HP SAS benchmark performance tests

HP SAS benchmark performance tests technology brief Abstract... 2 Introduction... 2 Test hardware... 2 HP ProLiant DL585 server... 2 HP ProLiant DL380 G4 and G4 SAS servers... 3 HP Smart Array P600 SAS controller... 3 HP Smart Array 6402 SCSI controller... 4 SAS SFF hot-plug hard drives... 4 SCSI hot-plug hard drives... 4 HP Modular Smart Array 50 Enclosure... 5 HP Modular Smart Array 30 Enclosure... 5 Test summaries... 6 TPC-H Benchmark test performance... 6 Jetstress Disk Performance Test... 7 LoadSim 2003... 7 Iometer... 8 Summary... 9 Appendix A: TPC-H Benchmark test... 11 Benchmark configuration... 11 TPC-H test results... 12 Appendix B: Exchange Server 2003 test... 13 Appendix C: Iometer benchmark test... 23 For more information... 27 Call to action... 27

Abstract HP performed benchmark tests of ProLiant servers configured with Serial Attached SCSI (SAS) and Ultra320 SCSI (U320) controllers and drives to compare their performance in various application environments (OLTP, web server, database, and e-mail). This paper presents the results of these performance benchmark tests and explains how customers can collect performance data to size SAS small form factor drive configurations for Exchange server deployments. Introduction Serial Attached SCSI (SAS) is a point-to-point topology that enables highly scalable storage systems internal, external, or a combination of both and gives manufacturers and customers the flexibility to design and deploy a range of solutions. SAS incorporates low-voltage differential (LVD) signaling, which reduces the effects of capacitance, inductance, and noise experienced by parallel SCSI at higher speeds. Unlike traditional parallel SCSI that shares the bandwidth of one bus (320 MB/s) with many devices, SAS provides maximum bandwidth (300 MB/s) to each device greatly improving scalability and performance. This paper presents the results of the TPC Benchmark H (TPC-H) test published by HP in August 2005 for the HP ProLiant DL585 4P Opteron Dual Core server. The results, which are available at http://www.tpc.org/tpch/default.asp, show that the 4-way HP ProLiant DL585 system has the best performance and price/performance among single (non-clustered) industry-standard servers. This paper also presents the results of Microsoft Exchange and Iometer benchmark tests performed by HP to compare the performance of SAS storage arrays to 3.5 inch U320 SCSI storage arrays. In Microsoft Exchange 2003 tests, the Small Form Factor 10K RPM SAS storage array consistently outperformed 10K RPM U320 SCSI storage arrays. Iometer benchmark tests show that the performance of a 10K RPM SAS storage array (with extra spindles) meets or exceeds 10K RPM and 15K RPM U320 SCSI storage array performance. Test hardware This section describes the servers, controllers, hard drives, and storage enclosures used in the HP benchmarks tests described in this paper. HP ProLiant DL585 server The 4-processor ProLiant DL585 x86 server supports the AMD Opteron model 865 and 875 Dual- Core processors running at 1.8 and 2.2 GHz with a 1MB L2 cache. The Dual-Core architecture provides simultaneous 32-bit and 64-bit computing capabilities. The server uses the AMD 8000 Series chipset, supports 64GB of memory (either PC2700 or PC3200), and has seven independent PCI-X bus segments. Figure 1. ProLiant DL585 server 2

HP ProLiant DL380 G4 and G4 SAS servers The flexibility of the ProLiant DL380 G4 server (Figure 2) makes it ideal for a wide variety of applications for businesses of all sizes. The DL380 G4 and G4 SAS servers use the Intel Xeon processor (up to 3.6 GHz) with 1-MB or 2-MB L2 cache and Intel E7520 chipset. The server supports up to 12 GB of memory using PC2-3200R DIMMs in six slots. 1 The DL380 G4 model supports up to six 3.5-inch SCSI drives while the G4 SAS model supports up to eight small form factor (SFF) SAS drives. For the benchmark performance tests, the Smart Array 6402 SCSI controller was installed in the DL380 G4 server and connected to the Modular Smart Array 30 enclosure. The Smart Array P600 SAS controller was installed in the DL380 G4 SAS server and connected to the Modular Smart Array 50 enclosure. Figure 2. ProLiant DL380 G4 and G4 SAS servers ProLiant DL380 G4 server ProLiant DL380 G4 SAS server HP Smart Array P600 SAS controller The HP Smart Array P600 (SA-P600) SAS controller (Figure 3) provides high levels of performance and reliability for HP ProLiant servers through its support of the latest SCSI technology and advanced RAID capabilities. This SA-P600 offers twice the bandwidth of a 4-channel U320 array controller. Eight SAS physical links are distributed across two internal x4 wide port connectors and one external x4 wide port connector. The SA-P600 controller supports drive mirroring (RAID 1, 1+0), distributed data guarding (RAID 5), and RAID 6 (HP Advanced Data Guarding) 2. Figure 3. Smart Array P600 Controller 1 See specific server configuration in the detailed test results in Appendix B and C. 2 For information about RAID ADG, go to http://h18004.www1.hp.com/products/servers/proliantstorage/arraycontrollers/adg/. 3

HP Smart Array 6402 SCSI controller The Smart Array 6402 (SA-6402) controller (Figure 4) provides high performance, flexibility, and reliable data protection for HP ProLiant servers. The SA-6402 is ideal for workgroup and departmental servers. The 64-bit, 133-MHz PCI-X interface delivers data bandwidth up to 320 MB/s bandwidth per channel. The SA-6402 controller used in the HP benchmark tests had a Double Data Rate (DDR) 256-MB battery-backed write cache (BBWC) architecture and a hardware RAID engine for RAID 0, 1+0, 5, and 6 support. Figure 4. Smart Array 6404 controller SAS SFF hot-plug hard drives SAS is a point-to-point architecture in which all storage devices connect directly to a SAS port rather than sharing a common bus, as traditional SCSI devices do. Point-to-point links increase data throughput and improve the ability to locate and fix disk failures. SAS links are full duplex; they send and receive information simultaneously, thereby reducing a major source of latency. The speed of the first-generation SAS link is 3.0 Gb/s (see Table 1). The SAS interface allows for combining multiple links to create 2x, 3x, or 4x connections for scalable bandwidth. Table 1. Characteristics of SAS small form factor drives 10K RPM SAS SFF Transfer rate (maximum) Capacity 3 Gb/s ; 300MB/s 72 GB, 36 GB Height 0.591 in (15.0 mm) Width Seek Time (average) 2.75 in (69.8 mm) 4.0 ms SCSI hot-plug hard drives HP Ultra320 hard drives provide a maximum transfer rate of 320 MB/s. The 15K RPM U320 SCSI drives access data up to 26 percent faster than 10K RPM parallel SCSI drives, depending on the application and the configuration (see Table 2). 4

Table 2. Characteristics of U320 SCSI drives 15K RPM U320 10K RPM U320 Transfer rate (max) 320 MB/s 320 MB/s Capacity 146GB, 72 GB, 36 GB 300, GB, 72 GB, Height 1.0 in (25.4 mm) 1.0 in (25.4 mm) Width 4.0 in (101.6 mm) 4.0 in (101.6 mm) Seek Time (average) 3.8 ms 5.2 ms HP Modular Smart Array 50 Enclosure The 1U Modular Smart Array 50 Enclosure (MSA50) is designed to support up to ten Universal form factor SAS and Serial ATA (SATA) SFF hard drives (Figure 5). This MSA50 provides optimized performance per U of rack space with lower power consumption than SCSI drive enclosures. The MSA50 enclosure supports HP ProLiant servers containing an SA-P600 array controller. Figure 5. MSA50 enclosure HP Modular Smart Array 30 Enclosure The 3U Modular Smart Array 30 Enclosure (MSA30) with single bus I/O module supports up to fourteen 1-inch Ultra2, Ultra3, or Ultra320 Universal hard disk drives (Figure 6). Its modular design allows the MSA30 to be used in any storage configuration, including JBOD (Just a Bunch of Disks), Smart Array storage enclosures, and external multi-vendor modular arrays. The MSA30 enclosure supports HP ProLiant servers containing an SA-6402, SA-642, or SA-6i array controller. Figure 6. MSA30 enclosure 5

Test summaries HP engineers tested the performance of SAS and U320 disk arrays using the TPC Benchmark H (TPC-H) test, Microsoft Exchange Server 2003 benchmark tests (Load Simulator 2003 and Jetstress), and the open source Iometer benchmark test. The configurations and results of each test are summarized below. Refer to Appendices A, B, and C for detailed test descriptions. TPC-H Benchmark test performance The TPC-H test) measures the performance of decision support systems that examine large volumes of data, execute highly complex queries, and answer critical business questions. The TPC-H test contains a collection of business oriented ad-hoc queries and simultaneous data modifications. The database size for the TPC-H test is 300 GB. The queries and the database contents are selected to have broad, industry-wide relevance. The primary performance metrics are the TPC-H Composite Query-per-Hour Performance Metric (QphH@300GB) and the TPC-H Price/Performance metric ($/QphH@300GB). The QphH@300GB metric indicates the size of the database against which the queries are executed, the query processing power when queries are submitted by a single stream, and the query throughput when queries are submitted by multiple concurrent users. The $/QphH@300GB metric indicates the total configuration cost, including all hardware systems and software used, calculated over a five-year ownership period. The configuration cost is then divided by the performance to give a simple measure of price/performance. Table 3 compares the results of the TPC-H tests conducted on HP ProLiant DL585 servers running IBM DB2 UDB 8.2 and connected to SAS and U320 disk arrays. The operating system used for the benchmark tests was Red Hat Enterprise Linux 4 AS. See Appendix A for detailed test results. Table 3. Summary of TPC-H Benchmark results Hardware Software Total System Cost QppH @ 300GB QthH @ 300GB QphH @ 300GB $ / QphH@ 300GB HP ProLiant DL585 with 10K RPM SAS disk array HP ProLiant DL585 with 15K RPM U320 SCSI disk array IBM DB2 UDB 8.2 Red Hat Enterprise Linux 4 AS IBM DB2 UDB 8.2 Red Hat Enterprise Linux 4 AS $288,751 15504.0 9157.2 11915.3 $24.24 USD $254,586 11511.2 6181.1 8,435.2 $30.18 US D 6

These published results demonstrate the power and price/performance leadership of HP Opteron processor-based servers and HP Smart Array SAS technology. As of the publication date of this paper, the 300 GB TPC-H benchmark test results published on the TPC web site (at http://www.tpc.org/tpch/results/tpch_perf_results.asp?resulttype=noncluster&version=2%&currencyi D=0) show that the 4-way HP ProLiant DL585 server with the SAS disk array has the best performance and price/performance among single (non-clustered) industry-standard servers, including competitors 8-way servers. Jetstress Disk Performance Test Jetstress verifies the performance of a disk subsystem by simulating the Exchange disk I/O load produced by a specified number of clients. Jetstress tests were run on a ProLiant DL380-G4 server using the following controller-disk array configurations: SA-P600 controller attached to an MSA50 enclosure populated with ten 36-GB, 10K RPM SAS drives SA-6402 controller attached to an MSA30 enclosure with ten 10K RPM, 36-GB, U320 SCSI drives SA-6402 controller attached to an MSA30 enclosure with ten 15K RPM, 36-GB, U320 SCSI drives The configurations were tested at RAID levels 0, 1+0, 5, and RAID 6. The default cache settings and stripe size were used. For each configuration and RAID level, HP used Jetstress to generate I/O requests simulating Exchange database read and write loads. Engineers loaded each configuration with additional threads and recorded the number of transfers per second until the disk read or write latency exceeded 20 milliseconds (ms) the threshold limit recommended by Microsoft. In each Jetstress test that HP conducted, the read latency was the first metric to exceed 20 ms. Therefore, the read latency was used as the key metric in determining the maximum transfers per second achieved by each configuration before it exceeded the 20 ms-threshold. The test results are summarized as follows: The SA-6402 controller configured with the 15K RPM U320 SCSI drives achieved the highest number of disk transfers per second below a read latency of 20 ms. The SA-P600 controller with 10K RPM SAS drives consistently outperformed the SA-6402 controller with the 10K RPM U320 SCSI drives. The SA-P600 controller and 10K RPM SAS drive configuration had significantly lower write latencies than the 15K RPM U320 and 10K RPM U320 SCSI configurations. See Appendix B for detailed test results. LoadSim 2003 HP engineers used LoadSim 2003 to determine how three disk storage subsystems (connected to a ProLiant DL380 G4 server running Exchange 2003) respond to e-mail loads generated by 1,200 simulated Messaging API (MAPI) clients. LoadSim uses MAPI Messaging Benchmark 3 (MMB-3) to send multiple messaging requests from the simulated clients to the Exchange server, thereby creating an e-mail load. Engineers configured the LoadSim tests to simulate the e-mail load of 1,200 MMB-3 clients on the Exchange server with the following controller-disk subsystems configurations: SA-P600 controller attached to an MSA50 enclosure populated with ten 36-GB, 10K RPM SAS drives SA-6402 controller attached to an MSA30 enclosure with ten 10K RPM, 36-GB, U320 SCSI drives SA-6402 controller attached to an MSA30 enclosure with ten 15K RPM, 36-GB, U320 SCSI drives 7

The LoadSim tests were conducted on each configuration at RAID 1+0 and RAID 5. These two RAID levels are prevalent in Microsoft Exchange deployments because they provide the best combination of performance and fault tolerance. At each RAID level, 15K RPM U320 SCSI drives achieved the lowest client response time, while the 10K RPM U320 SCSI drives had the highest client response time. As expected, the 15K RPM U320 SCSI drives achieved a lower client response time than an equal number of 10K RPM SAS drives because 15K RPM drives have a lower seek time. Iometer The Iometer benchmark tool generates controlled disk I/O loads and measures the corresponding performance (I/Os per second or MB/s) of a drive storage configuration. Iometer runs on the server and allows a storage configuration to be stressed by varying I/O parameters such as transfer request size, queue depth, and percent read/write distribution. One purpose of the Iometer test was to emulate the performance of a ProLiant DL380 G4 SAS server, which can support up to eight SFF SAS drives, and the performance of a DL380 G4 server, which can support up to six 3.5-inch U320 SCSI drives. To emulate the performance of the DL380 G4 SAS server, HP used an SA-P600 SAS controller connected to an MSA50 enclosure containing eight 10K RPM SAS drives. To emulate the performance of the DL380 G4 server, HP used an SA-6402 SCSI controller connected to an MSA30 enclosure containing six 15K RPM U320 SCSI drives. These configurations were subjected to OLTP and web server performance tests, maximum throughput tests, and random and sequential read/write performance tests. HP engineers also conducted drive-to-drive performance tests with the following controller-disk subsystem configurations at RAID 5: SA-P600 controller with 256-MB BBWC connected to an MSA50 enclosure containing ten 10K RPM, 72-GB, SFF SAS drives. SA-6402 controller with 256-MB BBWC connected to an MSA30 enclosure containing ten 10K RPM, 72-GB, U320 SCSI drives. SA-6402 controller with 256-MB BBWC connected to an MSA30 enclosure containing ten 15K RPM, 72-GB, U320 SCSI drives. For the various Iometer tests, engineers used the transfer request sizes shown in Table 4. Other test parameters, such as the Percent Read/Write Distribution and the Random/Sequential Distribution, are provided in Appendix C. Table 4. Transfer request sizes used for Iometer tests Iometer test OLTP Web server Random read/write Sequential read/write Maximum throughput Transfer request size 2 KB and 8 KB 8 KB 64 KB 64 KB and 512 KB 1 MB The Iometer test results are summarized as follows: Iometer tests results show that with low levels of demand (low queue depth) the faster rotational speed of the 15K RPM drives provided greater performance than the 10K RPM SFF SAS drives. However, when the demands on the storage subsystem were increased to a queue depth of 32 or 64, the performance of the eight 10K RPM SAS drives in parallel was higher than the six 15K RPM U320 SCSI drives. 8

The SAS drive array performed better in random 64 KB read/write tests primarily because the SA-P600 controller has twice the bandwidth of the SA-6402 SCSI controller. The results of the sequential read/write performance tests demonstrate the dominant performance of the SAS architecture. For block sizes of 64 KB, 512 KB, and 1 MB, the SA-P600 controller significantly outperformed the SA-6402 controller. The higher performance of SAS is attributed to: Higher available bandwidth than the U320 SCSI storage system Better scalability More powerful SA-P600 controller For thoroughness, the engineers also conducted performance tests to compare identical numbers of SAS and U320 drives. In drive-to-drive comparisons between configurations with ten 10K RPM SAS SFF drives, ten 10K RPM U320 SCSI drives, and ten 15K RPM U320 SCSI drives, engineers found that: For random workloads (OLTP 2K, OLTP 8K, Web server, 64 KB random read): 15K RPM U320 SCSI drives perform better than an equal number of 10K RPM SAS SFF drives 10K RPM U320 SCSI drives and 10K RPM SAS SFF drives have very similar performance capabilities. For sequential (streaming) workloads (64 KB, 512 KB, 1 MB sequential read or sequential write), the performance of the SA-P600 with 10K RPM SFF drives is substantially better than the SA-6402 with 10K RPM or 15K RPM U320 SCSI drives. Due to the complete saturation of the SCSI bus, increasing the number of U320 SCSI drives in the MSA30 enclosure will not result in higher SCSI performance. Summary As of the date of publication of this paper, TPC-H test results show that the 4-way HP ProLiant DL585 Opteron Dual-Core server with external SAS storage has the best performance and price/performance among single (non-clustered) industry-standard servers. The performance of 10K RPM SAS drives in Exchange (Jetstress and LoadSim 2003) benchmark tests demonstrate that first-generation 10K RPM SAS drives consistently perform better than an equal number of 10K RPM U320 drives. The higher SAS performance was achieved with a first-generation link speed of 3.0 gigabits/second (Gb/s). The benchmark performance results confirm the bright future of SAS, as the speed of the second-generation SAS link will be 6.0 Gb/s. At 6.0 Gb/s, wide links (2x, 3x, and 4x) will enable aggregate data transfers up to 24 Gb/s. HP conducted Iometer tests that emulated the performance of a fully configured ProLiant DL380 G4 SAS server with eight SFF SAS drives and a fully configured DL380 G4 server with six U320 SCSI drives. The test results showed that at higher queue depths, representing a heavily used server, the performance of the eight 10K RPM SAS drives in parallel was higher than that of the six 15K RPM U320 SCSI drives. The advantages of SAS technology are far reaching: Investment protection The SAS small form factor is the new universal drive. Scalability The SAS interface allows for combining multiple links to create 2x, 3x, or 4x connections. High performance Point-to-point links increase data throughput and improve the ability to locate and fix disk failures. Higher density Eighteen SFF SAS drive bays occupy the same amount of space as ten 3.5-inch Ultra320 SCSI drive bays in the ML570 server. Three MSA-50 enclosures with 30 SFF drive bays fit in the same rack space as one MSA-30 enclosure with fourteen 3.5-inch drive bays. (Each SA-P600 SAS controller supports up to two MSA-50 enclosures.) 9

High efficiency SFF drives enable better airflow and consume half as much power as 3.5-inch drives. Flexibility The SAS architecture enables system designs that deploy both SAS and SATA devices. 10

Appendix A: TPC-H Benchmark test In August 2005, HP published a full disclosure report documenting the methodology and results of the TPC Benchmark H test conducted on the HP ProLiant DL585 using IBM DB2 UDB 8.2, in conformance with the requirements of the TPC Benchmark H Standard Specification, Revision 2.3.0. The operating system used for the benchmark was Red Hat Enterprise Linux 4 AS. A representative of Performance Metrics, Inc. audited the: benchmark configuration environment and methodology used to produce and validate the test results the pricing model used to calculate the TPC-H Composite Query-per-Hour Performance Metric (QphH@300GB) and the TPC-H Price/Performance metric ($/QphH@300GB) The information in this Appendix summarizes the benchmark configuration and the TPC-H benchmark test results. A copy of the full disclosure report can be obtained from the Transaction Processing Performance Council at http://www.tpc.org/tpch/default.asp. Benchmark configuration The HP ProLiant DL585 server included four 2.2-GHz AMD Dual-Core Opteron Model 875 processors with 1-MB L2 cache and 64 GB of main memory (Figure A-1). The server was running IBM DB2 UDB 8.2 and the Red Hat Linux 4 AS operating system. The server had eight SA-P600 SAS controllers connected to sixteen MSA50 enclosures, each with ten 10K RPM SFF SAS drives. The total storage capacity was 5898.8 GB. Figure A-1. TPC-H benchmark hardware configuration 11

TPC-H test results Figure A-2 shows the report summary of the benchmark results for the HP ProLiant DL585. A detailed list of all hardware and software, including the 3-year price, is provided in the full disclosure report at http://www.tpc.org/tpch/default.asp. Figure A-2. TPC- H full disclosure report summary for the HP ProLiant DL585 2.2GHz 4P AMD Opteron Dual Core server 12

Appendix B: Exchange Server 2003 test Jetstress test results Microsoft s Jetstress 2004 utility was used to simulate Exchange I/0 against a storage subsystem to test the performance and determine the maximum number of Exchange IOPs that the subsystem can support. Engineers performed Jetstress tests on each controller-disk configuration at RAID levels 0, 1+0, 5, and 6. For each configuration, engineers increased the number of Jetstress threads until the I/O latency exceeded the acceptable threshold limit of 20 ms, as recommended by Microsoft. In each Jetstress test performed by HP, the read latency was the first metric to exceed 20 ms. Therefore, engineers used read latency as the key metric in determining the pass/fail status of each configuration. Storage array configurations Engineers performed Jetstress tests on one SAS array configuration and two SCSI array configurations connected to a ProLiant DL380 G4 server (Figure B-1). The controller-disk subsystem configurations were as follows: Smart Array P600 controller attached to an MSA50 enclosure populated with ten 36-GB, 10K RPM SAS drives Smart Array 6402 controller attached to an MSA30 enclosure with ten 10K RPM, 36-GB, U320 SCSI drives Smart Array 6402 controller attached to an MSA30 enclosure with ten 15K RPM, 36-GB, U320 SCSI drives Figure B-1. Controller-disk subsystem configurations for Jetstress tests 13

The tests were performed using RAID levels 0, 1+0, 5, and 6. Engineers used the default cache settings 50 percent read and 50 percent write as well as the default stripe sizes indicated in table B-1. Table B-1. Default stripe sizes used in Jetstress tests RAID level Default stripe size 0 128 KB 1+0 128 KB 5 64 KB 6 16 KB The Jetstress GUI (Figure B-2) was used to configure the test settings shown in Table B-2. Table B-2. Test settings used in Jetstress tests Test parameter Test duration Setting 2 hr Estimated IOPs per mailbox 1 Mailbox size limit 100 MB No. of mailboxes on server 1,000 Hardware storage cache 256 MB No. of storage groups 1 Figure B-2. Jetstress GUI displaying test configuration settings 14

The Jetstress GUI Advanced screen (Figure B-3) allowed engineers to further customize the test parameters. The default configuration allows Jetstress to self-tune to determine the maximum number of I/Os that a storage subsystem can support. Allowing Jetstress to self-tune would result in configuration parameters that varied between tests, which does not allow an accurate comparison of controller and disk configurations. Therefore, engineers specified a fixed workload for all tests and varied only the number of threads until the average latency for disk reads or writes exceeded 20 ms. The parameters in Advanced Settings were configured as shown in Table B-3. Table B-3. Test parameters defined in Jetstress advanced settings Test setting Threads Parameter Variable* Log buffer 64 Inserts 20 Replaces 75 Deletes 5 Lazy commits 93 * The number of threads was varied between tests to increase the I/O load. The values shown in Table B-3 and Figure B-3 result in an approximate 65:35 read/write ratio workload. This read/write ratio is typical of a corporate workload. Figure B-3. Jetstress Advanced settings 15

RAID 0 test results Raid 0 is the highest performing RAID level configuration because it performs striping across all members of the array. However, RAID 0 provides no fault tolerance capabilities and is susceptible to catastrophic data loss in the event of a single disk failure. Due to possible data loss, RAID 0 is seldom used for production deployments. The results of the Jetstress tests at RAID 0 are shown in Figure B-4 and summarized below. The SA-6402 with 15K RPM U320 SCSI drives averaged 1757 transfers per second at 16 Jetstress threads with an average read latency of 0.013 seconds and an average write latency of 0.0017 seconds. SA-P600 with 10K RPM SAS drives averaged 1526 transfers per second at 16 Jetstress threads with an average read latency of 0.015 seconds and an average write latency of 0.00018. The SA-6402 with the 10K RPM Ultra 3 SCSI drives averaged 1241 transfers per second at 16 Jetstress threads with an average read latency of 0.019 and an average write latency of 0.00167. Figure B-4. Jetstress test data at RAID 0 16

RAID 1+0 test results RAID 1+0 uses disk striping and data mirroring to provide the highest level of performance with the highest level of fault tolerance to protect against data loss in the event of a hard disk failure. In a RAID 1+0 logical drive, 50 percent of the available disk capacity is required for data mirroring. RAID 1+0 is commonly used in production environments when both performance and fault tolerance are required. The results of the Jetstress tests at RAID 1+0 are shown in Figure B-5 and summarized below. The SA-6402 with 15K RPM U320 SCSI drives averaged 1409 transfers per second at 14 Jetstress threads with an average read latency of 0.015 seconds and an average write latency of 0.0019 seconds. SA-P600 with 10K RPM SAS drives averaged 1133 transfers per second at 14 Jetstress threads with an average read latency of 0.0186 seconds and an average write latency of 0.000197. The SA-6402 with the 10K RPM Ultra 3 SCSI drives averaged 1075 transfers per second at 14 Jetstress threads with an average read latency of 0.0196 and an average write latency of 0.00178. Figure B-5. Jetstress test data at RAID1+0 17

RAID 5 test results RAID 5 uses striping with parity to provide increased performance and fault tolerance. RAID 5 logical drives do not provide the same level of performance achieved with a RAID 1+0 logical drives. However, RAID 5 logical drives only require 1/n (n = number of drives in the logical drive) of the total disk capacity to provide a level of fault tolerance for data protection. In general, a RAID 5 array is less costly than a RAID 1+0 array, which requires 50 percent of the total disk capacity of a logical drive for data mirroring. The results of the Jetstress tests at RAID 5 are shown in Figure B-6 and summarized below. The SA-6402 with 15K RPM U320 SCSI drives averaged 1218.5 transfers per second at 16 Jetstress threads with an average read latency of 0.0197 seconds and an average write latency of 0.00618 seconds. SA-P600 with 10K RPM SAS drives averaged 873.8 transfers per second at 10 Jetstress threads with an average read latency of 0.0173 seconds and an average write latency of 0.00018. The SA-6402 with the 10K RPM Ultra 3 SCSI drives averaged 665.3 transfers per second at 8 Jetstress threads with an average read latency of 0.0182 and an average write latency of 0.0017. 18

Figure B-6. Jetstress test data at RAID 5 19

RAID 6 test results RAID 6, unique to HP Smart Array controllers, provides additional data protection by recording two independent sets of parity data. The results of the Jetstress tests with RAID 6 are shown in Figure B-7 and summarized below. The SA-6402 with 15K RPM U320 SCSI drives averaged 825.95 transfers per second at 10 Jetstress threads with an average read latency of 0.018 seconds and an average write latency of 0.0033 seconds. SA-P600 with 10K RPM SAS drives averaged 544 transfers per second at 6 Jetstress threads with an average read latency of 0.016 seconds and an average write latency of 0.00029. The SA-6402 with the 10K RPM Ultra 3 SCSI drives averaged 492 transfers per second at 6 Jetstress threads with an average read latency of 0.018 and an average write latency of 0.003. Figure B-7. Jetstress test data with RAID 6 20

Conclusions The SA-6402 controller configured with the 15K RPM U320 SCSI drives achieved the highest number of disk transfers per second below the 20 ms threshold. The SA-P600 controller with 10K RPM SAS drives consistently outperformed the SA-6402 controller configured with the 10K RPM U320 SCSI drives. The SA-P600 controller with 10K SAS drives had significantly lower write latencies that the SCSI storage configurations. During several tests, the disk transfers per second decreased when additional load was placed on the controller. This performance curve is an indication that the controller throughput is saturated and optimal performance was achieved with less load. 21

LoadSim 2003 test results HP engineers performed LoadSim tests to validate the performance of SAS and SCSI storage subsystems connected to a ProLiant DL380 G4 server running Exchange 2003. The LoadSim tests were configured to run 1,200 simulated MMB-3 client profiles. The controller-disk subsystems configurations included the following: SA-P600 controller attached to an MSA50 enclosure populated with ten 36-GB, 10K RPM SAS drives. SA-6402 controller attached to an MSA30 enclosure with ten 10K RPM, 36-GB, U320 SCSI drives. SA-6402 controller attached to an MSA30 enclosure with ten 15K RPM, 36-GB, U320 SCSI drives. Engineers performed LoadSim tests on each configuration at RAID 1+0 and RAID 5. These two RAID levels are prevalent in Microsoft Exchange deployments because they provide the best combination of performance and fault tolerance. Given a fixed number of simulated users (1,200), the engineers used client response time as the key metric in evaluating the relative performance of the controller-disk subsystems at each RAID level. Figure B-8 shows that at each RAID level, the 15K RPM U320 SCSI drives achieved the fastest client response time and the 10K RPM U320 SCSI drives had the slowest response time. Figure B-8. LoadSim MMB-3 response time for 1200 simulated clients The LoadSim MMB-3 tests simulating 1,200 user profiles generated an average of 580 transfers per second during each test. Reviewing the results from the Jetstress tests at RAID 1+0, this average transfer rate is well within the capabilities of the SA-P600 and the SA-6402 controllers. For example, during the Jetstress test at RAID 1+0, the SA-P600 controller successfully supported 1133 transfers per second with an average latency below the 20 ms threshold. 22

Appendix C: Iometer benchmark test All Iometer tests were performed using an HP ProLiant DL380-G4 SAS server with a 3.60-GHz CPU (1-MB L2 Cache), an 800-MHz front side bus, 1024 MB of system memory, and server ROM: P51 (08/26/2004). For the SAS storage system, an SA-P600 controller 3 with 256-MB BBWC was installed in the server and connected to an MSA50 enclosure. For the U320 SCSI storage systems, an SA-6402 controller 4 with 256-MB BBWC was installed and connected to an MSA30 enclosure. OLTP, web server, and random read/write performance The OLTP and web server performance of a SAS storage array was compared to that of two U320 SCSI storage arrays. The controller in each array was set at RAID 5. The SAS storage array consisted of an MSA50 enclosure housing eight 72-GB, 10K RPM, SFF SAS drives. The SCSI storage arrays consisted of two MSA30 enclosures: one enclosure housing six 72-GB, 10K RPM, U320 drives and the other housing six 72-GB, 15K RPM, U320 drives. Eight SFF SAS drives and six U320 SCSI drives were used to match the maximum number of internal drives supported by the DL380 G4 SAS server and DL380 G4 server, respectively. Iometer tests results show that with low levels of demand (low queue depth), the faster rotational speed of the 15K RPM drives provided greater performance than the 10K RPM SFF SAS drives (Figure C-1). However, when the demands on the storage array were increased to a queue depth of 32 or 64, the performance of the eight 10K RPM SAS drives in parallel was higher than that of the six 15K RPM U320 SCSI drives. The higher queue depths represent an OLTP or web server that is heavily used. Figure C-2 provides a more in depth comparison that shows the relative performance of the SAS storage array as a percentage of the U320 SCSI storage array data rate. The SAS drive array performed better in 64 KB random read tests (at higher queue depths) and random write tests primarily because the SA-P600 controller has twice the bandwidth of the SA-6402 SCSI controller. Figure C-1. OLTP and web server performance for eight 10K RPM SFF SAS drives versus six 10K RPM U320 drives at various queue depths 3 Firmware version 2.36 4 Firmware version 1.14 23

Figure C-2. Relative performance of the SAS storage array and the U320 SCSI storage array at various queue depths Sequential read/write performance HP tested the sequential read/write performance and the maximum transfer rate of the SAS and U320 storage arrays. Both controllers were set at RAID 5. To test maximum throughput, engineers set the Transfer Request Size to 1 MB, the Read/Write Distribution to 100 percent Read, and the Random/Sequential Distribution to 100 percent Sequential. The greater bandwidth of the SAS architecture is evident in the results for sequential read/write performance tests (see Figures C-3 and C-4). For block sizes of 64 KB, 512 KB, and 1 MB, the SA- P600 controller significantly outperformed the SA-6402 controller. The higher performance of SAS can be attributed to: Higher available bandwidth than the U320 SCSI storage system Better SAS scalability More powerful SA-P600 controller 24

Figure C-3. Sequential read/write performance of eight 10K RPM SFF SAS drives and of six 10K RPM U320 drives at various queue depths Figure C-4. Relative sequential read/write performance of eight 10K RPM SFF SAS drives and six 15K RPM U320 drives at various queue depths 25

Drive-to-drive performance comparison In a direct drive-to-drive comparison, HP engineers performed the OLTP, web server, and sequential read/write tests on ten SAS, U320 10K RPM, and U320 15K RPM drives. As expected, equal numbers of 15K RPM U320 drives hold an I/O performance advantage over 10K RPM SFF SAS and 10K RPM U320 drives because the 15K RPM drives have lower latency (Figure C-5). Also because of their lower latency, 15K RPM U320 drives hold slight advantage over 10K RPM SFF SAS and 10K RPM U320 drives in random reads (Figure C-6). Increased SAS bandwidth and the improved array controller provide higher performance for sequential read/write operations. Figure C-5. OLTP and web server performance for equal number of SAS and U320 drives Figure C-6. Random and sequential read/write performance and maximum throughput for equal number of SAS and U320 drives 26

For more information TPC-H benchmark test results are published on the TPC web site at http://www.tpc.org/tpch/results/tpch_perf_results.asp?resulttype=noncluster&version=2%&currencyid=0 Call to action Please send comments about this paper to: TechCom@HP.com. Copyright 2005 Hewlett-Packard Development Company, L.P. The information contained herein is subject to change without notice. The only warranties for HP products and services are set forth in the express warranty statements accompanying such products and services. Nothing herein should be construed as constituting an additional warranty. HP shall not be liable for technical or editorial errors or omissions contained herein. Intel and Xeon are trademarks or registered trademarks of Intel Corporation or its subsidiaries in the United States and other countries. Microsoft and Windows are U.S. registered trademarks of Microsoft Corporation.IBM, DB2 and DB2 UDB are registered trademarks of International Business Machines Corporation. TPC Benchmark, TPC-H, QppH, QthH and QphH are trademarks of the Transaction Processing Performance Council. Linux is a registered trademark of Linus Torvalds in the United States, other countries, or both. TC050903TB, 09/2005