title: 集群中各服务器的配置与性能 nav_order: 6 —

集群中各服务器的配置与性能

August 16, 2023, Xiang Li, Yuejia Zhang

CPU / Memory

评测程序 Source: PassMark Performance Testing Linux

版本: 10.2.1003

参数: (-i 3) Test Iterations: 3, (-d 2) Test Duration: Medium.

测试结果仅供参考, 有一定的波动. (?) 为未经有效确认的结果.

CPU Info loginNode bigMem0 bigMem1 bigMem2 bigMem3
CPU Brand Intel Xeon CPU E5-2670 v3 @ 2.30GHz Intel Xeon Gold 6226R CPU @ 2.90GHz Intel Xeon Gold 6226R CPU @ 2.90GHz Hygon C86 7285 32-core Processor AMD EPYC 9754 128-Core Processor
Num of CPUs on Board 2 2 2 2 2
Total Threads 24 64 64 128 512
Base Clock Speed (GHz) 2.30 2.90 2.90 2.00 2.25
Boost Clock Speed (GHz) 2.30 3.90 3.90 3.00(?) 3.10
CPU Cache (MiB/CPU) 30 22 22 64 256
Lithography (Nanometer) 22 14 14 14(?) 5
CPU Speed loginNode bigMem0 bigMem1 bigMem2 bigMem3
Integer Math (Million Operations/s) 62,275 206,066 209,673 328,027 2,054,471
Floating Point Math (Million Operations/s) 48,122 129,819 130,277 146,501 1,149,133
Prime Numbers (Million Primes/s) 188 209 230 289 1,728
Sorting (Thousand Strings/s) 41,063 90,967 106,039 172,684 783,396
Encryption (MB/s) 5,959 25,616 27,013 99,718 492,442
Compression (MB/s) 315 817 865 1,330 6,933
CPU Single Threaded (Million Operations/s) 1,414 2,421 2,407 1,478 2,444
Physics (Frames/s) 1,229 2,336 3,002 7,420 22,446
Extended Instructions (SSE) (Million Matrices/s) 21,430 46,552 47,889 45,259 428,652
CPU Final Mark 19,471 42,705 46,047 53,223 138,716
Memory Info loginNode bigMem0 bigMem1 bigMem2 bigMem3
Total Available RAM (GiB) 125.5 1,006.4 1,006.3 485.6 1007.0
Memory Frequency (MHz) 2,133 2,666 2,933 3200 4800
Memory Speed loginNode bigMem0 bigMem1 bigMem2 bigMem3
Memory Latency (Nanoseconds) 50 53 52 62 70
Memory Read Cached (MB/s) 17,005 27,710 27,698 18,298 23,561
Memory Read Uncached (MB/s) 8,187 12,093 11,800 8700 23,334
Memory Write (MB/s) 7,665 10,100 8,890 7338 23,305
Memory Threaded (MB/s) 71,989 100,377 169,758 259,208 726,276
Database Operations (Thousand Operations/s) 10,649 18,742 19,791 16,254 29,555
Memory Final Mark 2,277 2,827 2,799 2,321 2,876

Reference:

E5 2670 v3 Specs

Gold 6226R Specs

Hygon IPO Specs pp.124-126

EPYC 9754 Specs(AMD) EPYC 9754 Specs(TechPowerUp)

GPU / DCU 计算加速卡

bigMem3 上无加速卡.

评测程序 Source: Mixbench

GPU/DCU loginNode bigMem0 bigMem1 bigMem2
Device NVIDIA GeForce GTX 1080 Ti Tesla T4 NVIDIA A30 HYGON Z100
CUDA driver version 11.60 11.60 11.60 -
GPU clock rate (MHz) 1721 1590 1440 1319
Memory clock rate 2752 MHz 2500 MHz 607 MHz - MHz
Memory bus width 352 bits 256 bits 3072 bits - bits
WarpSize 32 32 32 64
L2 cache size 2816 KB 4096 KB 24576 KB 8192 KB
Total global mem 11178 MB 14910 MB 24068 MB 32752 MB
ECC enabled No Yes Yes Yes
Compute Capability 6.1 7.5 8.0 -
Total SPs 3584 (28 MPs x 128 SPs/MP) 2560 (40 MPs x 64 SPs/MP) 7168 (56 MPs x 128 SPs/MP) 3840 (60 CUs x 64 SPs/CU)
Compute throughput (theoretical single precision FMAs) (GFlops) 12336.13 8140.80 20643.84 10129.92
Memory bandwidth (GB/sec) 484.44 320.06 933.12 1024.00

benchmark

磁盘读写速度

评测程序: fio-3.7

命令 $ fio --randrepeat=1 --ioengine=libaio --direct=1 --gtod_reduce=1 --name=test --filename={} --bs=4ki --iodepth=64 --size=400Mi --readwrite={method} --rwmixread={rwrate},

其中 method=randrw | rw, rwrate=0 | 75 | 100, filename 指定为对应硬盘下的位置.

Spec r/w loginNode:/ loginNode:/home, /scratch
randrw, 75% read + 25% write read 1.5 MiB/s 3.4 MiB/s
  write 0.5 MiB/s 1.1 MiB/s
randrw, 100% read read 4.1 MiB/s 12.5 MiB/s
randrw, 100% write write 1.4 MiB/s 1.6 MiB/s
rw, 75% read + 25% write read 21.8 MiB/s 27.5 MiB/s
  write 7.6 MiB/s 9.4 MiB/s
rw, 100% read read 176.0 MiB/s 407.0 MiB/s
rw, 100% write write 161.0 MiB/s 14.0 MiB/s