Re: [Myrinet] [Q] VIGM performance

Thanks Patrick,

With the pingpong program distributed by Myricom, Inc., I used the below 
command and the result is shown. (yes, it is a blocking test)

$./pingpong -s canxx.our-school.ac.kr 1 1000 4 65536
# VI PingPong (1000 iterations, size = from 4 to 65536 Bytes)
Size (Bytes)     Latency (us)    Bandwidth (MB/s) (bps)
4               62.60           0.06
8               62.58           0.12
16              62.74           0.24
32              63.68           0.48
64              65.16           0.94
128             70.25           1.74
256             76.20           3.20
512             84.48           5.78
1024            101.25          9.65
2048            136.19          14.34
4096            258.43          15.12
8192            301.29          25.93
16384           390.31          40.03
32768           564.89          55.32
65536           915.74          68.25

A non-blocking test result is also shown below.

$./pingpong -s canxx.our-school.ac.kr 0 1000 4 65536
# VI PingPong (1000 iterations, size = from 4 to 65536 Bytes)

Size (Bytes)     Latency (us)    Bandwidth (MB/s)
4               14.45           0.26
8               14.65           0.52
16              14.77           1.03
32              15.85           1.93
64              16.45           3.71
128             21.48           5.68
256             27.16           8.99
512             34.01           14.36
1024            48.55           20.12
2048            79.47           24.58
4096            147.41          26.50
8192            186.89          41.80
16384           276.80          56.45
32768           454.61          68.74
65536           803.42          77.79

I got a similar result to yours in case of a non-blocking test. 
But I think it is not fair to use a non-blocking test for measuring 
the latency. For a correct pingpong test, there should be only 
a single packet in the Myrinet link. Using a non-blocking test issues
as many packets as possible to the Myrinet link all at once. 
(it better fits to an effective bandwidth test, I think). 

I wonder how you executed your test. 

With regards,
sang


On Mon, Sep 09, 2002 at 09:58:29PM -0400, Patrick Geoffray wrote:
> Hi Moon-Sang,
> 
> On Sun, 2002-09-08 at 11:03, Moon-Sang Lee wrote:
> > Our cluster consists of P-III 850MHz dual-SMP nodes, and I'm testing 
> > VIGM on Myrinet NIC M2L-PCI64B. According to our measurement using the 
> > pingpong program in VIGM source tree, the latency shows 62 us for 4bytes 
> > messages (not the round-trip latency). I think it's terrible. :(
> 
> Which command line do you use to run the ping pong test ?
> 
> > Could anybody tell me the true latency of VIGM? I'm afraid of having 
> > a mistake to measure the latency. 
> 
> On my Dual PIII at 1 GHz, using PCI64C cards, I get 10 us for 4 Bytes.
> 
> > While I have tried to find out the performance of VIGM over the internet, 
> > it was not easy. 
> 
> Trying to find a ping pong code on the web was not easy either. I had to
> write my own, and it's not a perfect one.
> 
> > p.s. the bandwidth data would also be appreciated.
> 
> Peak bandwidth is roughly the same than MPICH-GM, aka 245 MB/s on L9 at
> 200 MHz.
> 
> Patrick
> -- 
> ----------------------------------------------------------
> |   Patrick Geoffray, Ph.D.      [email protected] 
> |   Myricom, Inc.                http://www.myri.com
> |   Cell:  865-389-8852          685 Emory Valley Rd (B)
> |   Phone: 626-821-5555          Oak Ridge, TN 37830
> ----------------------------------------------------------


------------Korea Advanced Institute of Science and Technology-------------
Moon-Sang Lee	    [email protected]
CS dept. KAIST      +82-42-869-5585@calab  
Sarang Kwan. 524    +82-42-879-4145@dormitory
Homepage URL        http://camars.kaist.ac.kr/~mslee
---------------------Computer Architecture Laboratory----------------------