In high performance computing (HPC), message passing interface (MPI) benchmarks are used to demonstrate the performance capability of the cluster network. While application performance is the most important result, benchmarking generally starts with standard micro-benchmarks used to determine best-case MPI latency, bandwidth and message rate. An HPC cluster is only as fast as the individual servers can communicate with each other. Intel® Omni-Path Architecture (Intel® OPA) has been designed to meet the requirements of clusters from small to large scale. This includes ground-breaking quality of service (QoS) features meant to keep latency low, bandwidth high, and message rate high, even at scale.1
This figure compares latency with 8 byte messages for Intel® OPA relative to Enhanced Data Rate* (EDR) InfiniBand* (IB), as measured with the Ohio State University (OSU) OMB osu_latency benchmark for both Open MPI and Intel® MPI. Intel® OPA latency has been measured up to 11% lower than EDR IB*. This latency includes a switch hop for both Intel OPA and EDR IB*.
This figure compares bandwidth with 1 MB messages for Intel® OPA relative to EDR as measured with the OSU OMB osu_bw benchmark for both Open MPI and Intel® MPI. Both Intel® OPA and EDR are capable of delivering nearly full wire rate of 100 Gbps.
This figure compares 8 byte message rate for Intel® OPA relative to EDR as measured with the OSU OMB osu_mbw_mr benchmark for both Open MPI and Intel® MPI. 32 MPI rank pairs are used in the measurement. Intel® OPA has been measured up to 64% higher than EDR without message coalescing at the MPI level. This is a true hardware message rate test without message coalescing in software.
Natural Order Ring (NOR) Latency
Natural order ring (NOR) latency is measured from the b_eff benchmark in the HPC Challenge* benchmark suite. These measurements demonstrate the ability of the fabric to sustain low latency as the cluster scales an HPC application. Intel® OPA has lower latency at 16 fully subscribed nodes using 32 MPI ranks per node.
Random Order Ring (ROR) Latency
Random order ring (ROR) latency is measured from the b_eff benchmark in the HPC Challenge* benchmark suite. These measurements demonstrate the ability of the fabric to sustain low latency as the cluster scales an HPC application. Intel® OPA has lower latency at 16 fully subscribed nodes using 32 MPI ranks per node.
|Test platform||Intel® Xeon® Processor E5-2697A v4 dual-socket servers (16 cores, 40 MB cache, 2.6 GHz, 9.6 GT/s Intel QuickPath Interconnect, 145 W TDP) with 64 GB DDR4 memory @ 2133 MHz. Intel® Turbo Boost Technology and Intel® Hyper-Thread Technology enabled|
Ohio State Micro Benchmarks* v. 5.0
|Operating system||Red Hat Enterprise Linux* 7.2|
|Intel® MPI||Intel® MPI 5.1.3|
|Open MPI*||Open MPI 1.10.0|
|Intel® OPA hardware and settings||shm:tmi fabrics, I_MPI_TMI_DRECV=1, Intel Corporation Device 24f0 – Series 100 HFI ASIC (B0 silicon). OPA Switch: Series 100 Edge Switch – 48 port (B0 silicon). IOU Non-posted Prefetch disabled in BIOS. Snoop hold-off timer = 9|
|EDR hardware and settings||shm:dapl fabric. -genv I_MPI_DAPL_EAGER_MESSAGE_AGGREGATION off (Intel® MPI Only). Mellanox* EDR ConnectX-4 Single Port Rev 3 MCX455A HCA. Mellanox SB7700 - 36 Port EDR InfiniBand switch. MLNX_OFED_LINUX-3.2-188.8.131.52 (OFED-3.2-2.0.0), Best of default, MXM_TLS=self,rc, and -mca pml yalla tunings|
a. osu_latency 8B message
b. osu_bw 1 MB message
c. osu_mbw_mr, 8 B message (uni-directional), 32 MPI rank pairs. Maximum rank pair communication time used instead of average time, introduced into Ohio State Micro Benchmarks as of v3.9 (2/28/13). EDR using shm:ofa fabric, since this returned better message rates than shm:dapl fabric with I_MPI_DAPL_EAGER_MESSAGE_COALESCING disabled
性能测试中使用的软件和工作负载可能仅可基于英特尔® 微处理器进行性能优化。性能测试（如 SYSmark* 和 MobileMark*）使用特定计算机系统、组件、软件、操作和功能进行测量。对这些因素的任何更改可能导致不同的结果。您应该查询其他信息和性能测试以帮助您对正在考虑的采购作出全面的评估，包括该产品在与其他产品结合使用时的性能。有关更多完整的信息，请访问 https://www.intel.cn/benchmarks。