拥抱万兆以太网时代 by intel

2012/07/07,ADC@hangzhou xeon 10Gb

Transcript

1. 拥抱万兆以太网时代 - 探索英特尔® 至强® 处理器平台上 的应用性能优化 Wang, Stanley stanley.wang@intel.com Ye, Dong dong.ye@intel.com
2. Legal Disclaimers • Software and workloads used in performance tests may have been optimized for performance only on Intel microprocessors. Performance tests, such as SYSmark and MobileMark, are measured using specific computer systems, components, software, operations and functions. Any change to any of those factors may cause the results to vary. You should consult other information and performance tests to assist you in fully evaluating your contemplated purchases, including the performance of that product when combined with other products. • Intel does not control or audit the design or implementation of third party benchmarks or Web sites referenced in this document. Intel encourages all of its customers to visit the referenced Web sites or others where similar performance benchmarks are reported and confirm whether the referenced benchmarks are accurate and reflect performance of systems available for purchase. • Relative performance is calculated by assigning a baseline value of 1.0 to one benchmark result, and then dividing the actual benchmark result for the baseline platform into each of the specific benchmark results of each of the other platforms, and assigning them a relative performance number that correlates with the performance improvements reported. • SPEC, SPECint, SPECfp, SPECrate. SPECpower, SPECjAppServer, SPECjbb, SPECjvm, SPECWeb, SPECompM, SPECompL, SPEC MPI, SPECjEnterprise* are trademarks of the Standard Performance Evaluation Corporation. See http://www.spec.org for more information. TPC-C, TPC-H, TPC-E are trademarks of the Transaction Processing Council. See http://www.tpc.org for more information. • Intel® Virtualization Technology requires a computer system with an enabled Intel® processor, BIOS, virtual machine monitor (VMM) and, for some uses, certain platform software enabled for it. Functionality, performance or other benefits will vary depending on hardware and software configurations and may require a BIOS update. Software applications may not be compatible with all operating systems. Please check with your application vendor. • Hyper-Threading Technology requires a computer system with a processor supporting HT Technology and an HT Technology-enabled chipset, BIOS and operating system. Performance will vary depending on the specific hardware and software you use. For more information including details on which processors support HT Technology, see here • Intel® Turbo Boost Technology requires a Platform with a processor with Intel Turbo Boost Technology capability. Intel Turbo Boost Technology performance varies depending on hardware, software and overall system configuration. Check with your platform manufacturer on whether your system delivers Intel Turbo Boost Technology. For more information, see http://www.intel.com/technology/turboboost • No computer system can provide absolute security under all conditions. Intel® Trusted Execution Technology (Intel® TXT) requires a computer system with Intel® Virtualization Technology, an Intel TXT-enabled processor, chipset, BIOS, Authenticated Code Modules and an Intel TXT-compatible measured launched environment (MLE). Intel TXT also requires the system to contain a TPM v1.s. For more information, visit http://www.intel.com/technology/security. In addition, Intel TXT requires that the original equipment manufacturer provides TPM functionality, which requires a TPM-supported BIOS. TPM functionality must be initialized and may not be available in all countries. • Intel ® AES-NI requires a computer system with an AES-NI enabled processor, as well as non-Intel software to execute the instructions in the correct sequence. AES-NI is available on Intel® Core™ i5-600 Desktop Processor Series, Intel® Core™ i7-600 Mobile Processor Series, and Intel® Core™ i5-500 Mobile Processor Series. For availability, consult your reseller or system manufacturer. For more information, see http://software.intel.com/en- us/articles/intel-advanced-encryption-standard-instructions-aes-ni/ • Intel processor numbers are not a measure of performance. Processor numbers differentiate features within each processor series, not across different processor sequences. See http://www.intel.com/products/processor_number for details. Intel products are not intended for use in medical, life saving, life sustaining, critical control or safety systems, or in nuclear facility applications. All dates and products specified are for planning purposes only and are subject to change without notice • Intel product plans in this presentation do not constitute Intel plan of record product roadmaps. Please contact your Intel representative to obtain Intel’s current plan of record product roadmaps. • Copyright © 2011 Intel Corporation. All rights reserved. Intel, the Intel logo, Xeon and Intel Core are trademarks or registered trademarks of Intel Corporation or its subsidiaries in the United States and other countries. All dates and products specified are for planning purposes only and are subject to change without notice 2
3. Agenda • Intel E5 platform introduction • Intel E5 platform optimization – AVX • Intel E5 platform optimization – AES-NI • Intel E5 platform optimization – DDIO • 10G Market Trend and Challenge • Intel 10GBASE-T Solution • Intel Readiness in I/O Virtualization 3
4. Delivering Computing Solutions This page is not all-inclusive of Intel’s offerings Services Software Platforms Silicon Intel ® Software Network Connect with developers and Intel engineers Intel ® Software Network Connect with developers and Intel engineers Infineon Wireless Solutions* Intel® Wireless Display Intel® vPro™ Technology Intel ® Insider *Other names and brands may be claimed as the property of others.
5. Tick-Tock Development Model: Sustained Microprocessor Leadership Intel® Core™ Microarchitecture TOCK New Micro- architecture Merom 65nm TICK Penryn New Process Technology 45nm Intel® Microarchitecture Codename Nehalem TOCK New Micro- architecture Nehalem 45nm TICK Westmere 32nm New Process Technology Intel® Microarchitecture Codename Sandy Bridge TOCK Sandy Bridge 32nm New Micro- architecture TICK Ivy Bridge 22nm New Process Technology 5
6. Intel® Xeon® Processor Family for Business Mainstream Enterprise Best combination of performance, power efficiency, and cost High Performance Computing & Workstations Bandwidth-optimized for high performance analytics & visualization Small Business Economical and more dependable vs. desktop Increasing capability Cloud Computing Efficient, secure, and open platforms for Internet datacenters and IAAS Entry Servers and Workstations More features and performance than traditional desktop systems Enterprise Server Versatility for infrastructure apps (up to 4S) Scalable Enterprise Top-of-the-line performance, scalability, and reliability Cloud Computing Highest virtualization density and advanced reliability for private cloud Mission Critical Performance and reliability for the most business critical workloads with outstanding economics High Performance Computing Greater scaling and memory capacity 6
7. E5服务器平台 Intel® Xeon® Processor E5-2600/2400 Product Families Intel® Xeon® processor E5-2400/2600 prod fam Intel® Xeon® Processor E5-2400/2600 prod fam Intel® C600 series chipset QPI QPI DDR3 DDR3 DDR3 DDR3 SAS, SATA Memory DDR3 & DDR3L RDIMMs & UDIMMs, LR DIMMs Socket R: 4 channels per socket, up to 3 DPC; speeds up to DDR3 1600 Socket B2: 3 channels per socket, up to 2 DPC; speeds up to DDR3 1600 Intel® LOM Options 1GbE: Hartwell Powerville 10GbE: Twinville (10GBase-T) Niantic (SfI/KR) PCI Express Socket R: 40 lanes per socket Socket B2: 24 lanes per socket DDR3 DDR3 DDR3 DDR3 PCIe x8 PCIe x8 PCIe x8 PCIe x8 PCIe x8 Intel® C600 series chipset (Patsburg PCH) Optimized Server & WS PCH Integrated Storage: Up to 8 ports SAS RAID 5 optional Sandy Bridge CPUs Socket R: Up to 8 cores / socket Socket B2: Up to 8 cores / socket DMI2 PCIe x8 PCIe x8 PCIe x8 PCIe x8 PCIe x8 PCIe2 x4 QPI Socket R: 2 QPI links Socket B2: 1 QPI link Manageability Node Manager Data Center Manager Interface Intel® vPro™ Technology (for WS) Intel® AMT Technology (for SVR)
8. 环形总线改进多核互联 • Ring Based Interconnect • Cores, LLC, PCIe, IMC, providing 96 GB/s1 per connection • Multi Banked Last Level Cache • scales with cores • Flexible removal of Cores/LLC • BIOS controls 1At ring frequency of 3GHz IMC -Memory Controller QPI PCIe Core 0 L1 + L2 Cache L1 + L2 Cache L1 + L2 Cache L1 + L2 Cache L1 + L2 Cache L3 Cache (2.5MB) 0 L3 Cache (2.5M) 1 L3 Cache (2.5M) 2 L3 Cache (2.5MB) 3 Core 1 L1 + L2 Cache Core 2 L1 + L2 Cache Core 3 L1 + L2 Cache L3 Cache (2.5MB) 7 L3 Cache (2.5M) 6 L3 Cache (2.5M) 5 L3 Cache (2.5MB) 4 Core 7 Core 6 Core 5 Core 4 High Bandwidth, Low Latency, Scalable
9. 9 Sandy Bridge Core Micro-architecture 32k L1 Instruction Cache Scheduler Memory Control Port 0 Port 1 Port 5 Port 2 Port 3 Port 4 32k L1 Data Cache 48 bytes/cycle Allocate/Rename/Retire Zeroing Idioms Load Buffers Store Buffers Reorder Buffers Load Load Store Address Store Address Store Data L2 Data Cache (MLC) Fill Buffers Pre decode Instruction Queue Decoders 1.5k uOP cache Decoders Decoders Decoders Branch Pred In order Out-of- order ALU VI ADD VI Shuffle AVX FP ADD ALU JMP AVX/FP Shuf AVX/FP Bool Imm Blend ALU VI MUL VI Shuffle DIV Imm Blend AVX FP MUL Sandy Bridge - Intel® Next Generation Microarchitecture AVX= Intel® Advanced Vector Extensions (Intel® AVX)
10. Intel® Xeon Advanced Technologies Different Technologies Optimized for Different Needs 10 Security Resilience Service Delivery Power Data Protection Advanced Encryption Standard—New Instructions Platform Security Trusted Execution Technology System Power Intel® Intelligent Power Node Manager Integrated Power Gates and Automated Low Power States Data Center Power Intel® Data Center Manager Processor Power Virtualization Near-Native VM Perf. & Seamless VM Migration Intel® VT (CPU, Chipset, and I/O) Intel® TXT Intel® AES-NI High Availability Mission Critical Class Reliability, Availability, and Serviceability Advanced RAS Performance Intel® Hyper- Threading Intel® Turbo Boost Automatically Adapt to the Workload Intel® VT-x, Intel® VT-d and Intel® VT-c
11. Agenda • Intel E5 platform introduction • Intel E5 platform optimization – AVX • Intel E5 platform optimization – AES-NI • Intel E5 platform optimization – DDIO • 10G Market Trend and Challenge • Intel 10GBASE-T Solution • Intel Readiness in I/O Virtualization 11
12. 12 Intel® 高级矢量扩展 (Intel® AVX) 关键特性 • 更宽的矢量 – 从128位增加到256位 – 两个128位load端口 关键特性 益处 • 最高达2倍的峰值FLOPs (每秒浮点操作 次数) 输出，具有更好的能效 • 增强的数据重排 – 使用新的256位元函数来广播、带掩码 load和排列数据 • 更快更有效率地组织，存取和提取必要 的数据 • 3个和4个操作数 • 对AVX 128和AVX 256使用非破 坏性的句法 • 更少的寄存器拷贝，对矢量和标量代码能 更好地使用寄存器 • 灵活的不对齐内存存取支持 • 更多机会进行fuse load和计算操作 • 可扩展的新操作码 (VEX) • 减小了代码的大小 Intel® AVX 是一种通用的的架构, 预期用来代替当今所有应用中的 SSE指令
13. AVX Performance 26 FP intensive workloads studied on our cycle-accurate performance simulators. AVX currently yields 1.6X average gains when compared to the same processor without AVX (this is hand tuned best vs. best in all cases). GSSE vs. GSR baseline - single threaded, 5/07 POR instructions (eas 0.94) 1.00 1.50 2.00 2.50 Alpha Blend (O paque/Trans opt.) SP N on-IEEE division SP N on-IEEE sqrt long IIR Black Scholes options pricing Alpha transparent Vert Dow nsam ple N oise generation Audio Encode D P VM L EXP() throughput collision physics short IIR D P VM L PO W () throughput Sphere-sphere collision physics C om plex SP FFT - 256 PT C om plex SP FFT - 1024 PT m any-tap FIR H orizontal Audio characterization Sandra FP M M benchm ark D G EM /M M M (D P) Binom ial Tree O ptions pricing Point transform Im aging H oriz D ow nsam ple threshold im age filtering Im age resam pling G SSE scaling Geomean 1.6X FP Heavy Workloads average 1.6X
14. 在线图片缩放的优化 - AVX 基于AVX 指令集优化的图像缩放应用，性能提升4倍以上
15. Agenda • Intel E5 platform introduction • Intel E5 platform optimization – AVX • Intel E5 platform optimization – AES-NI • Intel E5 platform optimization – DDIO • 10G Market Trend and Challenge • Intel 10GBASE-T Solution • Intel Readiness in I/O Virtualization 15
16. Intel® AES-NI: What is it? • Processor assistance for performing AES encryption − 7 new instructions • Makes enabled encryption software faster and stronger Meets Needs of Rapidly-Growing Data Encryption Workloads
17. Data Protection with Intel® AES-NI Internet Intranet Data in Motion Data at Rest Data in process Name: J.Doe SS#  Secure transactions used pervasively in ecommerce, banking, etc. Full disk encryption software protects data automatically during saving to disk Most enterprise and many cloud applications offer options to use encryption to secure information and protect confidentiality Efficient Ways to Use Encryption for Data Protection 1 2 3 1 2 3
18. OpenSSL 1.0.1 Improvements Hashing Cipher Cipher/Hash Stitching Public Key AES 128/256 CBC RC4 RSA-1024 RC4-HMAC-MD5 AES-128-CBC-HMAC-SHA1 AES-256-CBC-HMAC-SHA1 SHA-1 Released March 2012 - Added AES-NI, PCLMULQDQ, RDRAND support - Added SSSE3 support (SHA1, vector-permutation AES);
19. AES-NI指令对网络安全加解密应用性能的提升 英特尔® AES-NI 指令是一套 新的指令集 •始于英特尔®至强®5600系列 •使AES加密解密操作更加高效安全 •六条指令在硬件设计上实现对AES 操作的支持 –四条指令针对AES加解密操作 –两条指令用于支持密钥的扩展 在Sandybridge平台上测试 • 单线程模式 单核 6 倍性 能提升 • 多线程模式8-10 倍性能 提升
20. OpenSSL Performance comparison SSL Algorithm Thread versio n WSM E5645 SNB E5-2680 Speedup sign (s) verify (s) sign/s verify/s sign (s) verify (s) sign/s verify/s sign(s) verify( s) sign/s verify/ s rsa 512 bits 1 1.0.0 0.0001 05 0.0000 09 9528.4 115869. 6 0.0000 55 0.0000 05 18231.3 206348. 4 1.9090 91 1.8 1.9133 64 1.7808 67 1.0.1- beta2 0.0001 07 0.0000 09 9376.5 114819 0.0000 53 0.0000 04 19040.9 240098. 1 2.0188 7 2.25 2.0307 2.0911 24-32 1.0.0 0.0000 08 0.0000 01 122898. 3 1437500 0.0000 03 0 302261. 3 3200000 2.6666 67 2.4594 42 2.2260 87 1.0.1- beta2 0.0000 08 0.0000 01 122581. 8 1426470 .6 0.0000 03 0 310703. 1 3573845 .6 2.6666 7 2.5346 6 2.5053 8 rsa 1024 bits 1 1.0.0 0.0005 2 0.0000 27 1922 37256.7 0.0002 68 0.0000 14 3730.8 69045.8 1.9402 99 1.9285 71 1.9411 03 1.8532 45 1.0.1- beta2 0.0005 19 0.0000 27 1926.3 37331.2 0.0001 77 0.0000 12 5681.2 85987.2 2.9322 2.25 2.9492 8 2.3033 6 24-32 1.0.0 0.0000 4 0.0000 02 24707.5 461709. 2 0.0000 17 0.0000 01 59695.5 1101149 .4 2.3529 41 2 2.4160 88 2.3849 41 1.0.1- beta2 0.0000 4 0.0000 02 24735.1 459263. 4 0.0000 11 0.0000 01 87800 1344538 .3 3.6363 6 2 3.5496 1 2.9276 rsa 2048 bits 1 1.0.0 0.0032 94 0.0000 95 303.6 10478.3 0.0017 2 0.0000 51 581.4 19435.7 1.9151 16 1.8627 45 1.9150 2 1.8548 52 1.0.1- beta2 0.0033 0.0000 95 303 10472.2 0.0012 72 0.0000 39 786.4 25358.7 2.5943 4 2.4359 2.5953 8 2.4215 3 24-32 1.0.0 0.0002 62 0.0000 08 3814.1 128350. 1 0.0001 09 0.0000 03 9105.5 308187. 7 2.4036 7 2.6666 67 2.3873 26 2.4011 49 1.0.1- beta2 0.0002 62 0.0000 08 3810.5 127985 0.0000 81 0.0000 02 12447 412087. 9 3.2345 7 4 3.2665 3.2198 1 rsa 4096 bits 1 1.0.0 0.0232 79 0.0003 63 43 2752.7 0.0124 38 0.0002 04 80.4 4893 1.8716 03 1.7794 12 1.8697 67 1.7775 28 1.0.1- beta2 0.0232 79 0.0003 63 43 2753.6 0.0091 83 0.0001 48 108.7 6774.6 2.5350 1 2.4527 2.5279 1 2.4602 7 24-32 1.0.0 0.0019 02 0.0000 3 525.8 33352.4 0.0007 87 0.0000 13 1271.1 79305.1 2.4167 73 2.3076 92 2.4174 59 2.3777 93 1.0.1- beta2 0.0019 07 0.0000 3 524.3 33300.2 0.0005 76 0.0000 09 1747.2 110493. 3 3.3107 6 3.3333 3 3.3324 4 3.3181
21. OpenSSL Performance comparison - continued SSL Algorithm Thread version SNB E5-4650 sign (s) verify (s) sign/s verify/s RSA 2048 bits 1 0.9.8e 0.002522 0.000074 396.5 13589.2 0.002516 0.000073 397.5 13608.9 0.002513 0.000074 397.9 13553 64 0.000071 0.000002 14051.5 511402.9 0.000071 0.000002 14166.9 515233.1 0.000071 0.000002 14165.6 515238.3 1 1.0.1c 0.001348 0.000042 741.8 23848.6 0.001347 0.000042 742.6 23826.9 0.001348 0.000042 741.8 23848.4 64 0.000043 0.000001 23254.2 766224.6 0.000043 0.000001 23210.4 763986 0.000043 0.000001 23212.9 763689.2
22. CDN Web Server gzip 模块的优化 22 2.732 1.004 1.25124 2.769 1.357 1.002 1.25059 2.105 1.237 1.046 1.25203 0 0.5 1 1.5 2 2.5 3 SNB vs WSM SMT icc ipp web server gzip module speedup gzip level 6 gzip level 3 gzip level 1 以 100KB 左右的网页文件为例，在1/3/6 3种gzip的压缩级别下： • SMT: contributes the perf by 1.3X • icc: provide 1-4% speedup for gzip workload. • IPP: IPP gzip api can contribute further 25% performance improvement. • Sandy Bridge E5-2680 比 Westmere 5645 性能提高了 2 倍以上. • Sandy Bridge E5-2680 gzip 压缩的 QPS可以达到 8000 -12000。
23. Agenda • Intel E5 platform introduction • Intel E5 platform optimization – AVX • Intel E5 platform optimization – AES-NI • Intel E5 platform optimization – DDIO • 10G Market Trend and Challenge • Intel 10GBASE-T Solution • Intel Readiness in I/O Virtualization 23
24. 英特尔® Data Direct I/O 技术 (英特尔® DDIO) 英特尔® 至强®新架构 支持更高I/O性能及更低功耗 英特尔最新系统级I/O性能创新 允许英特尔以太网控制器直接访问处理器L3缓存，I/O数据的目的和源为L3缓存而不是主存 Performance tests and ratings are measured using specific computer systems and/or components and reflect the approximate performance of Intel products as measured by those tests. Any difference in system hardware or software design or configuration may affect actual performance. Buyers should consult other sources of information to evaluate the performance of systems or components they are considering purchasing. For more information on performance tests and on the performance of Intel products, visit http://www.intel.com/performance/resources/benchmark_limitations.htm. Intel, the Intel logo, Xeon and Xeon Inside are trademarks of Intel Corporation in the U.S. and/or other countries. Intel® Ethernet Controllers
25. 英特尔®E5-2600新平台I/O创新 英特尔® Data Direct I/O 技术支持快速存取数据 英特尔® 至强® 处理器 E5-2600 英特尔® 至强® 处理器 5500 核心 1 核心 1 核心 2 核心 2 核心 3 核心 3 核心 4 核心 4 核心 5 核心 5 核心 6 核心 6 英特尔® 至强 处理器 5500 缓存 缓存 核心 1 核心 1 核心 2 核心 2 核心 3 核心 3 核心 4 核心 4 核心 5 核心 5 核心 6 核心 6 英特尔® 至强 处理器 5500 缓存 缓存 内存 内存 内存 内存 核心 1 核心 1 核心 2 核心 2 核心 3 核心 3 核心 4 核心 4 核心 5 核心 5 核心 6 核心 6 核心 7 核心 7 核心 8 核心 8 英特尔® 至强 E5-2600 缓存 缓存 核心 1 核心 1 核心 2 核心 2 核心 3 核心 3 核心 4 核心 4 核心 5 核心 5 核心 6 核心 6 核心 7 核心 7 核心 8 核心 8 英特尔® 至强 处理器 E5-2600 缓存 缓存 PCI Express Gen 2 x4 高达36条 PCIe* Gen 2 lanes 英特尔® 5500 系列 (IOH) 英特尔® ICH 10 高达80条 PCIe* Gen 3 lanes 英特尔 至强 处理器 5500 • 集成内存控制器 • 英特尔® QuickPath 接口(英特尔® QPI) 英特尔 至强 处理器 E5-2600 • 英特尔® 集成I/O • 每个CPU插槽支持多达40条PCIe lane • 英特尔® Data Direct I/O 技术 PCI Express Gen 2 x4 约50Gbps 吞吐量 大于200Gbps 吞吐量 高达6.4 GT/s 高达8.0 GT/s 高达 20 MB 缓存 英特尔® C600 系列芯片组
26. Intel® QPI 1 Intel® QPI 2 Intel® QPI 1 Intel® QPI 2 核心 1 核心 1 核心 2 核心 2 核心 3 核心 3 核心 4 核心 4 核心 5 核心 5 核心 6 核心 6 核心 7 核心 7 核心 8 核心 8 英特尔® 至强® 处理器E5-2600 缓存 缓存 核心 1 核心 1 核心 2 核心 2 核心 3 核心 3 核心 4 核心 4 核心 5 核心 5 核心 6 核心 6 核心 7 核心 7 核心 8 核心 8 英特尔® 至强® 处理器 E5-2600 缓存 缓存 英特尔l® 以太网 控制器 PCI Express 不支持英特尔DDIO 英特尔® 以太网 控制器 PCI Express* 支持英特尔DDIO I/O 设备请求读I/O 数据 CPU核创建缓冲区供I/O 设置读取. 分配新缓存 引起内存读. 发送报文 源自IO设备的DMA读操作 引起对内存或对L3缓存 的猜测读(在上一代CPU 上) 发送报文 1 2 3 CPU核创建缓冲区供I/O设 备读取，将数据放入缓存( 缓存行已分配) 1 I/O设备请求读I/O数据 2 数据送至IO设备 4 数据送至IO设备 3 无内存访问 X 缓存行被移出缓存区引起写回操作 5 DDIO工作原理 英特尔® Data Direct I/O 技术(英特尔® DDIO)– 发送流程 (Tx) Intel® QuickPath Interconnect (Intel® QPI)
27. Intel® QPI 1 Intel® QPI 2 Intel® QPI 1 Intel® QPI 2 核心 1 核心 1 核心 2 核心 2 核心 3 核心 3 核心 4 核心 4 核心 5 核心 5 核心 6 核心 6 核心 7 核心 7 核心 8 核心 8 英特尔® 至强® 处理器 E5-2600 缓存 缓存 核心 1 核心 1 核心 2 核心 2 核心 3 核心 3 核心 4 核心 4 核心 5 核心 5 核心 6 核心 6 核心 7 核心 7 核心 8 核心 8 英特尔® 至强® 处理器 E5-2600 缓存 缓存 IO缓存 IO缓存 英特尔® 以太网 控制器 PCI Express* 不支持英特尔DDIO 英特尔® 以太网 控制器 PCI Express 支持英特尔DDIO DMA写数据到内存 缓存L3数据送至CPU核 CPU核读取数据 DMA直接将数据写入 分配给IO的L3缓存 接收报文 接收报文 1 2 4 无内存访问 CPU核读取数据 2 1 L3缓存数据送至CPU核 3 缓存不命中引起内存读 3 DDIO工作原理 英特尔® Data Direct I/O 技术(英特尔® DDIO)– 接收流程 (Rx) Intel® QuickPath Interconnect (Intel® QPI)
28. 英特尔® Data Direct I/O 技术 (英特尔® DDIO) 英特尔® 至强®新架构支持更高I/O性能 Test Configuration Netperf version 2.4.5 RedHat Enterprise Linux 6.2 Buffer Sizes= 8KB Interrupts Affinity – Socket Process Affinity – Socket Services Disabled - LLDPAD, IPTABLES, IP6TABLES, SELINUX, IRQBALANCE Network Configuration Cisco* Nexus* 5020 Clients connected @ 10000Mbps Software and workloads used in performance tests may have been optimized for performance only on Intel microprocessors. Performance tests, such as SYSmark* and MobileMark* are measured using specific computer systems, components, software, operations and functions. Any change to any of those factors may cause the results to vary. You should consult other information and performance tests to assist you in fully evaluating your contemplated purchases, including the performance of that product when combined with other products. Source: LAD Marketing - February 2012 SUT: GreenCity SDV Intel® Xeon® Processor X5680 (8M Cache, 3.33 GHz, 6.40 GT/s Intel® QPI) 12GB DDR3 RAM @ 1333MHz Intel® 5520 Series Chipset SUT: CanoePass SDV Intel® Xeon® Processor E5-2680 (8M Cache, 2.7 GHz, 8.0 GT/s Intel® QPI) 48GB DDR3 RAM @ 1333MHz Intel® C600 Series Chipset Clients Intel Server System SR1600UR 2 Intel® Xeon® Processors X5570 (8M Cache, 2.93GHz, 6.40 GT/s Intel® QPI) 8 GB RAM Intel® Ethernet Converged Network Adapter X520 RedHat Linux* 6.1 0 Gbps 50 Gbps 100 Gbps 150 Gbps 200 Gbps 250 Gbps Intel® Xeon® Processor 5600 Intel® Xeon® Processor E5-2600 Intel® Xeon® Processor E5-2600 16*10Gb Ports 8*10Gb Ports Red Hat* 6.2 – 双向性能 双倍I/O性能 双倍端口数量 Intel® Ethernet Controllers 8*10Gb Ports
29. 支持金融证券业务更高的交易率 英特尔® Data Direct I/O 技术 (英特尔® DDIO) 减小I/O延迟 测试配置: Netperf Benchmark, 1B UDP Payloads Rose City CRB, 8x2GB, DDR3-1333MHz Intel® Xeon® Processor E5-2680 (8C, B0, 2.8GHz (2.7GHz + turbo)) Linux* 2.6.36.2 Intel® Ethernet Converged Network Adapters X520-DA2 ixgbe-2.0.84.9P+LL-tx+tx-Fdir (Low-Latency PoC Driver) 来源: 英特尔以太网实验室, 2011年4月 12.1 µs 11.3 µs 10.5 µs 11.0 µs 11.5 µs 12.0 µs 12.5 µs Intel® Xeon Processor E5-2600 Intel DDIO Disabled Intel® Xeon Processor E5-2600 Intel DDIO Enabled UDP 往返延迟 延迟减小10-15% Intel® Ethernet Controllers Software and workloads used in performance tests may have been optimized for performance only on Intel microprocessors. Performance tests, such as SYSmark* and MobileMark* are measured using specific computer systems, components, software, operations and functions. Any change to any of those factors may cause the results to vary. You should consult other information and performance tests to assist you in fully evaluating your contemplated purchases, including the performance of that product when combined with other products.
30. L3 转发内存带宽/功耗 基于英特尔® Data Direct I/O 技术的英特尔® 数据面开发工具DPDK 测试配置: L3 Forwarding Benchmark DPDK rte-20110211 Rose City CRB, 8x4GB DDR3-1333MHz Intel® Xeon® Processor E5-26xx (8C B0, 2.0GHz) Linux* 2.6.33.6 4x Intel® Ethernet Converged Network Adapters X520-SR2 来源: 英特尔 ECG 实验室, 2011年2月 0 MB/s 10000 MB/s 20000 MB/s 30000 MB/s 64 Bytes 512 Bytes 1518 Bytes 总的内存带宽利用率 节省高达65W 英特尔 DDIO 关闭 英特尔 DDIO 使能 150 Watts 225 Watts 300 Watts 64 Bytes 512 Bytes 1518 Bytes 交流电源 减小占用内存带宽 报文大小 报文大小 英特尔® Data Direct I/O 技术 (英特尔® DDIO) 英特尔® 数据面开发工具 (英特尔® DPDK) Software and workloads used in performance tests may have been optimized for performance only on Intel microprocessors. Performance tests, such as SYSmark* and MobileMark* are measured using specific computer systems, components, software, operations and functions. Any change to any of those factors may cause the results to vary. You should consult other information and performance tests to assist you in fully evaluating your contemplated purchases, including the performance of that product when combined with other products.
31. Agenda • Intel E5 platform introduction • Intel E5 platform optimization – AVX • Intel E5 platform optimization – AES-NI • Intel E5 platform optimization – DDIO • 10G Market Trend and Challenge • Intel 10GBASE-T Solution • Intel Readiness in I/O Virtualization 31
32. 10GbE Market Forecast Source: Dell’Oro Group, Q4 ‘11 Update 10GbE NIC & LOM Ports Forecast • 5M 10GbE NIC/LOM ports shipped in 2011, expected to double to 10.8M in 2012 • The 10GbE NIC/LOM Forecasting target of 6.3M was not met for 2011, delta of -1.3M • By 2016, >70% of 10GbE ports will be stand alone LOM
33. Total Server-class Adapter and LOM Market Ports: Speeds Ports in Millions 0 9 18 27 36 45 54 2008A 2009A 2010A 2011E 2012E 2013E 2014E 2015E 4 Gbps FC 8 Gbps FC 1 GbE 10 GbE 40 GbE 16 Gbps FC Infiniband ©2011 Crehan Research Inc. 33
34. 10 GbE Server-class Adapter & LOM Port Interfaces – Ethernet Port Shipments in Millions 0 6 12 18 24 2008A 2009A 2010A 2011E 2012E 2013E 2014E 2015E KR (Internal Blad Other 10G BASE-T SFP+ ©2011 Crehan Research Inc. 34
35. 10GbE Cabling for Rack & Tower 35 Intel Confidential 10GBASE-SR Optical SFP+ Direct Attach 10GBASE-T Connector Type LC Duplex SFP+ RJ-45 Cable Type Multimode Fiber Twin-Ax Copper CAT 6/6A twisted pair Max Line Length 300m ~7m 100m Max PHY Power <2W/port (65nm) <1W/port (65nm) 3.5W/port (40nm) Energy Efficient Modes N/A N/A Short reach (~2.0W) 802.3az EEE (<1W idle) PHY Latency 100s of ns 100s of ns ~2us Compatibility 10GBASE-SR – regardless of module FF New technology Backward compatibility 100/1000BASE-T Data Center Deployment Full coverage Top of Rack only Top of Rack, Middle of Row, or End of Row
36. Agenda • Intel E5 platform introduction • Intel E5 platform optimization – AVX • Intel E5 platform optimization – AES-NI • Intel E5 platform optimization – DDIO • 10G Market Trend and Challenge • Intel 10GBASE-T Solution • Intel Readiness in I/O Virtualization 36
37. 10GBASE-T技术的最大挑战? 1000BASE-T在系统中具有更高的余量, 大多数5E类线质量好于规范要求, 因此大多 能够支持140-150m+距离 10GBASE-T 接近仙农(Shannon)信道容量, 需要成熟的PHY和很好的布线才能支持 100m网线并有余量 频率范围: • 1000BASE-T = 125M symbols/sec with Most energy <=80MHz • 10GBASE-T = 800M symbols/sec with Most energy <=500MHz 大多数噪声集中在500MHz附近: 在1000BASE-T频带外的噪声，落在10GBASE-T的频带内
38. • Similar to 1000BASE-T − Can be thought of as extensions of 1000BASE-T over a wider frequency range • Significantly greater NEXT + FEXT + Equalization challenge • Note how received the received signal is reconstructed from this “really, really, really” closed-eye system by applying successive DSP-based techniques 10GBASE-T技术的最大挑战?
39. • Single-die 10GBASE-T dual port MAC/PHY − Triple-speed 10G/1000/100BASE-T − Low power (10W Max) – enables passive heatsink − Short-reach-mode enabled for lower power at <30 meters − Package: 25mm (Powerville Dual Port Compatible) − Base features from Niantic • Adaptive Power Management − Automatic power management modes based on cable length – Data Center, Short Reach, and Full Reach • Optimized for Virtualization − 128 transmit and 128 receive queues per port − VMDq support (64 VMs) − PCI-SIG I/O Virtualization (64 Virtual function/port) • Unified networking − DCB: Priority grouping, Priority flow control, DCBx, QCN − iSCSI acceleration and remote boot • Security − IPSec & MACSec authentication/encryption • Manageability − BMC Pass-Through − Interfaces: NC-SI, SMB Path rough & MCTP over SMB − Host to BMC & BMC to Network pass-through − Inventory, monitoring and device configuration − IEEE 1588/802.1as timestamping PCIe with SR-IOV Traffic & Queue Mgmt IPMI Pass-Through PCI-E Gen2 (5GT/s) X8 Q1 Q2 Qm FCoE, iSCSI, TCP Acceleration 10GbE MAC 10G/1000/100BASE-T PHY IPSec/MACSec … NC-SI/SMBus Traffic & Queue Mgmt Q0 Q2 Qn FCoE, iSCSI, TCP Acceleration 10GbE MAC 10G/1000/100BASE-T PHY IPSec/MACSec To RJ-45 Port 0 … To RJ-45 Port 1 Virtual Ethernet Bridge 39 英特尔双端口 10GBASE-T 以太网控制器: X540 基于英特尔® 至强® 处理器E5-2600服务器
40. 40 10G BaseT PHY RJ45 10G BaseT PHY RJ45 RJ45 RJ45 XAUI ~$50/port discrete PHY + Additional Memory and Clocks 2x 10GBaseT Integrated 40nm MAC + PHY 25Mhz crystal Intel® 82599 + Ext 10GBase-T PHYs Intel® X540 Integrated 10GBase-T Making 10GbE cost effective for NIC, Mezz, and LOM designs NVRAM NVRAM NVRAM Clock NVRAM Clock NVRAM 25Mhz crystal Twinville (X540) 82599EB 英特尔双端口 10GBASE-T 以太网控制器: X540 基于英特尔® 至强® 处理器E5-2600服务器
41. 10GBASE-T 板载以太网 (LOM) 标准 PCI Express* 10GBASE-T 融合网络 适配器 定制化PCI Express* 10GBASE-T 融合 网络适配器 新一代10GBASE-T 控制器: • 低成本, 低功耗, 赋予整个数据中心万兆以太网(10GbE)性能 • 后向兼容1000BASE-T以太网，简化千兆到万兆迁移 • 统一网络支持LAN, iSCSI, 以及FCoE 英特尔双端口 10GBASE-T 以太网控制器: X540 基于英特尔® 至强® 处理器E5-2600服务器
42. 确保有足够的PCI Express* 带宽 在VMkernel.log文件中grep “Express” 以验证正确的PCI Express* 速度与带宽 双端口10Gb 全线速插槽要求至少 x8 Lanes PCI Express* 5.0GT/s (Gen 2) ~ # grep Express /var/log/VMkernel.log |more <6>ixgbe: 0000:06:00.0: ixgbe_probe: (PCI Express:5.0GT/s :Width x8) 00:1b:21:69:9e:30 <6>ixgbe: 0000:06:00.1: ixgbe_probe: (PCI Express:5.0GT/s :Width x8) 00:1b:21:69:9e:31 双端口 PCI Express* 10Gb 融合网络适配器 Test Configuration: Ixia* IxChariot* v7.1; 16 Clients Per Port Under Test; High Performance Throughput Script; File Size = 64-1K: 1,000,000 / 2K+:10,000,000 Bytes; Buffer Sizes=64 Bytes to 64 KB; Data Type – Zeroes; Data Verification Disabled; Nagles Disabled System Under Test: Intel® S5520HC (“Hanlan Creek”); two Intel® Xeon® Processors X5680 (12M Cache, 3.33 GHz, 6.40 GT/s Intel® QPI); Intel® 5520 Chipset (“Tylersburg”); RAM:12GB DDR3 @ 1333MHz; BIOS: 0050; Windows Server* 2008 R2 x64 Clients: SuperMicro* 6015T-TV; two Intel® Dual Core Xeon® processors 5160 @3.0GHz; 2 GB RAM; Intel® PRO/1000 PT Dual Port Server Adapter - v9.12.13.0 driver; Windows Server 2003 SP2 x64 Network Configuration: Force10 Networks* ExaScale* E1200i switch; Clients connected @ 1 Gbps Software and workloads used in performance tests may have been optimized for performance only on Intel microprocessors. Performance tests, such as SYSmark* and MobileMark* are measured using specific computer systems, components, software, operations and functions. Any change to any of those factors may cause the results to vary. You should consult other information and performance tests to assist you in fully evaluating your contemplated purchases, including the performance of that product when combined with other products. LAD Performance Lab 2010
43. X-Lab 实现万兆规范一致性,性能以及互操作性的测试 位于英特尔俄勒岗, 投资$5M 建有业界领先的10GBASE-T测试实验室 − 5x 10Gb 测试台 −自协商, 收发规范一致性, 接收性能 − 6类和6A类线缆设备 −4连接器和来自6个布线源的6-around-1 通道配置 −10 英里布线 − 遵从UNH IOL测试方法
44. Agenda • Intel E5 platform introduction • Intel E5 platform optimization – AVX • Intel E5 platform optimization – AES-NI • Intel E5 platform optimization – DDIO • 10G Market Trend and Challenge • Intel 10GBASE-T Solution • Intel Readiness in I/O Virtualization 44
45. 操作系统 基于虚拟机管理器的端口划分 英特尔®以太网控制器提供英特尔® VMDq技术 1. Intel® VT for Connectivity (Intel® VT-c); Intel® Ethernet Provides Virtual Machine Device Queues (VMDq) 单个万兆以太网连接 多个万兆虚拟队列 Rx Tx Rx Tx Rx Tx 物理网卡 功能 PCI Express 自动的端口分配和业务分流 动态带宽和服务质量控制 细粒度共享的虚拟机业务策略 更高性能和更低CPU利用率 单个万兆端口提供多个发/收队列 标准的万兆端口在虚拟机管理器里呈现为单个端口 网络队列（NetQueue） Virtual Machine Queues
46. VMware* vSphere* 5.0 上的万兆以太网 VMware 存储 I/O 控制 VMware vSphere 5.0 的 存储 I/O 控制加英特尔®万兆以太网给基于以太网技术的 iSCSI和FCoE云存储架构提供了非常划算的服务质量解决方案。 云构建者的参考架构库 www.intel.com/cloudbuilders 系统自带的FCoE支持 VMware* vSphere* 5.0 引入软件FCoE适配器，它不需要依赖特定的FCoE硬件卡就 能提供FCoE功能，因此能创建更灵活的网络基础架构。软件FCoE适配器和ESXi自带的 FCoE协议栈一起实现协议的处理，英特尔的网络融合网卡X520或X540提供了数据面的 处理，如数据中心桥接（DCB)和智能卸载. 系统自带的iSCSI支持 软件iSCSI适配器是包在Vmkernel里的Vmware代码。它允许你的主机通过英特尔以太 网卡连到iSCSI设备。软件的iSCSI适配器处理iSCSI协议，网卡处理底层通信。采用英 特尔Nehalem处理器的系统可以实现iSCSI digest的卸载处理，因此能减少对性能的影 响。 VMware vSphere 网络 I/O 控制 网络业务管理提供了统一网络环境中不同业务类型所要求的控制和保证。网络管理员可 以根据要求，通过网卡对不同的业务类型进行I/O的共享和限制。 性能和服务质量 NetQueue 和VMDq一起协作实现高业务量VM和以太网控制器硬件队列间的动态分配。 这主要是利用了控制器的网络业务分流和收/发循环调度功能，实现平衡的带宽分配并提 升性能。
47. 更高吞吐量 & 更低CPU占用率 英特尔®万兆以太网融合网络适配器X520 在VMware* ESXi 5.0下的性能测试 基于英特尔® 至强® 处理器 E5-2600服务器 与基于英特尔®至强®处理器5500服务器比较 英特尔® 至强处理器 E5-2600 英特尔至强处理器 5500 Test Configuration Ixia* IxChariot* v7.1 16 Clients Per Port Under Test (4 clients/VM; 4 VMs/port) High Perf. Throughput Script File Size = 64-1K: 1,000,000 / 2K+:10,000,000 Bytes Buffer Sizes=64 Bytes to 64 KB Data Type: Zeroes, Data Verification Disabled, Nagles Disabled Clients: SuperMicro* 6015T 2 Intel Xeon Processors 5160 @3.0GHz 2 GB RAM Intel® PRO/1000 PT Dual Port Server Adapter Windows* Server 2003 SP2 x64 英特尔®集成 I/O, 访问数据更快速 SUT: Dell R710 Intel® Xeon® Processor X5570 (8M Cache, 2.93 GHz, 6.40 GT/s Intel® QPI) 48GB DDR3 @ 1333 MHz Intel® 5520 Chipset 局域网接入部门软件性能实验室 2012年2月 IxChariot* 性能数据 Network Configuration Force10 Networks* ExaScale* E1200i switch Clients connected @1Gbps SUT: Intel® “Canoe Pass” 2 Intel Xeon Processors E5-2600 (4C/8T, 20M Cache, 2.9 GHz) Intel® C600 Series Chipset RAM: 16GB DDR3 @1333MHz Software and workloads used in performance tests may have been optimized for performance only on Intel microprocessors. Performance tests, such as SYSmark* and MobileMark* are measured using specific computer systems, components, software, operations and functions. Any change to any of those factors may cause the results to vary. You should consult other information and performance tests to assist you in fully evaluating your contemplated purchases, including the performance of that product when combined with other products.
48. 操作系统 基于PCI-SIG* SR-IOV 标准的端口划分技术 采用基于 SR-IOV 的灵活端口划分技术实现服务质量和业务控制 操作系统 Intel® Ethernet with VT-c Support PCI Express Rx Tx Rx Tx Rx Tx 物理网卡 功能 Rx Tx Rx Tx Rx Tx 虚拟网卡 功能 Rx Tx Rx Tx Rx Tx 虚拟网卡 功能 Rx Tx Rx Tx Rx Tx 虚拟网卡 功能 Rx Tx Rx Tx Rx Tx 虚拟网卡 功能 单个万兆连接 多个万兆虚拟连接 Rx Tx Rx Tx Rx Tx 物理网卡 功能 PCI Express 控制器内业务和队列管理实现服务质量 S 虚拟以太网桥和业务分流器 (L2 交换) 灵活的有弹性的端口分配 动态的带宽和 服务质量控制 带宽共享和业务隔离 更高的性能和更低的CPU利用率 1个万兆以太网端口识别为多个万兆 PCIe网络设备 标准的万兆以太网端口在OS里呈现为单端口 Coming in 2012
49. Linux 操作系统下部署 PCI-SIG* SR-IOV •采用基于标准的I/O虚拟化技术给虚拟机和Linux内核主 机处理提供保证带宽 灵活的有弹性的端口分配 •通过循环的收/发调度以及对每个虚拟功能发送速度的控 制实现自动化的发/收负载均衡 动态的带宽和服务质量控制 •把Linux内核处理直接分配给虚拟功能以实现业务隔离 带宽共享和业务隔离 •虚拟机跳过了虚拟机管理器直接控制了虚拟网卡功能 更高的性能 Linux* 主机操作系统 带KVM虚拟化的Linux 主机操作系统
50. Networks of Virtulization - Legacy SERVER 2 192.168.100.98 SERVER 1 X520 ‐ SR2 X520 ‐ SR2 VM01 192.168.100.51 1.0.100.51 … 192.168.100.99 Dom 0 Por t1 Hyper-V soft switch Port0 PF VF VF VF VF VM02 192.168.100.52 1.0.100.52 VM03 192.168.100.53 1.0.100.53 VM08 192.168.100.58 1.0.100.58 VM001 192.168.100.1 1.0.100.1 … VM002 192.168.100.2 1.0.100.2 VM003 192.168.100.3 1.0.100.3 VM008 192.168.100.8 1.0.100.8 192.168.100.98 Dom 0 Hyper-V soft switch X520 ‐ SR2 Por t1 Port0 PF VF VF VF VF
51. Networks of Virtulization - SRIOV SERVER 2 192.168.100.98 SERVER 1 X520 ‐ SR2 X520 ‐ SR2 VM01 192.168.100.51 1.0.100.51 … 192.168.100.99 Dom0 Por t1 Hyper-V soft switch Port0 PF VF VF VF VF VM02 192.168.100.52 1.0.100.52 VM03 192.168.100.53 1.0.100.53 VM08 192.168.100.58 1.0.100.58 VM001 192.168.100.1 1.0.100.1 … VM002 192.168.100.2 1.0.100.2 VM003 192.168.100.3 1.0.100.3 VM008 192.168.100.8 1.0.100.8 192.168.100.98 Dom0 Hyper-V soft switch X520 ‐ SR2 Por t1 Port0 PF VF VF VF VF
52. Performance Data from Netperf 4.6Gbps without IOV vs. 9.4Gbps with IOV
53. 微软* Windows* Server 8 with Hyper-V* PCI-SIG 单根的 I/O 虚拟化和共享 (SR-IOV) 万兆网络虚拟化 虚拟机队列 (VMQ): 通过英特尔VMDq技术由英特尔以太网控制器队列实现网络包的分 流。包被排序并送往相应的虚拟机，实现多个虚拟机对接收的并行 处理。通过控制器QoS循环调度器每个队列由一个不同的处理器 核进行处理。数据包被送往以太网控制器队列避免了从控制器接收 缓存到虚拟机地址空间的数据拷贝过程。 单根的 I/O虚拟化 （SR-IOV）: 可以把具有大量网络业务的虚拟机直接被分配给操作系统，服务器 平台以及英特尔以太网控制器所支持的SR-IOV所提供的虚拟网卡 功能（VF), 从而跳过虚拟机管理器和虚拟交换所带来的开销。 微软* Windows* Server 8 和英特尔® 以太网 • 支持英特尔®以太网产品的VMDq技术和 SR-IOV 技术 • 提供服务质量能够实现网络资源共享和网络的融合 • 利用硬件的优势来提升软件服务质量的性能
54. 英特尔® 至强® E5-2600处理器平台 Performance tests and ratings are measured using specific computer systems and/or components and reflect the approximate performance of Intel products as measured by those tests. Any difference in system hardware or software design or configuration may affect actual performance. Buyers should consult other sources of information to evaluate the performance of systems or components they are considering purchasing. For more information on performance tests and on the performance of Intel products, visit http://www.intel.com/performance/resources/benchmark_limitations.htm. Intel, the Intel logo, Xeon and Xeon Inside are trademarks of Intel Corporation in the U.S. and/or other countries. Intel® Ethernet Controllers 英特尔® 10GBASE-T 以太网控制器: X540 让我们真正地一起拥抱 万兆以太网 时代
55. Questions