[SCore-users-jp] Re: [SCore-users] trunking
Shinji Sumimoto
s-sumi @ flab.fujitsu.co.jp
2003年 3月 13日 (木) 21:47:36 JST
Hi.
Sorry for late response.
Are you using a switch that supports JUMBO Frame ?
If so, how about mpi_zerocopy=on option?
Here are results using Intel PRO/1000XTs on Supermicro mother boards.
These results are not so good compared with Myrinet 2000.
Broadcom 5701 based NIC has also good communication performace.
See: http://www.pccluster.org/score/dist/score/html/en/overview/pm-perf.html
maxnsend 24
backoff 2000
with mpi_zerocopy=on
***** Two Intel PRO/1000XTs.
#-----------------------------------------------------------------------------
# Benchmarking Sendrecv
# ( #processes = 2 )
#-----------------------------------------------------------------------------
#bytes #repetitions t_min[usec] t_max[usec] t_avg[usec] Mbytes/sec
0 1000 25.06 25.06 25.06 0.00
1 1000 24.97 25.03 25.00 0.08
2 1000 24.38 24.41 24.40 0.16
4 1000 25.18 25.20 25.19 0.30
8 1000 24.57 24.63 24.60 0.62
16 1000 25.52 25.54 25.53 1.19
32 1000 24.98 25.01 25.00 2.44
64 1000 25.09 25.11 25.10 4.86
128 1000 24.71 24.71 24.71 9.88
256 1000 29.80 29.84 29.82 16.36
512 1000 30.46 30.50 30.48 32.02
1024 1000 46.70 46.78 46.74 41.75
2048 1000 63.69 73.19 68.44 53.37
4096 1000 112.09 112.15 112.12 69.66
8192 1000 191.11 191.21 191.16 81.71
16384 1000 247.54 247.60 247.57 126.21
32768 1000 652.67 652.69 652.68 95.76
65536 1000 956.54 956.55 956.54 130.68
131072 1000 1559.37 1559.38 1559.38 160.32
262144 640 2737.14 2737.14 2737.14 182.67
524288 320 4946.09 4946.14 4946.12 202.18
1048576 160 9352.07 9352.14 9352.10 213.85
2097152 80 24004.33 24004.79 24004.56 166.63
4194304 40 48974.10 48975.73 48974.91 163.35
***** Three Intel PRO/1000XTs.
#-----------------------------------------------------------------------------
# Benchmarking Sendrecv
# ( #processes = 2 )
#-----------------------------------------------------------------------------
#bytes #repetitions t_min[usec] t_max[usec] t_avg[usec] Mbytes/sec
0 1000 30.05 30.06 30.06 0.00
1 1000 25.05 25.13 25.09 0.08
2 1000 21.71 21.77 21.74 0.18
4 1000 23.80 23.85 23.83 0.32
8 1000 32.45 32.48 32.46 0.47
16 1000 28.11 28.13 28.12 1.08
32 1000 24.24 24.30 24.27 2.51
64 1000 22.63 22.67 22.65 5.38
128 1000 25.27 25.28 25.27 9.66
256 1000 28.47 28.54 28.51 17.11
512 1000 30.25 30.30 30.28 32.23
1024 1000 45.22 45.29 45.26 43.12
2048 1000 62.34 71.10 66.72 54.94
4096 1000 96.77 96.86 96.81 80.66
8192 1000 178.64 178.79 178.71 87.39
16384 1000 564.16 564.17 564.16 55.39
32768 1000 625.56 625.58 625.57 99.91
65536 1000 798.51 798.52 798.52 156.54
131072 1000 1289.24 1289.24 1289.24 193.91
262144 640 2369.67 2369.68 2369.68 211.00
524288 320 4019.12 4019.12 4019.12 248.81
1048576 160 7143.38 7143.41 7143.39 279.98
2097152 80 14925.95 14926.55 14926.25 267.98
4194304 40 29971.10 29973.05 29972.07 266.91
PS: I am now re-writing PM/Ethernet to reduce communication cost.
Shinji.
From: Nick Birkett <nrcb @ streamline-computing.com>
Subject: [SCore-users] trunking
Date: Wed, 12 Mar 2003 23:51:02 +0000
Message-ID: <200303122351.02223.nrcb @ streamline-computing.com>
nrcb> Some first results.
nrcb>
nrcb> Hardware: 1U dual Xeon 2.6GHz Superservers with onboard dual gigabit.
nrcb> Each network via its own gigabit switch.
nrcb>
nrcb> Pallas benchmarks: Pingpong looks ok but Sendrecv is not good:
nrcb>
nrcb> backoff 1024
nrcb> maxnsend 16
nrcb>
nrcb> on both networks.
nrcb>
nrcb> See attached benchmarks.
nrcb>
nrcb>
------
Shinji Sumimoto, Fujitsu Labs
_______________________________________________
SCore-users mailing list
SCore-users @ pccluster.org
http://www.pccluster.org/mailman/listinfo/score-users
SCore-users-jp メーリングリストの案内