Presentation is loading. Please wait.

Presentation is loading. Please wait.

1 Copyright © 2011, Oracle and/or its affiliates. All rights reserved.

Similar presentations


Presentation on theme: "1 Copyright © 2011, Oracle and/or its affiliates. All rights reserved."— Presentation transcript:

1 1 Copyright © 2011, Oracle and/or its affiliates. All rights reserved.

2 2 Next Generation SPARC Processor An In-Depth Technical Review Rick Hetherington, Vice President of Microelectronics Greg Grohoski, Senior Director of Microelectronics

3 3 Copyright © 2011, Oracle and/or its affiliates. All rights reserved. The following is intended to outline our general product direction. It is intended for information purposes only, and may not be incorporated into any contract. It is not a commitment to deliver any material, code, or functionality, and should not be relied upon in making purchasing decisions. The development, release, and timing of any features or functionality described for Oracle’s products remains at the sole discretion of Oracle.

4 4 Copyright © 2011, Oracle and/or its affiliates. All rights reserved. Program Agenda Attributes of Commercial Workloads Introduce T4 Overview of the T4 Processor Focus on Core S3 SPARC T4 Systems Performance

5 5 Copyright © 2011, Oracle and/or its affiliates. All rights reserved. Attributes of Commercial Workloads On-Line ProcessingBatch Processing Data Warehouse Attribute Web SPECweb09 App Serv SPECjEnt PeopleSo ft Payroll Enterprise DSS (TPC-H) Application category Web Server Server Java OLTPERP DSS Instruction-level parallelismLow Medium High Thread-level parallelismHigh Instruction/ data working set Large Medium Large Data sharingLowMediumHighMedium

6 6 Copyright © 2011, Oracle and/or its affiliates. All rights reserved. Attributes of Commercial Workloads On-Line ProcessingBatch Processing Data Warehouse Attribute Web SPECweb09 App Serv SPECjEnt PeopleSo ft Payroll Enterprise DSS (TPC-H) Application category Web Server Server Java OLTPERP DSS Low Medium High Thread-level parallelismHigh Instruction/ data working set Large Medium Large Data sharingLowMediumHighMedium

7 7 Copyright © 2011, Oracle and/or its affiliates. All rights reserved. Attributes of Commercial Workloads On-Line ProcessingBatch Processing Data Warehouse Attribute Web SPECweb09 App Serv SPECjEnt PeopleSo ft Payroll Enterprise DSS (TPC-H) Application category Web Server Server Java OLTPERP DSS Instruction-level parallelismLow Medium High Instruction/ data working set Large Medium Large Data sharingLowMediumHighMedium CMT

8 8 Copyright © 2011, Oracle and/or its affiliates. All rights reserved. Attributes of Commercial Workloads On-Line ProcessingBatch Processing Data Warehouse Attribute Web SPECweb09 App Serv SPECjEnt PeopleSo ft Payroll Enterprise DSS (TPC-H) Application category Web Server Server Java OLTPERP DSS Low Medium High Instruction/ data working set Large Medium Large Data sharingLowMediumHighMedium SPARC T4

9 9 Copyright © 2011, Oracle and/or its affiliates. All rights reserved. SPARC T4 : The Fastest-ever SPARC Processor 8 SPARC S3 cores - 8 threads each Shared 4 MB Level 3 Cache Two dual-channel DDR3-1066 memory controllers Two PCI-Express x8 2.0 ports Two 10G Ethernet ports

10 10 Copyright © 2011, Oracle and/or its affiliates. All rights reserved. S3 Core Overview 8-way threaded, dual-issue, out-of-order (128 instructions in flight) Dynamically threaded with hardware-optimized resource sharing Deep pipeline for high frequency operation (3 GHz in 40 nm) Balanced single-thread and multi-thread performance -5X better single-thread than SPARC T3 with equivalent multi-thread performance Enhanced instruction set to accelerate Oracle SW stack -PAUSE, fused compare-branch Integrated user-level cryptographic acceleration -DES/3DES, AES, Kasumi, Camellia, MD5, SHA-1, SHA-224/256/384/512, RSA, DSA, CRC32c Foundation core for future technology / product nodes

11 11 Copyright © 2011, Oracle and/or its affiliates. All rights reserved. S3 Core Block Diagram

12 12 Copyright © 2011, Oracle and/or its affiliates. All rights reserved. Threaded S3 Core Pipeline View Vertical and horizontal threading Before Pick -Only 1 thread per pipe stage Pick to Commit -Multiple threads per pipe stage -Maximize throughput Commit -Only 1 thread per pipe stage

13 13 Copyright © 2011, Oracle and/or its affiliates. All rights reserved. Single-Threaded S3 Core Pipeline View Before Pick -Thread fills all pipe stages Pick to Commit -Thread fills all pipe stages Commit -Thread fills all pipe stages Critical Thread –Core Operates in Single Thread when Solaris Scheduler IDs a ‘Critical Thread’ (response time constraints)

14 14 Copyright © 2011, Oracle and/or its affiliates. All rights reserved. Dynamic Threading Core resources are shared between the active threads -Pick-queue, working-register-file, reorder-buffer, load-buffers, store-buffers, etc. Resources can be statically or dynamically allocated Dynamic allocation enables higher throughput by seamlessly adjusting resourcing based upon thread behavior -Increased throughput especially on heterogeneous workloads (mix of high and low-IPC threads) -Applications scale better Thread hog: If a thread occupies a resource, it must release the resource in a timely fashion -If not, thread is considered a “hog” and hardware limits the resources available to it -High and low watermarks for various core resources -High watermark reached by allocation, low watermark by de-allocation -Upon reaching high watermark, thread resource allocation stalls until low watermark reached

15 15 Copyright © 2011, Oracle and/or its affiliates. All rights reserved. Relative T4 Performance 8 S1 cores 32 threads 8 S2 cores 64 threads 16 S2 cores 128 threads *All performance estimates are relative to SPARC T1 performance T1 T2 T3 Relative Performance T4 vs. T1-T3

16 16 Copyright © 2011, Oracle and/or its affiliates. All rights reserved. Relative T4 Performance 8 S1 cores 32 threads 8 S2 cores 64 threads 16 S2 cores 128 threads 8 S3 cores 64 threads *All performance estimates are relative to SPARC T1 performance T1 T2 T3 T4 Relative Performance T4 vs. T1-T3

17 17 Copyright © 2011, Oracle and/or its affiliates. All rights reserved. T4 Cryptographic Performance User-level instruction access to cryptographic accelerators –Significant benefits on small payloads Enables transparent data encryption across Oracle SW stack Relative Core Performance T3/T4

18 18 Copyright © 2011, Oracle and/or its affiliates. All rights reserved. T4 RAS Features All program-visible register and main memory state protected by ECC -Integer and floating-point register files, scratchpad / MMU registers, etc. Critical cache state also ECC protected -MOESI state array: identify potential coherence loss Other arrays are parity protected -Instruction cache data, instruction buffer Hardware and software error recovery -Hardware or software recovery from parity errors via retry -Software recovery from ECC errors via correction & retry Graceful degradation in the face of hard errors –FMA: offline thread, group of threads (core), or socket

19 19 Copyright © 2011, Oracle and/or its affiliates. All rights reserved. T4 will be delivered as a product refresh in T3 systems –Offered in 1, 2, 4 socket systems and a 1-socket SB6000 blade –T4-4 systems also used in SPARC SuperCluster Solaris Support - Solaris 10 Update 10, Solaris 11 OVM for SPARC Support - 2.1 and beyond Java Run time Optimizations – JDK6 U24; JDK7 Studio Compiler Optimizations - Studio12.3 T4 Systems and Blades T4-1T4-2 T4-4 T4-1B

20 20 Copyright © 2011, Oracle and/or its affiliates. All rights reserved. SPARC T4-1B Blade Server Compute –1x SPARC T4 8-core 2.85GHz processor –16x DDR3-1066 DIMMs –Up to 256GB memory with 16GB DIMMs I/O –2x 2.5” SFF SAS SSD/HDD –4x PCIe 2.0 x8 interfaces –2x EM, 2x NEM with 1x FEM slots –4x GbE and 2x 10 GbE ports Availability and Management –No fans or PSU on blade, leveraging integrated Blade 6000 –Oracle Integrated Lights Out Manager Service Processor 1-Socket, SB6000-Series Form Factor, Solaris 10 U10, Solaris 11, OVM for SPARC

21 21 Copyright © 2011, Oracle and/or its affiliates. All rights reserved. SPARC T4-1 Server Compute –1x SPARC T4 8-core 2.85GHz processor –16x DDR3-1066 DIMMs –Up to 256GB memory with 16GB DIMMs I/O –8x 2.5” SAS2/SATA drives –6x PCIe2 slots (4 x8,2 x16) –4 1GbE ports –10GbE XAUI ports (optional) Availability and Management –Redundant hot-plug fans and power supplies –Oracle Integrated Lights Out Manager Service Processor 1-Socket, 2RU, Solaris 10 Update 10, Solaris 11, OVM for SPARC

22 22 Copyright © 2011, Oracle and/or its affiliates. All rights reserved. SPARC T4-2 Server Compute –2x SPARC T4 8-core 2.85GHz processor –32x DDR3-1066 DIMMs –Up to 512GB memory with 16GB DIMMs I/O –6x 2.5” SAS2/SATA drives –10x PCIe2 slots (8 x8,2 x16) –4x 1GbE ports –4x 10GbE XAUI ports (optional) Availability and Management –Redundant hot-plug fans and power supplies –Oracle Integrated Lights Out Manager Service Processor 2-Socket, 3RU, Solaris 10 Update 10, Solaris 11, OVM For SPARC

23 23 Copyright © 2011, Oracle and/or its affiliates. All rights reserved. 4-Socket, 5RU, Solaris 10 Update 10, Solaris 11, OVM for SPARC Compute –4x SPARC T4 8-core 3.0GHz processor –64x DDR3-1066 DIMMs –Up to 1TB memory with 16GB DIMMs I/O –8x 2.5” SAS2/SATA drives –16x PCIe2 EMs –4x 1GbE ports –8x 10GbE XAUI ports (optional) Availability and Management –Redundant hot-plug fans and power supplies –Oracle Integrated Lights Out Manager Service Processor SPARC T4-4 Server

24 24 Copyright © 2011, Oracle and/or its affiliates. All rights reserved. SPARC T4-4 System Topology

25 25 Copyright © 2011, Oracle and/or its affiliates. All rights reserved. SPARC T4 Eight #1's for Enterprise Applications Oracle Fusion & Unlimited Apps -#1 Fusion Middleware: SPECjEnterprise -#1 E-Business HR R12.1 -#1 Peoplesoft Payroll Batch -#1 Peoplesoft HR -#1 JD Edwards Industry Applications -#1 ASAP – Telco Activation -#1 Service Broker – Telco Database -#1 Oracle Database Security TDE

26 26 Copyright © 2011, Oracle and/or its affiliates. All rights reserved. SPARC T4 Leadership - Performance at Every Tier Oracle Database -DB for Unlimited Apps & Fusion -TPC-H -TimesTen & OLAP Applications: Online & Batch -E-Business -PeopleSoft -JD Edwards -Siebel Oracle Fusion Middleware -SPECjEnterprise Virtualization Benchmarks prove SPARC T4 leads at every level in stack

27 27 Copyright © 2011, Oracle and/or its affiliates. All rights reserved. Fast Java and Oracle 11g on SPARC T4-4 beat IBM Power7 World Record SPECjEnterprise 4x SPARC T4-4 application, 2x SPARC T4-4 database tier Achieved 40,104.86 SPECjEnterprise2010 EjOPS@Standard 2.4x faster than IBM Power7 Oracle's system focus gives more than IBM's “per-core focus” 1.5x faster than Cisco x86 Application TierDB TierJOPS@standardDatabase Ver 4x T4-42x T4-440,104.86Oracle11g 9.0 2x Cisco B440 M2Cisco C460 M226,118.67Oracle11g 9.0 IBM Power7 780IBM 750E16,646.34DB2 9.7 9.0

28 28 Copyright © 2011, Oracle and/or its affiliates. All rights reserved. 1TB TPC-H T4-4: Beats Power7 & Itanium Disproves IBM’s Claims of 4:1 Core Perf Advantages T4-4 22% faster than IBM Power7 780 4.14GHz that uses twice as many processors T4-4 has better $/perf than IBM T4-4 3.6x faster on database refresh System#proc.coreQphH$/QphHAvailDatabase Sun M800016, 64209,533.6$10.139/22/11Oracle 11g Sun T4-44, 32201,487.0$4.60NOWOracle 11g IBM Power7 7808, 32164,747.0$6.853/31/11Sybase HP Superdome16, 64140,181.0$12.1510/20/10Oracle 11g

29 29 Copyright © 2011, Oracle and/or its affiliates. All rights reserved. Proven throughput, high capacity, fast batch & fast response time SPARC T4 Leads on Unlimited Applications #1 PeopleSoft HR Self-Service 9.1 – online benchmark -14,000 online users supported -T4 PeopleSoft 9.1 is 3.8x faster IBM Power + mainframe PeopleSoft 8.9 #1 PeopleSoft Payroll - batch -T4-4 Payroll 9.1/unicode is 2.8x faster than IBM z10 EC 2097 mainframe (6,512 MIPS) Payroll 9.0/non-unicode -T4-4 Payroll 9.1/unicode is 3.1x faster than eight processor HP rx7640 Itanium2 on Payroll 9.0/non-unicode #1 JD Edwards - “Day in Life” online+batch benchmark -T4-2 (9.0.2 DIL) is 50% faster than IBM P750 POWER7 (9.0.1 DIL) -10,000 users #1 Oracle E-Business - R12.1.2 X-Large -Order management X-Large 2,400 users (online)

30 30 Copyright © 2011, Oracle and/or its affiliates. All rights reserved. SPARC T4 Leads on Industry Applications #1 ASAP Activation - Telco -Each SPARC T4 processor is 7.7x faster than 1.7GHz Itanium #1 Service Broker - Telco -Each SPARC T4 processor is 2.7x faster Intel Nehalem -Java Application BRM Prepaid Billing - Telco -SPARC T4-2 is 2.2x faster than legacy x86 -Times Ten implementation -Solaris Containers enables a 4x reduction in server count Proven throughput great response time

31 31 Copyright © 2011, Oracle and/or its affiliates. All rights reserved. Dramatic improvements with no change to software stack! SPARC T4 Single-threaded Performance Leading Batch Performance -JD Edwards: 2.5x faster vs. M3000/T3 -PeopleSoft 9.1: is 2.2x faster vs. IBM z10(6,512 MIPS) Payroll 9.0 -E-Business: ~25% faster vs. IBM x86 -PeopleSoft 9.1: is ~15% faster than M5000 -E-Business: similar to M5000 Sub-second response time -Unlimited Apps, SAP-SD, Fusion MW Database refresh -Up to 3.6x faster vs. Power7 -Up to 2.4x faster vs. IBM x86 -Up to 1.9x faster vs. HP Itanium -Faster startup time -2.5x faster Java App startup over T3 PeopleSoft JD Edwards DB Refresh Startup

32 32 Copyright © 2011, Oracle and/or its affiliates. All rights reserved. Query Processing Rate SPARC T4 TDE Performance TDE (Transparent Database Encryption) Performance -SPARC T4 is 44% faster on secure queries than x86 Westmere (AES/NI) -Combination of faster query processing and TDE performance -Tests 8 different queries on 2-socket configurations -Query time is nearly identical for AES 128-bit to 256-bit on SPARC T4 SystemClearAES-CFB-128AES-CFB-192AES-CFB-256 SPARC T4-2 2.85 GHz546 s585 s586 s 99%u / 1%sys X86 3.47GHz Westmere 780 s836 s841 s842 s99%u / 1%sys % T4 Advantage43% 44%

33 33 Copyright © 2011, Oracle and/or its affiliates. All rights reserved. SPARC T4 Summary Balanced single- and multi-thread performance –Applicable to all enterprise workloads –Numerous world records Outstanding cryptographic performance –Most complete cipher acceleration in the industry –Enables transparent encryption across SW stack Excellent RAS

34 34 Copyright © 2011, Oracle and/or its affiliates. All rights reserved. Benchmark Disclosure Statement T-series Copyright 2011, Oracle and/or its affiliates. All rights reserved. Oracle and Java are registered trademarks of Oracle and/or its affiliates. Other names may be trademarks of their respective owners. Results as of 9/26/2011. SPEC and the benchmark name SPECjEnterprise are registered trademarks of the Standard Performance Evaluation Corporation. Results from www.spec.org as of 9/26/2011. SPARC T4-4, 40,104.86 SPECjEnterprise2010 EjOPS; Cisco UCS B440 M1, 17,301.86 SPECjEnterprise2010 EjOPS; IBM Power 780, 16,646.34 SPECjEnterprise2010 EjOPS.Focusing on the critical JEE server hardware & OS, the IBM result includes a JEE server with a list price of $1.30 million. The Oracle JEE servers have a list price of $0.47 million. The JEE server price versus delivered EjOPs is $77.97/EjOP for IBM versus $11.67/EjOP for Oracle. Oracle's $/perf advantage is 6.7x better than IBM ($77.97/$11.67). Pricing details for IBM, IBM p780 512GB based on public pricing at http://tpc.org/results/FDR/TPCH/TPC-H_1TB_IBM780_Sybase- FDR.pdf. Adjusted hardware costs to license all 64 cores. AIX pricing at: http://www-304.ibm.com/easyaccess3/fileserve?contentid=214347 and AIX Standard Edition V7.1 per processor (5765-G98-0017 64*2,600=$166,400). This gives application tier hardware & OS Price/perf: $77.97/EjOPS (1297956/16646.34) Pricing details for Oracle, four SPARC T4-4 512 GB, HW acquisition price from Oracle's price list: $467,856 http://www.oracle.com. This gives application tier hardware & OS Price/perf: $11.67/EjOPS (467856/40104.86) The Oracle application tier servers occupy 20U of space, 40,140.86/20=2007 EjOPS/U. The IBM application tier server occupies 16U of space, 16,646.34/16=1040 EjOPS/RU. 2007/1040=1.9x TPC-H, QphH, $/QphH are trademarks of Transaction Processing Performance Council (TPC). For more information, see www.tpc.org. SPARC T4-4 201,487 QphH@1000GB, $4.60/QphH@1000GB, avail 10/30/2011; SPARC Enterprise M8000 209,533.6 QphH@1000GB, $9.53/QphH@1000GB, avail 09/22/11; IBM Power 780 QphH@1000GB, 164,747.2 QphH@1000GB, $6.85/QphH@1000GB, avail 03/31/11; HP Integrity Superdome 2 140,181.1 QphH@1000GB, $12.15/QphH@1000GB avail 10/20/10; Dell PowerEdge R710 using EXASolution 4.0 4,253,937 QphH@1000GB, $0.10/QphH@1000GB, avail 10/01/11.

35 35 Copyright © 2011, Oracle and/or its affiliates. All rights reserved.

36 36 Copyright © 2011, Oracle and/or its affiliates. All rights reserved.


Download ppt "1 Copyright © 2011, Oracle and/or its affiliates. All rights reserved."

Similar presentations


Ads by Google