To MIPS or Not to MIPS. That is the CP Question!

Transcription

1 To MIPS or Not to MIPS That is the CP Question! SHARE Seattle EWCP Gary King IBM March 4,

2 2 Trademarks Systems & Technology Group The following are trademarks of the International Business Machines Corporation in the United States and/or other countries. AlphaBlox* APPN* CICS* CICS/VSE* Cool Blue DB2* DFSMS DFSMShsm DFSMSrmm DirMaint DRDA* DS6000 DS8000 ECKD ESCON* FICON* FlashCopy* * Registered trademarks of IBM Corporation GDPS* HiperSockets HyperSwap IBM* IBM eserver IBM logo* IMS Language Environment* Lotus* Large System Performance Reference (LSPR ) Multiprise* MVS OMEGAMON* Parallel Sysplex* Performance Toolkit for VM PowerPC* PR/SM Processor Resource/Systems Manager The following are trademarks or registered trademarks of other companies. RACF* Redbooks* Resource Link RETAIN* REXX RMF S/390* Scalable Architecture for Financial Reporting Sysplex Timer* Systems Director Active Energy Manager System/370 System p* System Storage System x* System z System z9* System z10 Tivoli* Tivoli Storage Manager TotalStorage* VSE/ESA VTAM* WebSphere* zenterprise xseries* z9* z10 z10 BC z10 EC z/architecture* z/os* z/vm* z/vse zseries* Adobe, the Adobe logo, PostScript, and the PostScript logo are either registered trademarks or trademarks of Adobe Systems Incorporated in the United States, and/or other countries. Cell Broadband Engine is a trademark of Sony Computer Entertainment, Inc. in the United States, other countries, or both and is used under license therefrom. Java and all Java-based trademarks are trademarks of Sun Microsystems, Inc. in the United States, other countries, or both. Microsoft, Windows, Windows NT, and the Windows logo are trademarks of Microsoft Corporation in the United States, other countries, or both. Intel, Intel logo, Intel Inside, Intel Inside logo, Intel Centrino, Intel Centrino logo, Celeron, Intel Xeon, Intel SpeedStep, Itanium, and Pentium are trademarks or registered trademarks of Intel Corporation or its subsidiaries in the United States and other countries. UNIX is a registered trademark of The Open Group in the United States and other countries. Linux is a registered trademark of Linus Torvalds in the United States, other countries, or both. ITIL is a registered trademark, and a registered community trademark of the Office of Government Commerce, and is registered in the U.S. Patent and Trademark Office. IT Infrastructure Library is a registered trademark of the Central Computer and Telecommunications Agency, which is now part of the Office of Government Commerce. * All other products may be trademarks or registered trademarks of their respective companies. Notes: Performance is in Internal Throughput Rate (ITR) ratio based on measurements and projections using standard IBM benchmarks in a controlled environment. The actual throughput that any user will experience will vary depending upon considerations such as the amount of multiprogramming in the user's job stream, the I/O configuration, the storage configuration, and the workload processed. Therefore, no assurance can be given that an individual user will achieve throughput improvements equivalent to the performance ratios stated here. IBM hardware products are manufactured from new parts, or new and serviceable used parts. Regardless, our warranty terms apply. All customer examples cited or described in this presentation are presented as illustrations of the manner in which some customers have used IBM products and the results they may have achieved. Actual environmental costs and performance characteristics will vary depending on individual customer configurations and conditions. This publication was produced in the United States. IBM may not offer the products, services or features discussed in this document in other countries, and the information may be subject to change without notice. Consult your local IBM business contact for information on the product or services available in your area. All statements regarding IBM's future direction and intent are subject to change or withdrawal without notice, and represent goals and objectives only. Information about non-ibm products is obtained from the manufacturers of those products or their published announcements. IBM has not tested those products and cannot confirm the performance, compatibility, or any other claims related to non-ibm products. Questions on the capabilities of non-ibm products should be addressed to the suppliers of those products. Prices subject to change without notice. Contact your IBM representative or Business Partner for the most current pricing in your geography.

3 MIPS - can one number fit all? It's commonplace to assign IBM System z processors a capacity rating called MIPS The MIPS rating is very often used ƒto track and set workload capacity requirements ƒto select the proper size processor for the workload Today, we will discuss ƒjust what are MIPS and where do they come from? ƒfor a given processor, do all workloads run at the same MIPS? ƒhow much trouble can using MIPS get us into? and what to do about it 3

4 Just what are MIPS? Once upon a time, MIPS really meant Millions of Instructions Per Second As commonly used today, MIPS has become a RELATIVE indicator of AVERAGE processor CAPACITY MIPS are based on capacity RATIOS between processors MIPS are still in the ballpark of real Mi/sec Generally speaking, MIPS of new processor = MIPS of old processor x the AVERAGE CAPACITY RATIO new:old 4

5 Average Capacity Ratio IBM System z sets average capacity ratios among processors based on a variety of measured workloads which are published in the Large System Performance Reference (LSPR) ƒ Old and new processors are measured in the same environment with the same workloads at high utilizations (>=90%) Over time, workloads and environment are updated to stay current with customer profiles ƒold processors measured with new workloads/environment may have different average capacity ratios compared to when they were originally measured 5

6 So, can one number (MIPS) fit all? To find out we have to ask ƒwhen is it okay to use an average and when is it not? Sources of variation from average capacity ratio ƒsystem design ƒworkload characteristics ƒworkload scaling ƒ utilization ƒlpar configurations ƒcoupling technology 6

7 System Design: Processor Processor Design (core) cycle time pipeline branch prediction hardware vs. millicode memory hierarchy (nest) high speed buffers (caches) on chip, on module private, shared buses number, bandwidth latency distance speed of light Example: z196 Single Book Memory 1 L3 Cache L4 Cache 4 1 L3 Cache 4 7

8 System Design: Hipervisor and OS Hipervisor (PR/SM) virtualization layer at OS level distributes physical resources memory processors logicals dispatched on physicals dedicated shared affinities OS virtualization layer at addrspc level distributes logical resources memory processors tasks dispatched on logcials Enhanced cooperation HiperDispatch starting with z10 EC z/os + PR/SM Logical View of 2 books Memory Cache PR/SM LP1 LP2 LPn TCBa TCBb z/osi TCBx Memory Cache LP1 LP2 LPm TCBa TCBb z/osj TCBx 8

9 Workload Characteristics Workload Characteristics instructions mix sequence branch characteristics task dispatch profile locked in or chatty memory size locality of reference multiprogramming level I/O rate 1 L3 Cache Memory L4 Cache 4 1 L3 Cache 4 9

10 LSPR z/os workload primitives CB-L commercial batch long job steps WASDB WebSphere-focus application server and data base OLTP-T traditional online transaction processing OLTP-W webenabled access to legacy data CHARACTERISTICS MORE IMPORTANT THAN NAME use profile I/O Memory Hierarchy CB-L heavy appl, light OS light light WASDB medium appl and OS light light/moderate OLTP-T medium appl and OS heavy moderate OLTP-W medium appl and OS moderate stress NOW RUN IN VARIOUS MIXES TO PRODUCE WORKLOADS MATCHING CUSTOMER PROFILE OF MEMORY HIERARCHY STRESS OR RELATIVE "NEST" INTENSITY (RNI) Relative Nest Intensity (RNI) - Activity beyond private cache(s) is the most sensitive area - Reflects distribution and latency of sourcing from shared caches and memory - Data for calculation available from MF (SMF 113) starting with z10 10

11 LSPR Workload Categories Categories developed to match the profile of data gathered on customer systems over 100 data points (LPARs) used in the profiling Various combinations of prior workload primitives are measured to reflect the new workload categories Applications include CICS, DB2, IMS, OSAM, VSAM, WebSphere, COBOL, utilities LOW (relative nest intensity) Workload curve representing light use of the memory hierarchy Similar to past high Nway scaling workload primitives AVERAGE (relative nest intensity) Workload curve expected to represent the majority of customer workloads Similar to the past LoIO-mix curve HIGH (relative nest intensity) Workload curve representing heavy use of the memory hierarchy Similar to the past DI-mix curve zpcr extends these published categories Low-Avg: 50% LOW and 50% AVERAGE Avg-High: 50% AVERAGE and 50% HIGH 11

12 System Design + Workload Characteristics Variation from Average: sometimes small Example: z990 to z9 EC z GHz superscalar Caches private 256k i, 256k d shared 32 MB / book book interconnect: ring Memory Cache z9 EC 1.7 GHz superscalar Caches private 256k i, 256k d shared 40 MB / book book interconnect: ring Memory Cache 12

13 znext Capacity Performance Highlights All performance numbers are not final and subject to change Full speed capacity models capacity ratio to z10 1.3x to 1.5x at equal Nway 2 1.6x max capacity (80w znext versus 64w z10) Capacity Ratio LSPR Single Image Capacity Ratios 10way: z9 EC versus z990 ITR Ratio Subcapacity models capacity ratio to full speed znext 0.20x x 0.64x JOB A CB-L individual job capacity ratios JOB B JOB C JOB D JOB E JOB F JOB G JOB H JOB I JOB J JOB K JOB L JOB M JOB N JOB O JOB P JOB Q JOB R 1.33 LOW RNI AVERAGE RNI HIGH RNI LSPR Workload 13

14 System Design + Workload Characteristics Variation from Average: sometimes large Example: z9 EC to z10 EC z9 EC 1.7 GHz superscalar Caches private 256k i, 256k d shared 40 MB / book book interconnect: ring Memory Cache z10 EC 4.4 GHz redesigned pipeline superscalar Caches private 64k i, 128k d.5 private 3 MB shared 48 MB / book book interconnect: star.5 Memory Cache

15 znext Capacity Performance Highlights All performance numbers are not final and subject to change 2 Full speed capacity models capacity ratio to z10 1.3x to 1.5x at equal Nway 1 1.6x max capacity (80w znext versus 64w z10) Capacity Ratio LSPR Single Image Capacity Ratios 10way: z10 EC versus z9 EC Subcapacity models capacity ratio to full speed znext x 0.49x x ITR Ratio 0 CB-L individual job capacity ratios JOB A JOB C JOB E JOB G JOB I JOB K JOB M JOB O JOB Q JOB B JOB D JOB F JOB H JOB J JOB L JOB N JOB P JOB R 1.42 LOW RNI AVERAGE RNI HIGH RNI LSPR Workload 15

16 16 z196 5 System Design Workload + Workload Scalability Characteristics Variation from Variation Average: from sometimes Average inbetween Example: Example: z10 EC z9 to ECz196 z10 EC z/os 1.8 Multi-Image 4.4 GHz Wide range around average Caches - intensive scale higher 10 private 64k i, 128k d - Memory intensive scale lower.5 private - size 3 of MB variation grows shared with 48 MB higher / book Nway book interconnect: star Capacity Ratio Memory 5.2 GHz Out-Of-Order execution L4 Cache Caches z9 EC ODE-B z9 EC WASDB z9 EC OLTP-W z9 EC MIXED private 64k i, 128k dz9 EC CB-L z9 EC OLTP-T L3 Cache L3 Cache 0 private 1.5 MB L3 shared 24 MB / chip L4 shared 192 MB / book book interconnect: star 1 Nway.5 4 Memory Cache

17 znext Capacity Performance Highlights All performance numbers are not final and subject to change Full speed capacity models capacity 2 ratio to z10 1.3x to 1.5x at equal Nway 1.6x max capacity (80w znext 1 versus 64w z10) Capacity Ratio LSPR Single Image Capacity Ratios 10way: z196 versus z10 EC Subcapacity models capacity 0 ratio to full speed znext JOB A JOB C JOB E JOB G JOB I JOB K JOB MJOB O JOB Q JOB B JOB D JOB F JOB H JOB J JOB L JOB N JOB P JOB R 0.20x 0.49x 0.64x ITR Ratio CB-L individual job capacity ratios LOW RNI AVERAGE RNI HIGH RNI LSPR Workload 17

18 System Design + Workload Characteristics Variation from Average: sometimes fairly small Example: z196 to zec12 z GHz Out-Of-Order execution Caches private 64k i, 128k d private 1.5 MB L3 shared 24 MB / chip L4 shared 192 MB / book zec GHz Enhanced Out-Of-Order Caches private 64k i, 96k d private 1 MB i + 1 MB d L3 shared 48 MB / chip L4 shared 384 MB / book 1 1 L3 Cache L3 Cache Memory L4 Cache 4 1 Memory L4 Cache 6 1 L3 Cache L3 Cache

19 znext Capacity Performance Highlights All performance numbers are not final and subject to change Full speed capacity models capacity 2 ratio to z10 1.3x to 1.5x at equal Nway 1.6x max capacity (80w znext 1 versus 64w z10) Capacity Ratio LSPR Single Image Capacity Ratios 10way: zec12 versus z196 Subcapacity models capacity 0 ratio to full speed znext JOB A JOB C JOB E JOB G JOB I JOB KJOB M JOB O JOB Q JOB B JOB D JOB F JOB H JOB J JOB L JOB N JOB P JOB R 0.20x 0.49x 0.64x ITR Ratio CB-L individual job capacity ratios LOW RNI AVERAGE RNI HIGH RNI LSPR Workload 19

20 z13 versus zec12 hardware comparison zec GHz Enhanced Out-Of-Order Caches private 64k i, 96k d private 1 MB i + 1 MB d L3 shared 48 MB / chip L4 shared 384 MB / book z GHz Major pipeline enhancements Caches private 96k i, 128k d private 2 MB i + 2 MB d L3 shared 64 MB / chip L4 shared 480 MB / node plus 224 MB NIC 1 L3 Cache PU1 L3 Cache Single Book View Memory L4 Cache 6 1 L3 Cache 6 Single Drawer View Memory Memory L4 Cache PU8 L3 Cache L3 Cache PU1 PU8 PU1 L4 Cache PU8 L3 Cache PU1 PU8

21 LSPR Single Image Capacity Ratios 16way: z13 versus zec12 Example of Workload Variability CB-L individual job capacity ratios 2 Capacity Ratio ITR Ratio 1 0 JOB A JOB C JOB E JOB G JOB I JOB K JOB MJOB O JOB Q JOB B JOB D JOB F JOB H JOB J JOB L JOB N JOB P JOB R LOW RNI AVERAGE RNI HIGH RNI LSPR Workload

22 Capacity Ratio Workload Scalability Variation from Average Example: z9 EC Range around average - Low RNI workloads scale higher - High RNI workloads scale lower - Size of variation grows with higher Nway - more pressure on shared caches Systems & Technology Group 0 z9 EC LOW z9 EC AVERAGE z9 EC HIGH Nway 22

23 Capacity Ratio Workload Scalability Variation from Average Example: moving from z990 to z9 EC Equal capacity upgrade - due to workload variation - could land on wrong machine - could be short or overconfigured - upgrade could appear to deliver too few or too many MIPS - for this move not much variation Systems & Technology Group 0 z990 LOW z990 AVERAGE z990 HIGH z9 EC LOW Nway z9 EC AVERAGE z9 EC HIGH 23

24 Workload Scalability Variation from Average Example: moving from z9 EC to z10 EC Systems & Technology Group 20 Equal capacity upgrade - due to workload variation - wider spread in upgrade scenarios than z990 to z9 EC Capacity Ratio z9 EC LOW z9 EC AVERAGE z9 EC HIGH z10 EC LOW Nway z10 EC AVERAGE z10 EC HIGH 24

25 Workload Scalability Variation from Average Example: moving from z10 EC to z196 Systems & Technology Group Equal capacity upgrade - due to workload variation - similar spread to that of z9 EC to z10 EC - but workload effect is reversed Capacity Ratio z10 EC LOW z10 EC AVERAGE Nway z10 EC HIGH z196 LOW z196 AVERAGE z196 HIGH 25

26 Workload Scalability Variation from Average Example: moving from z196 to zec12 Systems & Technology Group 40 Equal capacity upgrade - due to workload variation - for this move not much variation 30 Capacity Ratio z196 LOW z196 AVERAGE z196 HIGH zec12 LOW Nway zec12 AVERAGE zec12 HIGH 26

27 Workload Scalability Variation from Average Example: moving from zec12 to z13 Systems & Technology Group 40 Equal capacity upgrade - due to workload variation - for this move a fair amount of variation 30 Capacity Ratio zec12 LOW zec12 AVERAGE Nway zec12 HIGH z13 LOW z13 AVERAGE z13 HIGH 27

28 Utilization Source of Variation utilization generally reflects the amount of work flowing through a fixed HW/SW configuration the higher the workload rate, the higher the utilization As more work flows through a fixed HW/SW configuration, the efficiency of the HW and SW is reduced less shared HW resources (caches, buses) available to each work unit SW manages more work units - longer queues, more contention time per transaction or job will grow Magnitude of the effect is related to workload characteristics higher RNI workloads (as measured at higher utilizations) see higher impact size of the processor smallest Nways (say 1-4way) are somewhat less sensitive 28

29 time/trans OLTP Client Workload Example Growth in time/trans as busy increases Approx. 4% growth in /trx for each 10% growth in busy Busy 29

30 Utilization Impact to Capacity Planning When Using MIPS Impact to capacity planning comes in two flavors may have less headroom on the box than you think when moving a workload, it may not fit in the new container Example assume a workload is running at 50% busy on a 2000 MIPS box without factoring in utilization effect, it will be called a 1000 MIPS workload in fact, it may be an 1200 MIPS workload when running at the efficiency of a 90% busy box caution #1: there is NOT room to double this workload on the current box caution #2: if moved to a new box or LPAR, it will likely need a 1200 MIPS container (not 1000 MIPS) to fit Estimating the impact - conservative approach For a change in utilization of 10%, plan for the capacity effect to be 3% for LOW RNI workloads 4% for AVERAGE RNI workloads 5% for HIGH RNI workloads 30

31 LPAR Configurations Variation from Average MIPS LPAR configurations affect the efficiency of the HW and SW key factors workload characteristics number of LPARs number of logical processors and weight of each LPAR overall ratio of logical to physical processors MIPS ratings are based on AVERAGE wkld and median client LPAR config median client LPAR configuration varies by Nway number of LPARs 5 at low-end, 9 at high-end generally 2 are major (>20% of weight), rest are minor size of major LPARs close to Nway of box for low/mid-range Nways, well less than Nway at high-end logical:physical ratio 5:1 at low end, 2:1 for most, 1.3:1 at high end 31

32 Example LPAR Configurations Effect on MIPS 30 Holding workload constant - vary LPAR configuration 2x1/2 = only 2 "half-size" images MI = median client LPAR configuration SI = 1 large z/os image 2xMI = 2x complexity of median config MIPS Thousands z10 EC Physical Nway 32

33 Coupling Technology Impact on MIPS Sysplex configurations affect the efficiency of the HW and SW key factors workload characteristics - rate of operations to the coupling facility speed of coupling technology ( and links) versus speed of host technology example host effects 2% for light coupling workload 5-7% for medium coupling workload with speed-matched CF technology 9% for medium coupling workload with "slow" CF technology 10-14% for heavy coupling workload with speed-matched CF technology 18% for heavy coupling workload with "slow" CF technology When upgrading the host, must consider impact of CF technology on MIPS requirement 33

34 So, what have we learned about MIPS? When there is a big change in a sensitive factor - be careful move to new processor technology change in workload characteristics change in utilization change in LPAR configuration change in coupling technology But, most of the time, the items above are stable or change only a little for example, adding an engine to an existing processor And over the long run many variations tend to "even out" for example, when moving to a new technology, a below average workload this time is often an above average workload the next time 34

35 Conclusions about MIPS MIPS are fine for long term workload trending MIPS are okay for short term planning where there are only minor changes in any of the sensitive factors But whenever there is to be a major change, there is a risk of significant variation from average (MIPS) and additional analysis should be done Useful tool to help with "additional analysis" - zpcr 35

36 zpcr Capacity sizing tool available for download from or just search for "zpcr download" Through customization, zpcr can provide insight on many of the sensitive factors discussed in this presentation system design workload characteristics workload scaling LPAR configurations 36