGY RX350 S7 vs. PRIMERGY TX300 S SPECint SPECint_base PRIMERGY TX300 S6 2 x Xeon X5687 PRIMERGY RX350 S7 2 x Xeon E SPECcpu2006: integer performance PRIMERGY RX350 S7 vs. PRIMERGY TX300 S SPECint_rate SPECint_rate_base PRIMERGY TX300 S6 2 x Xeon X5690 PRIMERGY RX350 S7 2 x Xeon E Fujitsu Technology Solutions Page 11 (54)

12 SPECcpu2006: floating-point performance PRIMERGY RX350 S7 vs. PRIMERGY TX300 S SPECfp2006 SPECfp_base PRIMERGY TX300 S6 2 x Xeon X5687 PRIMERGY RX350 S7 2 x Xeon E SPECcpu2006: floating-point performance PRIMERGY RX350 S7 vs. PRIMERGY TX300 S SPECfp_rate SPECfp_rate_base PRIMERGY TX300 S6 2 x Xeon X5690 PRIMERGY RX350 S7 2 x Xeon E Page 12 (54) Fujitsu Technology Solutions

13 The two diagrams below reflect how the performance of the PRIMERGY RX350 S7 scales from one to two processors when using the Xeon E SPECcpu2006: integer performance PRIMERGY RX350 S7 (2 sockets vs. 1 socket) SPECint_rate SPECint_rate_base x Xeon E x Xeon E SPECcpu2006: floating-point performance PRIMERGY RX350 S7 (2 sockets vs. 1 socket) SPECfp_rate SPECfp_rate_base x Xeon E x Xeon E Fujitsu Technology Solutions Page 13 (54)

14 SPECjbb2005 Benchmark description SPECjbb2005 is a Java business benchmark that focuses on the performance of Java Server platforms. SPECjbb2005 is essentially a modernized SPECjbb2000. The main differences are: The transactions have become more complex in order to cover a greater functional scope. The working set of the benchmark has been enlarged to the extent that the total system load has increased. SPECjbb2000 allows only one active Java Virtual Machine instance (JVM) whereas SPECjbb2005 permits several instances, which in turn achieves greater closeness to reality, particularly with large systems. On the software side SPECjbb2005 primarily measures the performance of the JVM used with its just-in-time compiler as well as their thread and garbage collection implementation. Some aspects of the operating system used also play a role. As far as hardware is concerned, it measures the efficiency of the CPUs and caches, the memory subsystem and the scalability of shared memory systems (SMP). Disk and network I/O are irrelevant. SPECjbb2005 emulates a 3-tier client/server system that is typical for modern business process applications with the emphasis on the middle-tier system: Clients generate the load, consisting of driver threads, which on the basis of TPC-C benchmark generate OLTP accesses to a database without thinking times. The middle tier system implements the business processes and the updating of the database. The database takes on the data management and is emulated by Java objects that are in the memory. Transaction logging is implemented on an XML basis. The major advantage of this benchmark is that it includes all three tiers that run together on a single host. The performance of the middle-tier is measured. Large-scale hardware installations are thus avoided and direct comparisons between the SPECjbb2005 results from the various systems are possible. Client and database emulation are also written in Java. SPECjbb2005 only needs the operating system as well as a Java Virtual Machine with J2SE 5.0 features. The scaling unit is a warehouse with approx. 25 MB Java objects. Precisely one Java thread per warehouse executes the operations on these objects. The business operations are assumed by TPC-C: New Order Entry Payment Order Status Inquiry Delivery Stock Level Supervision Customer Report However, these are the only features SPECjbb2005 and TPC-C have in common. The results of the two benchmarks are not comparable. SPECjbb2005 has 2 performance metrics: bops (business operations per second) is the overall rate of all business operations performed per second. bops/jvm is the ratio of the first metrics and the number of active JVM instances. In comparisons of various SPECjbb2005 results, both metrics must be specified. The following rules, according to which a compliant benchmark run has to be performed, are the basis for these three metrics: A compliant benchmark run consists of a sequence of measuring points with an increasing number of warehouses (and thus of threads) with the number in each case being increased by one warehouse. The run is started at one warehouse up through 2*MaxWh, but not less than 8 warehouses. MaxWh is the number of warehouses with the highest rate per second the benchmark expects. Per default the benchmark equates MaxWh with the number of CPUs visible by the operating system. The metric bops is the arithmetic average of all measured operation rates with MaxWh warehouses up to 2*MaxWh warehouses. Page 14 (54) Fujitsu Technology Solutions

15 Benchmark environment System Under Test (SUT) Hardware Model Power Supply Unit PRIMERGY RX350 S7 2 Power supply 800W (hot-plug) Processor 2 Xeon E Memory Software BIOS settings Operating system Operating system settings JVM 16 8GB (1x8GB) 2Rx4 L DDR R ECC Hardware Prefetch = Disable Adjacent Sector Prefetch = Disable DCU Streamer Prefetch = Disable SAS/SATA OpROM = LSI MegaRAID Microsoft Windows Server 2008 R2 Enterprise SP1 Using the local security settings console, lock pages in memory was enabled for the user running the benchmark. Oracle Java HotSpot(TM) 64-Bit Server VM on Windows, version 1.6.0_31 JVM settings start /HIGH /AFFINITY [0xFFFF,0xFFFF0000] /B java -server -Xmx29g -Xms29g -Xmn24g - XX:BiasedLockingStartupDelay=200 -XX:ParallelGCThreads=16 -XX:SurvivorRatio=60 - XX:TargetSurvivorRatio=90 -XX:InlineSmallCode=3900 -XX:MaxInlineSize=270 - XX:FreqInlineSize=2500 -XX:AllocatePrefetchDistance=256 -XX:AllocatePrefetchLines=4 - XX:InitialTenuringThreshold=12 -XX:MaxTenuringThreshold=15 -XX:LoopUnrollLimit=45 - XX:+UseCompressedStrings -XX:+AggressiveOpts -XX:+UseLargePages - XX:+UseParallelOldGC -XX:-UseAdaptiveSizePolicy Some components may not be available in all countries or sales regions. Benchmark results SPECjbb2005 bops = SPECjbb2005 bops/jvm = The following diagrams illustrate the throughput of the PRIMERGY RX350 S7 in comparison to its predecessor PRIMERGY TX300 S6, in their respective most performant configuration. SPECjbb2005 bops: PRIMERGY RX350 S7 vs. TX300 S6 SPECjbb2005 bops: PRIMERGY RX350 S7 vs. TX300 S6 Fujitsu Technology Solutions Page 15 (54)

16 SPECpower_ssj2008 Benchmark description SPECpower_ssj2008 is the first industry-standard SPEC benchmark that evaluates the power and performance characteristics of a server. With SPECpower_ssj2008 SPEC has defined standards for server power measurements in the same way they have done for performance. The benchmark workload represents typical server-side Java business applications. The workload is scalable, multi-threaded, portable across a wide range of platforms and easy to run. The benchmark tests CPUs, caches, the memory hierarchy and scalability of symmetric multiprocessor systems (SMPs), as well as the implementation of Java Virtual Machine (JVM), Just In Time (JIT) compilers, garbage collection, threads and some aspects of the operating system. SPECpower_ssj2008 reports power consumption for servers at different performance levels from 100% to active idle in 10% segments over a set period of time. The graduated workload recognizes the fact that processing loads and power consumption on servers vary substantially over the course of days or weeks. To compute a power-performance metric across all levels, measured transaction throughputs for each segment are added together and then divided by the sum of the average power consumed for each segment. The result is a figure of merit called overall ssj_ops/watt. This ratio provides information about the energy efficiency of the measured server. The defined measurement standard enables customers to compare it with other configurations and servers measured with SPECpower_ssj2008. The diagram shows a typical graph of a SPECpower_ssj2008 result. The benchmark runs on a wide variety of operating systems and hardware architectures and does not require extensive client or storage infrastructure. The minimum equipment for SPEC-compliant testing is two networked computers, plus a power analyzer and a temperature sensor. One computer is the System Under Test (SUT) which runs one of the supported operating systems and the JVM. The JVM provides the environment required to run the SPECpower_ssj2008 workload which is implemented in Java. The other computer is a Control & Collection System (CCS) which controls the operation of the benchmark and captures the power, performance and temperature readings for reporting. The diagram provides an overview of the basic structure of the benchmark configuration and the various components. Page 16 (54) Fujitsu Technology Solutions

17 Benchmark environment System Under Test (SUT) Hardware Model Model version PRIMERGY RX350 S7 SFF base unit Processor 2 Xeon E Memory Network-Interface Disk-Subsystem Power Supply Unit Software BIOS BIOS settings 6 4GB (1x4GB) 2Rx8 L DDR U ECC Onboard LAN-Controller (1 port used) Onboard HDD-Controller Measurement with Oracle Java HotSpot VM: 1 SSD SATA 3G 32GB SLC HOT PLUG 2.5" EP Measurement with IBM J9 VM: 1 HD SATA 6G 250GB 7.2K HOT PL 2.5" BC 1 Power supply 450W (hot-plug) Measurement with Oracle Java HotSpot VM: R1.9.0 Measurement with IBM J9 VM: R Adjacent Sector Prefetch = Disabled Hardware Prefetch = Disabled DCU Streamer Prefetch = Disabled Memory Speed = Low-Voltage optimized USB Port Control = Enable internal ports only QPI Link Speed = 6.4GT/s P-State coordination = SW_ANY Intel Virtualization Technology = Disabled SAS/SATA OpROM = LSI MegaRAID ASPM Support = Auto LAN Controller = LAN 1 Firmware Measurement with Oracle Java HotSpot VM: 6.52A Measurement with IBM J9 VM: 6.53A Operating system Operating system settings JVM JVM settings Microsoft Windows Server 2008 R2 Enterprise SP1 Using the local security settings console, lock pages in memory was enabled for the user running the benchmark. Power Management: Enabled ( Fujitsu Enhanced Power Settings power plan) Set Turn off hard disk after = 1 Minute in OS. Benchmark was started via Windows Remote Desktop Connection. Measurement with Oracle Java HotSpot VM: Oracle Java HotSpot(TM) 64-Bit Server VM on Windows, version 1.6.0_30 Measurement with IBM J9 VM: IBM J9 VM (build 2.6, JRE Windows Server 2008 R2 amd _ (JIT enabled, AOT enabled) start /NODE [0,1] /AFFINITY [0x3,0xC,0x30,0xC0,0x300,0xC00,0x3000,0xC000] Measurement with Oracle Java HotSpot VM: -server -Xmx1024m -Xms1024m -Xmn853m -XX:ParallelGCThreads=2 -XX:SurvivorRatio=60 -XX:TargetSurvivorRatio=90 -XX:InlineSmallCode=3900 -XX:MaxInlineSize=270 -XX:FreqInlineSize=2500 -XX:AllocatePrefetchDistance=256 -XX:AllocatePrefetchLines=4 -XX:InitialTenuringThreshold=12 -XX:MaxTenuringThreshold=15 -XX:LoopUnrollLimit=45 -XX:+UseCompressedStrings -XX:+AggressiveOpts -XX:+UseLargePages -XX:+UseParallelOldGC Measurement with IBM J9 VM: -Xaggressive -Xcompressedrefs -Xgcpolicy:gencon -Xmn800m -Xms1024m -Xmx1024m -XlockReservation -Xnoloa -XtlhPrefetch -Xlp -Xconcurrentlevel0 Fujitsu Technology Solutions Page 17 (54)

18 Other software Measurement with Oracle Java HotSpot VM: none Measurement with IBM J9 VM: IBM SDK Java Technology Edition Version 7.0 for Windows x64 Some components may not be available in all countries or sales regions. Benchmark results Measurement with Oracle Java HotSpot VM The PRIMERGY RX350 S7 achieved the following result: SPECpower_ssj2008 = 5,035 overall ssj_ops/watt The adjoining diagram shows the result of the configuration described above. The red horizontal bars show the performance to power ratio in ssj_ops/watt (upper x-axis) for each target load level tagged on the y-axis of the diagram. The blue line shows the run of the curve for the average power consumption (bottom x-axis) at each target load level marked with a small rhomb. The black vertical line shows the benchmark result of 5,035 overall ssj_ops/watt for the PRIMERGY RX350 S7. This is the quotient of the sum of the transaction throughputs for each load level and the sum of the average power consumed for each measurement interval. The following table shows the benchmark results for the throughput in ssj_ops, the power consumption in watts and the resulting energy efficiency for each load level. Performance Power Energy Efficiency Target Load ssj_ops Average Power (W) ssj_ops/watt 100% 1,306, ,436 90% 1,178, ,576 80% 1,046, ,825 70% 914, ,112 60% 785, ,087 50% 653, ,755 40% 524, ,101 30% 391, ,216 20% 262, ,139 10% 131, ,777 Active Idle ssj_ops / power = 5,035 Page 18 (54) Fujitsu Technology Solutions

19 Measurement with IBM J9 VM The PRIMERGY RX350 S7 achieved the following result: SPECpower_ssj2008 = 5,347 overall ssj_ops/watt The adjoining diagram shows the result of the configuration described above. The red horizontal bars show the performance to power ratio in ssj_ops/watt (upper x-axis) for each target load level tagged on the y-axis of the diagram. The blue line shows the run of the curve for the average power consumption (bottom x-axis) at each target load level marked with a small rhomb. The black vertical line shows the benchmark result of 5,347 overall ssj_ops/watt for the PRIMERGY RX350 S7. This is the quotient of the sum of the transaction throughputs for each load level and the sum of the average power consumed for each measurement interval. The following table shows the benchmark results for the throughput in ssj_ops, the power consumption in watts and the resulting energy efficiency for each load level. Performance Power Energy Efficiency Target Load ssj_ops Average Power (W) ssj_ops/watt 100% 1,432, ,838 90% 1,286, ,966 80% 1,147, ,208 70% 1,003, ,452 60% 855, ,429 50% 714, ,081 40% 571, ,378 30% 432, ,484 20% 286, ,311 10% 144, ,890 Active Idle ssj_ops / power = 5,347 The PRIMERGY RX350 S7 achieved a new class record with this result (date: September 19, 2012). Thus, the PRIMERGY RX350 S7 proves itself to be the most energy-efficient 2-socket 4U rack server in the world. The current results can be found at Fujitsu Technology Solutions Page 19 (54)

20 The following diagram shows for each load level the power consumption (on the right y-axis) and the throughput (on the left y-axis) of the PRIMERGY RX350 S7 compared to the predecessor the PRIMERGY TX300 S6. SPECpower_ssj2008: PRIMERGY RX350 S7 vs. PRIMERGY TX300 S6 Thanks to the new Sandy Bridge microarchitecture and the 9% higher-performing IBM J9 VM the PRIMERGY RX350 S7 has in comparison with the PRIMERGY TX300 S6 a substantially higher throughput and considerably lower power consumption. Both result in an overall increase in energy efficiency in the PRIMERGY RX350 S7 of 79%. SPECpower_ssj2008 overall ssj_ops/watt: PRIMERGY RX350 S7 vs. PRIMERGY TX300 S6 Page 20 (54) Fujitsu Technology Solutions

21 Disk I/O Benchmark description Performance measurements of disk subsystems for PRIMERGY servers are used to assess their performance and enable a comparison of the different storage connections for PRIMERGY servers. As standard, these performance measurements are carried out with a defined measurement method, which models the hard disk accesses of real application scenarios on the basis of specifications. The essential specifications are: Share of random accesses / sequential accesses Share of read / write access types Block size (kb) Number of parallel accesses (# of outstanding I/Os) A given value combination of these specifications is known as load profile. The following five standard load profiles can be allocated to typical application scenarios: Standard load profile Access Type of access Block size read write [kb] Application File copy random 50% 50% 64 Copying of files File server random 67% 33% 64 File server Database random 67% 33% 8 Streaming sequential 100% 0% 64 Database (data transfer) Mail server Database (log file), Data backup; Video streaming (partial) Restore sequential 0% 100% 64 Restoring of files In order to model applications that access in parallel with a different load intensity, the # of Outstanding I/Os is increased, starting with 1, 3, 8 and going up to 512 (from 8 onwards in increments to the power of two). The measurements of this document are based on these standard load profiles. The main results of a measurement are: Throughput [MB/s] Throughput in megabytes per second Transactions [IO/s] Transaction rate in I/O operations per second Latency [ms] Average response time in ms The data throughput has established itself as the normal measurement variable for sequential load profiles, whereas the measurement variable transaction rate is mostly used for random load profiles with their small block sizes. Data throughput and transaction rate are directly proportional to each other and can be transferred to each other according to the formula Data throughput [MB/s] Transaction rate [IO/s] = Transaction rate [IO/s] Block size [MB] = Data throughput [MB/s] / Block size [MB] This section specifies hard disk capacities on a basis of 10 (1 TB = bytes) while all other capacities, file sizes, block sizes and throughputs are specified on a basis of 2 (1 MB/s = 2 20 bytes/s). All the details of the measurement method and the basics of disk I/O performance are described in the white paper Basics of Disk I/O Performance. Fujitsu Technology Solutions Page 21 (54)

22 Benchmark environment All the measurement results discussed in this chapter were determined using the hardware and software components listed below: System Under Test (SUT) Hardware Controller Drive Software Operating system Administration software Initialization of RAID arrays File system 1 LSI SW RAID on Intel C600 (Onboard SATA) 1 LSI SW RAID on Intel C600 (Onboard SAS) 1 RAID Ctrl SAS 6G 0/1 (D2607) 1 RAID Ctrl SAS 5/6 512MB (D2616) 1 RAID Ctrl SAS 6G 5/6 1GB (D3116) 24 EP HDD SAS 6 Gbit/s rpm 146 GB 12 EP HDD SAS 6 Gbit/s rpm 300 GB 24 EP SSD SAS 6 Gbit/s GB MLC 4 BC HDD SATA 6 Gbit/s rpm 3 TB Microsoft Windows Server 2008 Enterprise x64 Edition SP2 ServerView RAID Manager RAID arrays are initialized before the measurement with an elementary block size of 64 kb ( stripe size ) NTFS Measuring tool Iometer Measurement data Measurement files of 32 GB with 1 8 hard disks; 64 GB with 9 16 hard disks; 128 GB with 17 or more hard disks Some components may not be available in all countries / sales regions. Page 22 (54) Fujitsu Technology Solutions

23 Benchmark results The results presented here are designed to help you choose the right solution from the various configuration options of the PRIMERGY RX350 S7 in the light of disk-i/o performance. The selection of suitable components and the right settings of their parameters is important here. These two aspects should therefore be dealt with as preparation for the discussion of the performance values. Components The hard disks are the first essential component. If there is a reference below to hard disks, this is meant as the generic term for HDDs ( hard disk drives, in other words conventional hard disks) and SSDs ( solid state drives, i.e. non-volatile electronic storage media). When selecting the type of hard disk and number of hard disks you can move the weighting in the direction of storage capacity, performance, security or price. In order to enable a pre-selection of the hard disk types depending on the required weighting the hard disk types for PRIMERGY servers are divided into three classes: Economic (ECO): low-priced hard disks Business Critical (BC): very failsafe hard disks Enterprise (EP): very failsafe and very high-performance hard disks The following table is a list of the hard disk types that have been available for the PRIMERGY RX350 S7 since system release. Drive class Data medium type Interface Form factor krpm Business Critical HDD SATA 6G 2.5" 7.2 Business Critical HDD SATA 6G 3.5" 7.2 Enterprise HDD SAS 6G 3.5" 15 Enterprise HDD SAS 6G 2.5" 10, 15 Enterprise SSD SATA 6G 2.5" - Enterprise SSD SAS 6G 2.5" - Mixed drive configurations of SAS and SATA hard disks in one system are permitted, unless they are excluded in the configurator for special hard disk types. The SATA-HDDs offer high capacities right up into the terabyte range at a very low cost. The SAS-HDDs have shorter access times and achieve higher throughputs due to the higher rotational speed of the SAS- HDDs (in comparison with the SATA-HDDs). SAS-HDDs with a rotational speed of 15 krpm have better access times and throughputs than comparable HDDs with a rotational speed of 10 krpm. The 6G interface has in the meantime established itself as the standard among the SAS-HDDs. Of all the hard disk types SSDs offer on the one hand by far the highest transaction rates for random load profiles, and on the other hand the shortest access times. In return, however, the price per gigabyte of storage capacity is substantially higher. More hard disks per system are possible as a result of using 2.5" hard disks instead of 3.5" hard disks. Consequently, the load that each individual hard disk has to overcome decreases and the maximum overall performance of the system increases. More detailed performance statements about hard disk types are available in the white paper Single Disk Performance. Fujitsu Technology Solutions Page 23 (54)

24 The maximum number of hard disks in the system depends on the system configuration. The following table lists the essential cases. Form factor Interface Connection type Number of PCIe controllers Maximum number of hard disks 3.5" SATA 3G, SAS 3G direct ", 3.5" SATA 3G/6G, SAS 6G direct " SATA 6G, SAS 6G Expander " SATA 3G/6G, SAS 6G direct " SATA 3G/6G, SAS 6G Expander 1 24 After the hard disks the RAID controller is the second performance-determining key component. In the case of these controllers the modular RAID concept of the PRIMERGY servers offers a plethora of options to meet the various requirements of a wide range of different application scenarios. The following table summarizes the most important features of the available RAID controllers of the PRIMERGY RX350 S7. A short alias is specified here for each controller, which is used in the subsequent list of the performance values. Controller name Alias Cache Supported interfaces LSI SW RAID on Intel C600 (Onboard SATA) LSI SW RAID on Intel C600 (Onboard SAS) RAID Ctrl SAS 6G 0/1 (D2607) RAID Ctrl SAS 6G 5/6 512MB (D2616) RAID Ctrl SAS 6G 5/6 1GB (D3116) Max. # disks in the system RAID levels in the system Patsburg A - SATA 3G " 0, 1, 10 -/- Patsburg B - SATA 3G SAS 3G LSI SATA 3G/6G SAS 3G/6G LSI MB SATA 3G/6G SAS 3G/6G LSI2208-1G 1 GB SATA 3G/6G SAS 3G/6G BBU/ FBU " 0, 1, 10 -/- PCIe 2.0 x8 PCIe 2.0 x8 PCIe 2.0 x " 8 3.5" " " " " 0, 1, 1E, 10 -/- 0, 1, 5, 6, 10, 50, 60 0, 1, 1E, 5, 6, 10, 50, 60 /- -/ The onboard RAID controller is implemented in the chipset Intel C600 on the motherboard of the server and uses the CPU of the server for the RAID functionality. This controller is a simple solution that does not require a PCIe slot. In addition to the invariably available connection option of SATA hard disks, the additional SAS functionality can be activated via an SAS enabling key. System-specific interfaces The interfaces of a controller to the motherboard and to the hard disks have in each case specific limits for data throughput. These limits are listed in the following table. The minimum of these two values is a definite limit, which cannot be exceeded. This value is highlighted in bold in the following table. Controller alias Effective in the configuration # Disk channels Limit for throughput of disk interface PCIe version PCIe width Limit for throughput of PCIe interface Connection via expander Patsburg A 4 SATA 3G 973 MB/s Patsburg B 4 SAS 3G 973 MB/s LSI SAS 6G 3890 MB/s 2.0 x MB/s - LSI SAS 6G 3890 MB/s 2.0 x MB/s LSI2208-1G 8 SAS 6G 3890 MB/s 2.0 x MB/s Page 24 (54) Fujitsu Technology Solutions

25 An expander makes it possible to connect more hard disks in a system than the SAS channels that the controller has. An expander cannot increase the possible maximum throughput of a controller, but makes it available in total to all connected hard disks. More details about the RAID controllers of the PRIMERGY systems are available in the white paper RAID Controller Performance. Settings In most cases, the cache of the hard disks has a great influence on disk-i/o performance. This is particular valid for HDDs. It is frequently regarded as a security problem in case of power failure and is thus switched off. On the other hand, it was integrated by hard disk manufacturers for the good reason of increasing the write performance. For performance reasons it is therefore advisable to enable the hard disk cache. This is particular valid for SATA-HDDs. The performance can as a result increase more than tenfold for specific access patterns and hard disk types. More information about the performance impact of the hard disk cache is available in the document Single Disk Performance. To prevent data loss in case of power failure you are recommended to equip the system with a UPS. In the case of controllers with a cache there are several parameters that can be set. The optimal settings can depend on the RAID level, the application scenario and the type of data medium. In the case of RAID levels 5 and 6 in particular (and the more complex RAID level combinations 50 and 60) it is obligatory to enable the controller cache for application scenarios with write share. If the controller cache is enabled, the data temporarily stored in the cache should be safeguarded against loss in case of power failure. Suitable accessories are available for this purpose (e.g. a BBU or FBU). For the purpose of easy and reliable handling of the settings for RAID controllers and hard disks it is advisable to use the RAID-Manager software ServerView RAID that is supplied for PRIMERGY servers. All the cache settings for controllers and hard disks can usually be made en bloc specifically for the application by using the pre-defined modi Performance or Data Protection. The Performance mode ensures the best possible performance settings for the majority of the application scenarios. More information about the setting options of the controller cache is available in the white paper RAID Controller Performance. Performance values In general, disk-i/o performance of a RAID array depends on the type and number of hard disks, on the RAID level and on the RAID controller. If the limits of the system-specific interfaces are not exceeded, the statements on disk-i/o performance are therefore valid for all PRIMERGY systems. This is why all the performance statements of the document RAID Controller Performance also apply for the PRIMERGY RX350 S7 if the configurations measured there are also supported by this system. The performance values of the PRIMERGY RX350 S7 are listed in table form below, specifically for different RAID levels, access types and block sizes. Substantially different configuration versions are dealt with separately. The performance values in the following tables use the established measurement variables, as already mentioned in the subsection Benchmark description. Thus, transaction rate is specified for random accesses and data throughput for sequential accesses. To avoid any confusion among the measurement units the tables have been separated for the two access types. The table cells contain the maximum achievable values. This has three implications: On the one hand hard disks with optimal performance were used (the components used are described in more detail in the subsection Benchmark environment). Furthermore, cache settings of controllers and hard disks, which are optimal for the respective access scenario and the RAID level, are used as a basis. And ultimately each value is the maximum value for the entire load intensity range (# of outstanding I/Os). In order to also visualize the numerical values each table cell is highlighted with a horizontal bar, the length of which is proportional to the numerical value in the table cell. All bars shown in the same scale of length have the same color. In other words, a visual comparison only makes sense for table cells with the same colored bars. Since the horizontal bars in the table cells depict the maximum achievable performance values, they are shown by the color getting lighter as you move from left to right. The light shade of color at the right end of the bar tells you that the value is a maximum value and can only be achieved under optimal prerequisites. The darker the shade becomes as you move to the left, the more frequently it will be possible to achieve the corresponding value in practice. Fujitsu Technology Solutions Page 25 (54)

26 RAID Controller Hard disk type Form factor #Disks RAID level HDDs random 8 kb blocks 67% read [IO/s] HDDs random 64 kb blocks 67% read [IO/s] SSDs random 8 kb blocks 67% read [IO/s] SSDs random 64 kb blocks 67% read [IO/s] WHITE PAPER PERFORMANCE REPORT PRIMERGY RX350 S7 VERSION: Random accesses (performance values in IO/s): Configuration version Patsburg A BC SATA HDD 3.5" Patsburg B EP SAS HDD 3.5" LSI2008 EP SAS HDD EP SAS SSD 2.5" LSI2008 EP SAS HDD 3.5" LSI2108 LSI2108 EP SAS HDD 3.5" LSI2208-1G EP SAS HDD EP SAS SSD EP SAS HDD EP SAS SSD 2.5" 2.5" LSI2208-1G EP SAS HDD 3.5" N/A N/A N/A N/A N/A N/A N/A N/A N/A N/A N/A N/A N/A N/A N/A N/A N/A N/A N/A N/A N/A N/A N/A N/A N/A N/A N/A N/A N/A N/A N/A N/A N/A N/A Page 26 (54) Fujitsu Technology Solutions

27 RAID Controller Hard disk type Form factor #Disks RAID level HDDs sequential 64 kb blocks 100% read [MB/s] HDDs sequential 64 kb blocks 100% write [MB/s] SSDs sequential 64 kb blocks 100% read [MB/s] SSDs sequential 64 kb blocks 100% write [MB/s] WHITE PAPER PERFORMANCE REPORT PRIMERGY RX350 S7 VERSION: Sequential accesses (performance values in MB/s): Configuration version Patsburg A BC SATA HDD 3.5" Patsburg B EP SAS HDD 3.5" LSI2008 EP SAS HDD EP SAS SSD 2.5" LSI2008 EP SAS HDD 3.5" LSI2108 EP SAS HDD EP SAS SSD 2.5" LSI2108 EP SAS HDD 3.5" LSI2208-1G EP SAS HDD EP SAS SSD 2.5" LSI2208-1G EP SAS HDD 3.5" N/A N/A N/A N/A N/A N/A N/A N/A N/A N/A N/A N/A N/A N/A N/A N/A N/A N/A N/A N/A N/A N/A N/A N/A N/A N/A N/A N/A N/A N/A N/A N/A N/A N/A The use of one controller at its maximum configuration with powerful hard disks (configured as RAID 0) enables the PRIMERGY RX350 S7 to achieve a throughput of up to 2958 MB/s for sequential load profiles and a transaction rate of up to IO/s for typical, random application scenarios. Fujitsu Technology Solutions Page 27 (54)

28 OLTP-2 Benchmark description OLTP stands for Online Transaction Processing. The OLTP-2 benchmark is based on the typical application scenario of a database solution. In OLTP-2 database access is simulated and the number of transactions achieved per second (tps) determined as the unit of measurement for the system. In contrast to benchmarks such as SPECint and TPC-E, which were standardized by independent bodies and for which adherence to the respective rules and regulations are monitored, OLTP-2 is an internal benchmark of Fujitsu. OLTP-2 is based on the well-known database benchmark TPC-E. OLTP-2 was designed in such a way that a wide range of configurations can be measured to present the scaling of a system with regard to the CPU and memory configuration. Even if the two benchmarks OLTP-2 and TPC-E simulate similar application scenarios using the same load profiles, the results cannot be compared or even treated as equal, as the two benchmarks use different methods to simulate user load. OLTP-2 values are typically similar to TPC-E values. A direct comparison, or even referring to the OLTP-2 result as TPC-E, is not permitted, especially because there is no priceperformance calculation. Further information can be found in the document Benchmark Overview OLTP-2. Benchmark environment The measurement set-up is symbolically illustrated below: Driver Tier A Tier B Network Network Application Server Database Server Disk subsystem Clients System Under Test (SUT) All results were determined by way of example on a PRIMERGY RX300 S7. Page 28 (54) Fujitsu Technology Solutions

29 Database Server (Tier B) Hardware Model Processor PRIMERGY RX300 S7 Xeon E processor series Memory 1 processor: 8 32GB (1x32GB) 4Rx4 L DDR LR ECC 2 processors: 16 32GB (1x32GB) 4Rx4 L DDR LR ECC Network interface Disk subsystem Software BIOS Operating system Database 2 onboard LAN 1 Gb/s RX300 S7: Onboard RAID Ctrl SAS 6G 5/6 1024MB (D3116) 2 73 GB 15k rpm SAS Drive, RAID1 (OS), GB 15k rpm SAS Drive, RAID10 (LOG) 3 LSI MegaRAID SAS 9286CV-8e 6 JX40: GB SSD Drive each, RAID5 (data) Version V R1.0.5 Microsoft Windows Server 2008 R2 Enterprise SP1 Microsoft SQL Server 2008 R2 Enterprise SP1 Application Server (Tier A) Hardware Model Processor Memory Network interface Disk subsystem Software Operating system 1 PRIMERGY RX200 S6 2 Xeon X GB, 1333 MHz registered ECC DDR3 2 onboard LAN 1 Gb/s 2 Dual Port LAN 1Gb/s 1 73 GB 15k rpm SAS Drive Microsoft Windows Server 2008 R2 Standard Client Hardware Model 1 PRIMERGY RX200 S5 Processor 2 Xeon X5570 Memory 24 GB, 1333 MHz registered ECC DDR3 Network interface 2 onboard LAN 1 Gb/s Disk subsystem 1 73 GB 15k rpm SAS Drive Software Operating system Microsoft Windows Server 2008 R2 Standard Benchmark OLTP-2 Software EGen version Some components may not be available in all countries / sales regions. Fujitsu Technology Solutions Page 29 (54)

30 Benchmark results Database performance greatly depends on the configuration options with CPU, memory and on the connectivity of an adequate disk subsystem for the database. In the following scaling considerations for the processors we assume that both the memory and the disk subsystem has been adequately chosen and is not a bottleneck. A guideline in the database environment for selecting main memory is that sufficient quantity is more important than the speed of the memory accesses. This why a configuration with a total memory of 512 GB was considered for the measurements with two processors and a configuration with a total memory of 256 GB for the measurements with one processor. Both memory configurations have memory access of 1333 MHz. Further information about memory performance can be found in the White Paper Memory Performance of Xeon E (Sandy Bridge-EP) Based Systems. The following diagram shows the OLTP-2 transaction rates that can be achieved with one and two processors of the Intel Xeon E series. OLTP-2 tps E Core, HT E Core, HT E Core, HT E Core, HT E Core, HT E Core, HT E5-2650L - 8 Core, HT E Core, HT E Core, HT E Core, HT E5-2630L - 6 Core, HT E Core, HT E Core, HT E Core E Core E Core, HT HT: Hyper-Threading CPUs 512GB RAM CPU 256GB RAM tps bold: measured cursive: calculated Page 30 (54) Fujitsu Technology Solutions

31 It is evident that a wide performance range is covered by the variety of released processors. If you compare the OLTP-2 value of the processor with the lowest performance (Xeon E5-2603) with the value of the processor with the highest performance (Xeon E5-2690), the result is a 4-fold increase in performance. Based on the results achieved the processors can be divided into different performance groups: The start is made with Xeon E and E as processors with four cores, but without Hyper- Threading and without turbo mode. Although the Xeon E only has two cores, it is nevertheless Hyper- Threading-capable and on account of the clock frequency lies, as far as performance is concerned, between these two processors. Due to its high clock frequency and the high QPI speed of 8.00 GT/s the throughput rates of the 6-core processors with the lowest frequencies (Xeon E and E5-2630L) are almost achieved with the performance-optimized 4-core processor Xeon E However, the processors with 95 Watt and 60 Watt respectively also have distinctly lower power consumption than the Xeon E with 130 Watt. The 6-core processors are all Hyper-Threading-capable, have with 7.20 GT/s a higher QPI speed than the group of 4-core processors with 6.40 GT/s and they have a 50% larger L3 cache of 15 MB. At the upper end of the performance scale of the 6-core processors is the Xeon E (130 Watt) with its especially high frequency, which on the other hand achieves an OLTP performance that is slightly above the 8-core processor with the lowest performance, Xeon E5-2650L (70 Watt). The group of processors with eight cores, a QPI speed of 8.00 GT/s and a 20 MB L3 cache is to be found at the upper end of the performance scale. Due to the graduated CPU clock frequencies an OLTP performance of between 1145 tps (2 Xeon E5-2650L) and 1696 tps (2 Xeon E5-2690) is achieved. If you compare the maximum achievable OLTP-2 values of the current system generation with the values that were achieved on the predecessor systems, the result is an increase of about 34%. tps Maximum OLTP-2 tps Comparison of system generations + ~ 34% X GB Predecessor System 2 E GB Current System Current System TX300 S7 RX200 S7 RX300 S7 RX350 S7 BX924 S3 Predecessor System TX300 S6 RX200 S6 RX300 S6 TX300 S6 BX924 S2 Fujitsu Technology Solutions Page 31 (54)

32 vservcon Benchmark description vservcon is a benchmark used by Fujitsu Technology Solutions to compare server configurations with hypervisor with regard to their suitability for server consolidation. This allows both the comparison of systems, processors and I/O technologies as well as the comparison of hypervisors, virtualization forms and additional drivers for virtual machines. vservcon is not a new benchmark in the true sense of the word. It is more a framework that combines already established benchmarks (or in modified form) as workloads in order to reproduce the load of a consolidated and virtualized server environment. Three proven benchmarks are used which cover the application scenarios database, application server and web server. Application scenario Benchmark No. of logical CPU cores Memory Database Sysbench (adapted) GB Java application server SPECjbb (adapted, with 50% - 60% load) 2 2 GB Web server WebBench GB Each of the three application scenarios is allocated to a dedicated virtual machine (VM). Add to these a fourth machine, the so-called idle VM. These four VMs make up a tile. Depending on the performance capability of the underlying server hardware, you may as part of a measurement also have to start several identical tiles in parallel in order to achieve a maximum performance score. System Under Test Database VM Java VM Web VM Idle VM Tile n Database VM Database VM Database VM Java VM Java VM Java VM Web VM Web VM Web VM Idle VM Idle VM Idle VM Tile 3 Tile 2 Tile 1 Each of the three vservcon application scenarios provides a specific benchmark result in the form of application-specific transaction rates for the respective VM. In order to derive a normalized score, the individual benchmark results for one tile are put in relation to the respective results of a reference system. The resulting relative performance values are then suitably weighted and finally added up for all VMs and tiles. The outcome is a score for this tile number. Starting as a rule with one tile, this procedure is performed for an increasing number of tiles until no further significant increase in this vservcon score occurs. The final vservcon score is then the maximum of the vservcon scores for all tile numbers. This score thus reflects the maximum total throughput that can be achieved by running the mix defined in vservcon that consists of numerous VMs up to the possible full utilization of CPU resources. This is why the measurement environment for vservcon measurements is designed in such a way that only the CPU is the limiting factor and that no limitations occur as a result of other resources. The progression of the vservcon scores for the tile numbers provides useful information about the scaling behavior of the System under Test. Moreover, vservcon also documents the total CPU load of the host (VMs and all other CPU activities) and, if possible, electrical power consumption. A detailed description of vservcon is in the document: Benchmark Overview vservcon. Page 32 (54) Fujitsu Technology Solutions

33 Benchmark environment The measurement set-up is symbolically illustrated below: Framework controller Server Disk subsystem Multiple 1Gb or 10Gb networks Load generators System Under Test (SUT) All results were determined by way of example on a PRIMERGY RX350 S7. System Under Test (SUT) Hardware Model Processor PRIMERGY RX350 S7 Xeon E processor series Memory 1 processor: 8 8GB (1x8GB) 2Rx4 L DDR R ECC 2 processors: 16 8GB (1x8GB) 2Rx4 L DDR R ECC Network interface Disk subsystem Software 1 dual port 1GbE adapter 1 dual port 10GbE server adapter 1 dual-channel FC controller Emulex LPe12002 ETERNUS DX80 storage systems: Each tile: 50 GB LUN Each LUN: RAID 0 with 2 Seagate ST SS disks (15 krpm) Operating system VMware ESX Build Load generator (incl. Framework controller) Hardware (Shared) Enclosure PRIMERGY BX900 Hardware Model 18 PRIMERGY BX920 S1 server blades Processor 2 Xeon X5570 Memory 12 GB Network interface 3 1 Gbit/s LAN Software Operating system Microsoft Windows Server 2003 R2 Enterprise with Hyper-V Fujitsu Technology Solutions Page 33 (54)

34 Load generator VM (per tile 3 load generator VMs on various server blades) Hardware Processor 1 logical CPU Memory 512 MB Network interface 2 1 Gbit/s LAN Software Operating system Microsoft Windows Server 2003 R2 Enterprise Edition Some components may not be available in all countries or sales regions. Page 34 (54) Fujitsu Technology Solutions

35 Xeon E Series RX200 S7 RX300 S7 RX350 S7 TX300 S7 BX924 S3 CX250 S1 CX270 S1 WHITE PAPER PERFORMANCE REPORT PRIMERGY RX350 S7 VERSION: Benchmark results The PRIMERGY dual-socket systems dealt with here are based on Intel Xeon series E processors. The features of the processors are summarized in the section Technical data. The available processors of these systems with their results can be seen in the following table. Processor #Tiles Score 2 Cores, HT, TM E Cores E E Cores, HT, TM E Cores HT, TM E E5-2630L E E E Cores HT, TM E5-2650L E E E E E E HT = Hyper-Threading, TM = Turbo Mode These PRIMERGY dual-socket systems are very suitable for application virtualization thanks to the progress made in processor technology. Compared with a system based on the previous processor generation an approximate 40% higher virtualization performance can be achieved (measured in vservcon score in their maximum configuration). The relatively large performance differences between the processors can be explained by their features. The values scale on the basis of the number of cores, the size of the L3 cache and the CPU clock frequency and as a result of the features of Hyper-Threading and turbo mode, which are available in most processor types. Furthermore, the data transfer rate between processors ( QPI Speed ) also determines performance. As a matter of principle, the memory access speed also influences performance. A guideline in the virtualization environment for selecting main memory is that sufficient quantity is more important than the speed of the memory accesses. More information about the topic Memory Performance and QPI architecture can be found in the White Paper Memory Performance of Xeon E (Sandy Bridge-EP) Based Systems. Fujitsu Technology Solutions Page 35 (54)

36 Final vservcon Score tiles tiles E E E E E E5-2630L E E E E5-2650L E E E E E E Final vservcon Score WHITE PAPER PERFORMANCE REPORT PRIMERGY RX350 S7 VERSION: The first diagram compares the virtualization performance values that can be achieved with the processors reviewed here. Xeon E Processor Series #Tiles Core 4 Core 6 Core 8 Core The Xeon E as the processor with two cores only makes the start. A similarly low performance can be seen in the Xeon E and E processors, as they have to manage without Hyper-Threading (HT) and turbo mode (TM). In principle, these weakest processors are only to a limited extent suitable for the virtualization environment. A further increase in performance is achieved by the processor with four cores, which supports both Hyper- Threading and the turbo mode (Xeon E5-2643). In addition to the number of cores, the L3 cache and the data transfer rate make a considerable contribution to the respective increase in performance in the 8-core versions compared with the 6-core versions. Within a group of processors with the same number of cores scaling can be seen via the CPU clock frequency Until now we have looked at the virtualization performance of a fully configured system. However, with a server with two sockets the question also arises as to how good performance scaling is from one to two processors. The better the scaling, the lower the overhead usually caused by the shared use of resources within a server. The scaling factor also depends on the application. If the server is used as a virtualization platform for server consolidation, the system scales with a factor of When operated with two processors, the system thus almost achieves twice the performance as with one processor, as is illustrated in the diagram opposite using the processor version Xeon E as an example. 0 1 x E x E Page 36 (54) Fujitsu Technology Solutions

37 vservcon score WHITE PAPER PERFORMANCE REPORT PRIMERGY RX350 S7 VERSION: The next diagram illustrates the virtualization performance for increasing numbers of VMs based on the Xeon E (6-Core) and E (8-Core) processors. The respective CPU loads of the host have also been entered. The number of tiles with optimal CPU load is typically at about 90%; beyond that you have overload, which is where virtualization performance no longer increases, or E E sinks again % In addition to the increased number of CPU Util % physical cores, Hyper-Threading, which is supported by almost all Xeon 10 processors of the E series, is 90% 80% an additional reason for the high 70% 8 number of VMs that can be operated. 60% As is known, a physical processor core is consequently divided into two logical cores so that the number of cores available for the hypervisor is 6 50% 40% doubled. This standard feature thus 4 generally increases the virtualization performance of a system. 2 30% 20% 10% % #Tiles The scaling curves for the number of tiles as seen in the previous diagram are specifically for systems with Hyper-Threading. 16 physical and thus 32 logical cores are available with the Xeon E processors; approximately four of them are used per tile (see Benchmark description). This means that a parallel use of the same physical cores by several VMs is avoided up to a maximum of about four tiles. That is why the performance curve in this range scales almost ideal. For the quantities above the growth is flatter up to CPU full utilization. The previous diagram examined the total performance of all application VMs of a host. However, studying the performance from an individual application VM viewpoint is also interesting. This information is in the previous diagram. For example, the total optimum is reached in the above Xeon E situation with 24 application VMs (eight tiles, not including the idle VMs); the low load case is represented by three application VMs (one tile, not including the idle VM). Remember: the vservcon score for one tile is an average value across the three application scenarios in vservcon. This average performance of one tile drops when changing from the low load case to the total optimum of the vservcon score - from 2.02 to 10.4/8=1.3, i.e. to 64%. The individual types of application VMs can react very differently in the high load situation. It is thus clear that in a specific situation the performance requirements of an individual application must be balanced against the overall requirements regarding the numbers of VMs on a virtualization host. Fujitsu Technology Solutions Page 37 (54)

38 vservcon Score WHITE PAPER PERFORMANCE REPORT PRIMERGY RX350 S7 VERSION: The virtualization-relevant progress in processor technology since 2008 has an effect on the one hand on an individual VM and, on the other hand, on the possible maximum number of VMs up to CPU full utilization. The following comparison shows the proportions for both types of improvements. Four systems are compared with approximately the same processor frequency: a system from 2008 with 2 Xeon E5420, a system from 2009 with 2 Xeon E5540, a system from 2011 with 2 Xeon E5649 and a current system with 2 Xeon E Virtualization relevant improvements Few VMs (1 Tile) Score at optimum Tile count E GHz 4C 2009 E GHz 4C 2011 E GHz 6C 2012 E GHz 8C 2008 E GHz 4C 2009 E GHz 4C 2011 E GHz 6C 2012 E GHz 8C Year CPU Freq. #Cores 2012 TX300 S7 RX200 S7 RX300 S7 RX350 S7 - - BX924 S3 CX250 S1 CX270 S TX300 S6 RX200 S6 RX300 S6 TX300 S6 BX620 S6 BX922 S2 BX924 S TX300 S5 RX200 S5 RX300 S5 - BX620 S TX300 S4 RX200 S4 RX300 S4 - BX620 S The clearest performance improvements arose from 2008 to 2009 with the introduction of the Xeon 5500 processor generation (e. g. via the feature Extended Page Tables (EPT) 1 ). One sees an increase of the vservcon score by a factor of 1.30 with a few VMs (one tile). With full utilization of the systems with VMs there was an increase by a factor of The one reason was the performance increase that could be achieved for an individual VM (see score for a few VMs). The other reason was that more VMs were possible with total optimum (via Hyper-Threading). However, it can be seen that the optimum was bought with a triple number of VMs with a reduced performance of the individual VM. Where exactly is the technology progress between 2009 and 2012? The performance for an individual VM in low-load situations has basically remained the same for the processors compared here with approximately the same clock frequency but with different cache size and speed of memory connection. The decisive progress is in the higher number of physical cores and associated with it in the increased values of pure performance (factor 1.47 and 1.64 in the diagram). We must explicitly point out that the increased virtualization performance as seen in the score cannot be completely deemed as an improvement for one individual VM. More than approximately 30% to 50% increased throughput compared to an identically clocked processor of the Xeon 5400 generation from 2008 is not possible here. Performance increases in the virtualization environment since 2009 are mainly achieved by increased VM numbers due to the increased number of available logical or physical cores. 1 EPT accelerates memory virtualization via hardware support for the mapping between host and guest memory addresses. Page 38 (54) Fujitsu Technology Solutions

39 VMmark V2 Benchmark description VMmark V2 is a benchmark developed by VMware to compare server configurations with hypervisor solutions from VMware regarding their suitability for server consolidation. In addition to the software for load generation, the benchmark consists of a defined load profile and binding regulations. The benchmark results can be submitted to VMware and are published on their Internet site after a successful review process. After the discontinuation of the proven benchmark VMmark V1 in October 2010, it has been succeeded by VMmark V2, which requires a cluster of at least two servers and covers data center functions, like Cloning and Deployment of virtual machines (VMs), Load Balancing, as well as the moving of VMs with vmotion and also Storage vmotion. VMmark V2 is not a new benchmark in the actual sense. It is in fact a framework that consolidates already established benchmarks, as workloads in order to simulate the load of a virtualized consolidated server environment. Three proven benchmarks, which cover the application scenarios mail server, Web 2.0, and e-commerce were integrated in VMmark V2. Application scenario Load tool # VMs Mail server LoadGen 1 Web 2.0 Olio client 2 E-commerce DVD Store 2 client 4 Standby server (IdleVMTest) 1 Each of the three application scenarios is assigned to a total of seven dedicated virtual machines. Then add to these an eighth VM called the standby server. These eight VMs form a tile. Because of the performance capability of the underlying server hardware, it is usually necessary to have started several identical tiles in parallel as part of a measurement in order to achieve a maximum overall performance. A new feature of VMmark V2 is an infrastructure component, which is present once for every two hosts. It measures the efficiency levels of data center consolidation through VM Cloning and Deployment, vmotion and Storage vmotion. The Load Balancing capacity of the data center is also used (DRS, Distributed Resource Scheduler). The result of VMmark V2 is a number, known as a score, which provides information about the performance of the measured virtualization solution. The score reflects the maximum total consolidation benefit of all VMs for a server configuration with hypervisor and is used as a comparison criterion of various hardware platforms. This score is determined from the individual results of the VMs and an infrastructure result. Each of the five VMmark V2 application or front-end VMs provides a specific benchmark result in the form of applicationspecific transaction rates for each VM. In order to derive a normalized score the individual benchmark results for one tile are put in relation to the respective results of a reference system. The resulting dimensionless performance values are then averaged geometrically and finally added up for all VMs. This value is included in the overall score with a weighting of 80%. The infrastructure workload is only present in the benchmark once for every two hosts; it determines 20% of the result. The number of transactions per hour and the average duration in seconds respectively are determined for the score of the infrastructure workload components. In addition to the actual score, the number of VMmark V2 tiles is always specified with each VMmark V2 score. The result is thus as follows: Score@Number of Tiles, for example 4.20@5 tiles. A detailed description of VMmark V2 is available in the document Benchmark Overview VMmark V2. Fujitsu Technology Solutions Page 39 (54)

40 Benchmark environment The measurement set-up is symbolically illustrated below: Clients & Management Server(s) Storage System Multiple 1Gb or 10Gb networks Load Generators incl. Prime Client and Datacenter Management Server vmotion network System under Test (SUT) System Under Test (SUT) Hardware Number of servers 2 Model PRIMERGY RX350 S7 Processor 2 Xeon E Memory Network interface Disk subsystem Software BIOS BIOS settings 256 GB: GB (1x16GB) 2Rx4 L DDR R ECC 1 dual port 1GbE adapter 1 dual port 10GbE server adapter 1 dual-channel FC controller Emulex LPe12002 ETERNUS DX80 storage systems: Each tile: 241 GB Each DX80: RAID 0 with several LUNs Total: 114 disks (incl. SSDs) Version V R1.0.6 See details Operating system VMware ESX U2 Build Operating system settings ESX settings: see details Page 40 (54) Fujitsu Technology Solutions

41 Prime Client/Datacenter Management Server (DMS) Hardware (Shared) Enclosure PRIMERGY BX600 Network Switch 1 PRIMERGY BX600 GbE Switch Blade 30/12 Hardware Model Processor Memory Network interface Software 1 server blade PRIMERGY BX620 S4 2 Xeon X GB 2 1 Gbit/s LAN Operating system Prime Client: Microsoft Windows Server 2003 R2 Enterprise Edition SP2, KB DMS: Microsoft Windows Server 2003 R2 Enterprise x64 Edition SP2, KB Load generator Hardware Model 1 PRIMERGY RX600 S6 Processor 4 Xeon E Memory Network interface Software 512 GB 1 1 Gbit/s LAN 2 10 Gbit/s LAN Operating system VMware ESX U2 Build Load generator VM (per tile 1 load generator VM) Hardware Processor Memory Network interface Software Operating system 4 logical CPU 4 GB 1 1 Gbit/s LAN Microsoft Windows Server 2008 Enterprise x64 Edition SP2 Details See disclosure Some components may not be available in all countries or sales regions. Fujitsu Technology Solutions Page 41 (54)

42 tiles tiles tiles tiles tiles tiles tiles tiles VMmark V2 Score WHITE PAPER PERFORMANCE REPORT PRIMERGY RX350 S7 VERSION: Benchmark results On March 6, 2012 Fujitsu achieved with a PRIMERGY RX350 S7 with Xeon E processors and VMware ESX U2 a VMmark V2 score of 10.52@10 tiles in a system configuration with a total of 2 16 processor cores and when using two identical servers in the System under Test (SUT). With this result the PRIMERGY RX350 S7 is in the official VMmark V2 ranking the most powerful 2-socket server in a matched pair configuration consisting of two identical hosts (valid as of benchmark results publication date). All comparisons for the competitor products reflect the status of 7th March The current VMmark V2 results as well as the detailed results and configuration data are available at The diagram shows the result of the PRIMERGY RX350 S7 in comparison with the best 2-socket systems. 2-socket systems Fujitsu PRIMERGY RX350 S7 2 Xeon E HP ProLiant DL380p G8 2 Xeon E HP ProLiant BL460c Gen8 2 Xeon E HP ProLiant BL620c G7 2 Xeon E HP ProLiant BL620c G7 2 Xeon E Fujitsu PRIMERGY RX300 S6 2 Xeon X IBM BladeCenter HS22V 2 Xeon X Dell PowerEdge R710 2 Xeon X5690 The table opposite shows the difference in the score (in %) between the Fujitsu system and comparable hardware. 2-socket systems VMmark V2 score Difference Fujitsu PRIMERGY RX350 S % HP ProLiant DL380p G The processors used, which with a good hypervisor setting could make optimal use of their processor features, were the essential prerequisites for achieving the PRIMERGY RX350 S7 result. These features include Hyper-Threading. All this has a particularly positive effect during virtualization. In comparison with a PRIMERGY system of the predecessor generation with Xeon X5690 processors an increase in performance of about 38% is achieved with VMmark V2. All VMs, their application data, the host operating system as well as additionally required data were on a powerful fibre channel disk subsystem from ETERNUS DX80 systems. As far as possible, the configuration of the disk subsystem takes the specific requirements of the benchmark into account. The use of SSDs (Solid State Disk) resulted in advantages in the number and response times of the hard disks used. The network connection of the load generators and the infrastructure workload connection between the hosts were implemented with the 10Gb LAN ports. All the components used were optimally attuned to each other. Page 42 (54) Fujitsu Technology Solutions

43 STREAM Benchmark description STREAM is a synthetic benchmark that has been used for many years to determine memory throughput and which was developed by John McCalpin during his professorship at the University of Delaware. Today STREAM is supported at the University of Virginia, where the source code can be downloaded in either Fortran or C. STREAM continues to play an important role in the HPC environment in particular. It is for example an integral part of the HPC Challenge benchmark suite. The benchmark is designed in such a way that it can be used both on PCs and on server systems. The unit of measurement of the benchmark is GB/s, i.e. the number of gigabytes that can be read and written per second. STREAM measures the memory throughput for sequential accesses. These can generally be performed more efficiently than accesses that are randomly distributed on the memory, because the CPU caches are used for sequential access. Before execution the source code is adapted to the environment to be measured. Therefore, the size of the data area must be at least four times larger than the total of all CPU caches so that these have as little influence as possible on the result. The OpenMP program library is used to enable selected parts of the program to be executed in parallel during the runtime of the benchmark, consequently achieving optimal load distribution to the available processor cores. During implementation the defined data area, consisting of 8-byte elements, is successively copied to four types, and arithmetic calculations are also performed to some extent. Type Execution Bytes per step Floating-point calculation per step COPY a(i) = b(i) 16 0 SCALE a(i) = q b(i) 16 1 SUM a(i) = b(i) + c(i) 24 1 TRIAD a(i) = b(i) + q c(i) 24 2 The throughput is output in GB/s for each type of calculation. The differences between the various values are usually only minor on modern systems. In general, only the determined TRIAD value is used as a comparison. The measured results primarily depend on the clock frequency of the memory modules; the CPUs influence the arithmetic calculations. The accuracy of the results is approximately 5%. This chapter specifies throughputs on a basis of 10 (1 GB/s = 10 9 Byte/s). Benchmark environment System Under Test (SUT) Hardware Model Processor Memory Software BIOS settings PRIMERGY RX350 S7 2 processors of Xeon E processor series 16 8GB (1x8GB) 2Rx4 L DDR R ECC Processors other than Xeon E5-2603, E5-2609: Hyper-Threading = Disabled Operating system Red Hat Enterprise Linux Server release 6.2 Operating system settings Compiler Intel C Compiler 12.1 Benchmark Stream.c Version 5.9 echo never > /sys/kernel/mm/redhat_transparent_hugepage/enabled Some components may not be available in all countries or sales regions. Fujitsu Technology Solutions Page 43 (54)

44 Benchmark results Processor Cores Processor Frequency [Ghz] Max. Memory Frequency [MHz] TRIAD [GB/s] 2 Xeon E Xeon E Xeon E Xeon E Xeon E5-2630L Xeon E Xeon E Xeon E Xeon E Xeon E5-2650L Xeon E Xeon E Xeon E Xeon E Xeon E Xeon E The results depend primarily on the maximum memory frequency. The Xeon E5-2637, which with only 2 cores does not use all 4 channels of the memory controller in the STREAM benchmark, is the exception. The smaller differences with processors with the same maximum memory frequency are a result in arithmetic calculation of the different processor frequencies. The following diagram illustrates the throughput of the PRIMERGY RX350 S7 in comparison to its predecessor, the PRIMERGY TX300 S6, in their most performant configuration. STREAM TRIAD: PRIMERGY RX350S7 vs. PRIMERGY TX300S6 GB/s PRIMERGY TX300 S6 2 Xeon X PRIMERGY RX350 S7 2 Xeon E Page 44 (54) Fujitsu Technology Solutions

45 LINPACK Benchmark description LINPACK was developed in the 1970s by Jack Dongarra and some other people to show the performance of supercomputers. The benchmark consists of a collection of library functions for the analysis and solution of linear system of equations. A description can be found in the document LINPACK can be used to measure the speed of computers when solving a linear equation system. For this purpose, an n n matrix is set up and filled with random numbers between -2 and +2. The calculation is then performed via LU decomposition with partial pivoting. A memory of 8n² bytes is required for the matrix. In case of an n n matrix the number of arithmetic operations required for the solution is 2 / 3 n 3 + 2n 2. Thus, the choice of n determines the duration of the measurement: a doubling of n results in an approximately eight-fold increase in the duration of the measurement. The size of n also has an influence on the measurement result itself: as n increases, the measured value asymptotically approaches a limit. The size of the matrix is therefore usually adapted to the amount of memory available. Furthermore, the memory bandwidth of the system only plays a minor role for the measurement result, but a role that cannot be fully ignored. The processor performance is the decisive factor for the measurement result. Since the algorithm used permits parallel processing, in particular the number of processors used and their processor cores are - in addition to the clock rate - of outstanding significance. LINPACK is used to measure how many floating point operations were carried out per second. The result is referred to as Rmax and specified in GFlops (Giga Floating Point Operations per Second). An upper limit, referred to as Rpeak, for the speed of a computer can be calculated from the maximum number of floating point operations that its processor cores could theoretically carry out in one clock cycle: Rpeak = Maximum number of floating point operations per clock cycle Number of processor cores of the computer Maximum processor frequency[ghz] LINPACK is classed as one of the leading benchmarks in the field of high performance computing (HPC). LINPACK is one of the seven benchmarks currently included in the HPC Challenge benchmark suite, which takes other performance aspects in the HPC environment into account. Manufacturer-independent publication of LINPACK results is possible at The use of a LINPACK version based on HPL is prerequisite for this (see: Intel offers a highly optimized LINPACK version (shared memory version) for individual systems with Intel processors. Parallel processes communicate here via "shared memory", i.e. jointly used memory. Another version provided by Intel is based on HPL (High Performance Linpack). Intercommunication of the LINPACK processes here takes place via OpenMP and MPI (Message Passing Interface). This enables communication between the parallel processes - also from one computer to another. Both versions can be downloaded from Manufacturer-specific LINPACK versions also come into play when graphics cards for General Purpose Computation on Graphics Processing Unit (GPGPU) are used. These are based on HPL and include extensions which are needed for communication with the graphics cards. Fujitsu Technology Solutions Page 45 (54)

46 Benchmark environment Measurements with Xeon E processor series System Under Test (SUT) Hardware Model Processor Memory Software BIOS settings PRIMERGY RX350 S7 2 processors of Xeon E processor series 16 8GB (1x8GB) 2Rx4 L DDR R ECC Processors other than Xeon E5-2603, E5-2609: Hyper-Threading = Disabled Operating system Red Hat Enterprise Linux Server release 6.2 Benchmark Shared memory version: Intel Optimized LINPACK Benchmark 10.3 Update 11 for Linux OS Measurements with 2 Xeon E and up to 2 PY NVIDIA Tesla C2075 GPGPU System Under Test (SUT) Hardware Model PRIMERGY RX350 S7 Processor 2 Xeon E Memory GPGPU/Coprocessor Software BIOS-settings 16 8GB (1x8GB) 2Rx4 L DDR R ECC 1 / 2 PY NVIDIA Tesla C2075 GPGPU Hyper-Threading = Disabled Operating system Red Hat Enterprise Linux Server release 6.3 Benchmark Compiler HPL version: CUDA-enabled version of HPL optimized for Tesla 20-series GPUs version 1.3 Intel MPI Library 4.0 Update 3 for Linux OS Intel Math Kernel Library 10.3 Update 11 for Linux OS CUDA 4.0 Intel C++ Compiler XE 12.1 Update 5 for Linux Page 46 (54) Fujitsu Technology Solutions

47 Measurements with 2 Xeon E and up to 2 PY NVIDIA Tesla K20 GPGPU Measurements with 2 Xeon E and up to 2 PY NVIDIA Tesla K20X GPGPU System Under Test (SUT) Hardware Model PRIMERGY RX350 S7 Processor 2 Xeon E Memory GPGPU/Coprocessor Software BIOS-settings 16 8GB (1x8GB) 2Rx4 L DDR R ECC 1 / 2 PY NVIDIA Tesla K20 GPGPU 1 / 2 PY NVIDIA Tesla K20X GPGPU Hyper-Threading = Disabled Turbo Mode = Enabled (default) = Disabled Operating system Red Hat Enterprise Linux Server release 6.3 Benchmark Compiler HPL version: CUDA-enabled version of HPL optimized for Tesla 20-series GPUs version 1.5 Intel MPI Library 4.0 Update 3 for Linux OS Intel Math Kernel Library 11.0 Update 2 for Linux OS CUDA 5.0 Intel C++ Compiler XE 13.1 for Linux Measurements with 2 Xeon E and up to 2 Intel Xeon Phi Co-Processor 5110P System Under Test (SUT) Hardware Model PRIMERGY RX350 S7 Processor 2 Xeon E Memory GPGPU/Coprocessor Software BIOS-settings 16 8GB (1x8GB) 2Rx4 L DDR R ECC 1 / 2 Intel Xeon Phi Co-Processor 5110P Hyper-Threading = Disabled Turbo Mode = Enabled (default) = Disabled Operating system Red Hat Enterprise Linux Server release 6.3 Benchmark Compiler HPL version: Intel Optimized LINPACK Benchmark 11.0 Update 3 for Linux OS Intel MPI Library for Linux OS Intel Math Kernel Library 11.0 Update 3 for Linux OS Intel C++ Compiler XE 13.1 for Linux Some components may not be available in all countries or sales regions. Fujitsu Technology Solutions Page 47 (54)

48 Processor Cores Processor frequency [Ghz] Maximum turbo frequency at full load [Ghz] Number of processors WHITE PAPER PERFORMANCE REPORT PRIMERGY RX350 S7 VERSION: Benchmark results Measurements with Xeon E processor series Without Turbo Mode With Turbo Mode Rpeak [GFlops] Rmax [GFlops] Rpeak [GFlops] Rmax [GFlops] Xeon E Xeon E n/a Xeon E n/a Xeon E Xeon E5-2630L Xeon E Xeon E Xeon E Xeon E Xeon E5-2650L Xeon E Xeon E Xeon E Xeon E Xeon E Xeon E Rmax = Measurement result Rpeak = Maximum number of floating point operations per clock cycle Number of processor cores of the computer Maximum processor frequency[ghz] The following applies for processors without Turbo mode and for those with Turbo mode disabled: Maximum processor frequency[ghz] = Nominal processor frequency[ghz] Processors with Turbo mode enabled are not limited by the nominal processor frequency and therefore do not provide a constant processor frequency. Instead the actual processor frequency swings - depending on temperature and power consumption - between the nominal processor frequency and maximum turbo frequency at full load. Therefore, the following applies for these processors: Maximum processor frequency[ghz] = Maximum turbo frequency at full load[ghz] Page 48 (54) Fujitsu Technology Solutions

49 Processor Cores Processor frequency [Ghz] Maximum turbo frequency at full load [Ghz] Number of processors GPGPU/Coprocessor Number of GPGPUs/Coprocessors WHITE PAPER PERFORMANCE REPORT PRIMERGY RX350 S7 VERSION: Measurements with 2 Xeon E and up to 2 PY NVIDIA Tesla C2075 GPGPU Theoretical maximum performance of a single graphics card according to the data sheet: PY NVIDIA Tesla C2075 GPGPU: 515 GFlops During runtime the computer load was distributed over the system processors and the processors of the graphics cards by means of a given ratio. The LINPACK result is this made up of the sum of the performance values of the system processors and the graphics cards. Without Turbo Mode With Turbo Mode Rpeak [GFlops] Rmax [GFlops] Rpeak [GFlops] Rmax [GFlops] Xeon E Xeon E PY NVIDIA Tesla C2075 PY NVIDIA Tesla C LINPACK: PRIMERGY RX350 S7 2 Xeon E PY NVIDIA Tesla C Xeon E PY NVIDIA Tesla C % +88% 2 Xeon E GFlops Fujitsu Technology Solutions Page 49 (54)

50 Processor Cores Processor frequency [Ghz] Maximum turbo frequency at full load [Ghz] Number of processors GPGPU/Coprocessor Number of GPGPUs/Coprocessors WHITE PAPER PERFORMANCE REPORT PRIMERGY RX350 S7 VERSION: Measurements with 2 Xeon E and up to 2 PY NVIDIA Tesla K20 GPGPU Measurements with 2 Xeon E and up to 2 PY NVIDIA Tesla K20X GPGPU Theoretical maximum performance of a single graphics card according to the data sheet: PY NVIDIA Tesla K20 GPGPU: 1170 GFlops PY NVIDIA Tesla K20X GPGPU: 1310 GFlops During runtime the computer load was distributed over the system processors and the processors of the graphics cards by means of a given ratio. The LINPACK result is this made up of the sum of the performance values of the system processors and the graphics cards. Without Turbo Mode With Turbo Mode Rpeak [GFlops] Rmax [GFlops] Rpeak [GFlops] Rmax [GFlops] Xeon E Xeon E Xeon E Xeon E PY NVIDIA Tesla K20 PY NVIDIA Tesla K20 PY NVIDIA Tesla K20X PY NVIDIA Tesla K20X LINPACK: PRIMERGY RX350 S7 2 Xeon E PY NVIDIA Tesla K20X 2 Xeon E PY NVIDIA Tesla K Xeon E PY NVIDIA Tesla K20X 2 Xeon E PY NVIDIA Tesla K Xeon E GFlops Page 50 (54) Fujitsu Technology Solutions

51 Processor Cores Processor frequency [Ghz] Maximum turbo frequency at full load [Ghz] Number of processors GPGPU/Coprocessor Number of GPGPUs/Coprocessors WHITE PAPER PERFORMANCE REPORT PRIMERGY RX350 S7 VERSION: Measurements with 2 Xeon E and up to 2 Intel Xeon Phi Co-Processor 5110P Theoretical maximum performance of a single coprocessor according to the data sheet: Intel Xeon Phi Co-Processor 5110P: 1011 GFlops During runtime the computer load was distributed over the system processors and the coprocessors by the benchmark. The LINPACK result is this made up of the sum of the performance values of the system processors and the coprocessors. Without Turbo Mode With Turbo Mode Rpeak [GFlops] Rmax [GFlops] Rpeak [GFlops] Rmax [GFlops] Xeon E Xeon E Xeon E Intel Xeon Phi Co-Processor 5110P Intel Xeon Phi Co-Processor 5110P LINPACK: PRIMERGY RX350 S7 2 Xeon E Intel Xeon Phi Co-Processor 5110P Xeon E Intel Xeon Phi Co-Processor 5110P Xeon E GFlops Fujitsu Technology Solutions Page 51 (54)

52 System comparison The following diagram illustrates the throughput of the PRIMERGY RX350 S7 in comparison to its predecessor, the PRIMERGY TX300 S6, in their most performant configuration. LINPACK: PRIMERGY RX350 S7 vs. PRIMERGY TX300 S6 GFlops % w/o GPU % GPU +2 GPUs PRIMERGY TX300 S6 2 Xeon X % 363 PRIMERGY RX350 S7 2 Xeon E Page 52 (54) Fujitsu Technology Solutions