[SCore-users-jp] Re: [SCore-users] Something strange in PM/Ethernet Network Trunking
Shinji Sumimoto
s-sumi ＠ flab.fujitsu.co.jp
2004年 5月 7日 (金) 18:40:47 JST
Hi.
From: Chen Hao <hchen ＠ mdl.ipc.pku.edu.cn>
Subject: Re: [SCore-users] Something strange in PM/Ethernet Network Trunking
Date: Fri, 07 May 2004 17:31:15 +0800
Message-ID: <409B5763.8090705 ＠ mdl.ipc.pku.edu.cn>
hchen> I didn't change the values maxnsend and backoff options in
hchen> pm-ethernet.conf(maxnsend=16, backoff=4800). And as I posted early,
hchen> pingpong and pingping's results had great different, It seemed very strange.
Could you change the value maxnsend=24, backoff=2400, and test?
The PingPing is a burst test, so communicaiton performance is
decreased if some re-transmission has occurred.
If the situation does not change, please test PM level communication
bandwidth using rpmtest.
http://www.pccluster.org/score/dist/score/html/en/installation/pm-testethernet.html
Shinji.
hchen> Shinji Sumimoto wrote:
hchen>
hchen> >Hi.
hchen> >
hchen> >Sorry for late response because of holidays in Japan.
hchen> >
hchen> >Have you changed the values maxnsend and backoff options in
hchen> >pm-ethernet.conf? If so, how about pmtest communication performace?
hchen> >
hchen> >Shinji.
hchen> >
hchen> >From: Chen Hao <hchen ＠ mdl.ipc.pku.edu.cn>
hchen> >Subject: [SCore-users] Something strange in PM/Ethernet Network Trunking
hchen> >Date: Wed, 28 Apr 2004 20:50:12 +0800
hchen> >Message-ID: <408FA884.2020508 ＠ mdl.ipc.pku.edu.cn>
hchen> >
hchen> >hchen> Hello all, I used PMB 2.2.1 to measure SCORE/MPI performance and found
hchen> >hchen> something string. When I used two 3C905B NICs, I got the results of
hchen> >hchen> PingPong and Pingping as following:
hchen> >hchen> #---------------------------------------------------
hchen> >hchen> # Benchmarking PingPong
hchen> >hchen> # ( #processes = 2 )
hchen> >hchen> # ( 14 additional processes waiting in MPI_Barrier)
hchen> >hchen> #---------------------------------------------------
hchen> >hchen> #bytes #repetitions t[usec] Mbytes/sec
hchen> >hchen> 0 1000 49.13 0.00
hchen> >hchen> 1 1000 49.89 0.02
hchen> >hchen> 2 1000 50.40 0.04
hchen> >hchen> 4 1000 50.54 0.08
hchen> >hchen> 8 1000 50.86 0.15
hchen> >hchen> 16 1000 52.92 0.29
hchen> >hchen> 32 1000 56.37 0.54
hchen> >hchen> 64 1000 62.83 0.97
hchen> >hchen> 128 1000 75.88 1.61
hchen> >hchen> 256 1000 98.53 2.48
hchen> >hchen> 512 1000 151.50 3.22
hchen> >hchen> 1024 1000 247.95 3.94
hchen> >hchen> 2048 1000 347.78 5.62
hchen> >hchen> 4096 1000 463.81 8.42
hchen> >hchen> 8192 1000 594.80 13.13
hchen> >hchen> 16384 1000 1056.45 14.79
hchen> >hchen> 32768 1000 1754.30 17.81
hchen> >hchen> 65536 640 3122.35 20.02
hchen> >hchen> 131072 320 5992.52 20.86
hchen> >hchen> 262144 160 11469.87 21.80
hchen> >hchen> 524288 80 22598.54 22.13
hchen> >hchen> 1048576 40 44952.29 22.25
hchen> >hchen> 2097152 20 89360.53 22.38
hchen> >hchen> 4194304 10 178679.39 22.39
hchen> >hchen> #---------------------------------------------------
hchen> >hchen> # Benchmarking PingPing
hchen> >hchen> # ( #processes = 2 )
hchen> >hchen> # ( 14 additional processes waiting in MPI_Barrier)
hchen> >hchen> #---------------------------------------------------
hchen> >hchen> #bytes #repetitions t[usec] Mbytes/sec
hchen> >hchen> 0 1000 55.40 0.00
hchen> >hchen> 1 1000 54.75 0.02
hchen> >hchen> 2 1000 52.09 0.04
hchen> >hchen> 4 1000 55.12 0.07
hchen> >hchen> 8 1000 53.07 0.14
hchen> >hchen> 16 1000 58.13 0.26
hchen> >hchen> 32 1000 57.33 0.53
hchen> >hchen> 64 1000 67.17 0.91
hchen> >hchen> 128 1000 79.80 1.53
hchen> >hchen> 256 1000 98.88 2.47
hchen> >hchen> 512 1000 164.32 2.97
hchen> >hchen> 1024 1000 252.47 3.87
hchen> >hchen> 2048 1000 1901.03 1.03
hchen> >hchen> 4096 1000 3520.96 1.11
hchen> >hchen> 8192 1000 2870.28 2.72
hchen> >hchen> 16384 1000 10062.60 1.55
hchen> >hchen> 32768 1000 10718.80 2.92
hchen> >hchen> 65536 640 20858.36 3.00
hchen> >hchen> 131072 320 26830.07 4.66
hchen> >hchen> 262144 160 39868.81 6.27
hchen> >hchen> 524288 80 62836.03 7.96
hchen> >hchen> 1048576 40 108759.45 9.19
hchen> >hchen> 2097152 20 194836.70 10.27
hchen> >hchen> 4194304 10 379412.10 10.54
hchen> >hchen> we could see that pingpong's bandwidth is very poor when packet is from
hchen> >hchen> 2KB to 64KB, what's the matter?
hchen> >hchen>
hchen> >hchen> ChenHao
hchen> >hchen> _______________________________________________
hchen> >hchen> SCore-users mailing list
hchen> >hchen> SCore-users ＠ pccluster.org
hchen> >hchen> http://www.pccluster.org/mailman/listinfo/score-users
hchen> >hchen>
hchen> >hchen>
hchen> >------
hchen> >Shinji Sumimoto, Fujitsu Labs
hchen> >_______________________________________________
hchen> >SCore-users mailing list
hchen> >SCore-users ＠ pccluster.org
hchen> >http://www.pccluster.org/mailman/listinfo/score-users
hchen> >
hchen> >
hchen> >
hchen>
hchen>
------
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