[SCore-users] Score 4.2.1 on new Athlon systems.

[Shinji Sumimoto]

Hi, Nick. 

From: Nick Birkett <nrcb at streamline-computing.com>
Subject: Re: [SCore-users] Score 4.2.1 on new Athlon systems.
Date: Mon, 11 Mar 2002 20:37:12 +0000
Message-ID: <02031120371200.01590 at pecan.comlab.ox.ac.uk>

nrcb> On Thursday 07 March 2002 05:49, you wrote:
nrcb> > Hi.
nrcb> >
nrcb> > From: Nick Birkett <nrcb at streamline-computing.com>
nrcb> > Subject: [SCore-users] Score 4.2.1 on new Athlon systems.
nrcb> > Date: Thu, 7 Mar 2002 01:16:58 +0000
nrcb> > Message-ID: <0203070116580I.19905 at pecan.comlab.ox.ac.uk>
nrcb> > nrcb> Hi , we have been running both Score and Myricom GM.
nrcb> > nrcb>
nrcb> > nrcb>  We now have some new systems based on Tyan Tiger MPX multi-processor
nrcb> > boards nrcb> (1.6 GHz Athlon with DDR memory).
nrcb> > nrcb>
nrcb> > nrcb> Pallas benchmark PMB-MPI1 compiled using -O4 -fomit-frame-pointer
nrcb> > nrcb> -funroll-all-loops  and with -DnoCHECK gives very poor bandwidth
nrcb> > result for nrcb> Score Myrinet2k, whereas GM gives good one.
nrcb> > nrcb>
nrcb> > nrcb> Score 4.2.1 gives 90Mbytes/sec point to point over Myrinet2k.
nrcb> > nrcb>
nrcb> > nrcb> GM/MPICH (GM 1.5, MPICH 1.2..8 and 1.2.7.. ) gives twice this
nrcb> > bandwidth nrcb> over Myrinet.
nrcb> > nrcb>
nrcb> > nrcb> GM is using kernel 2.4.18 and  Score uses 2.4.10smp (the one that
nrcb> > comes with nrcb> Score 4.2.1 binary release).
nrcb> > nrcb>
nrcb> > nrcb> We have compared GM and Score on other Athlon systems and do not have
nrcb> > this nrcb> big difference (usually bandwidth is same within a few %).
nrcb> >
nrcb> > MPICH/GM uses zero-copy at default, but MPICH/SCore does not use
nrcb> > zero-copy at default.
nrcb> >
nrcb> > Could you test the program using mpi_zerocopy=on?
nrcb> >
nrcb> >
nrcb> 
nrcb> Ok many thanks. We got our Myrinet cards back from another benchmark (at 
nrcb> last). Sorry for the delay.
nrcb> 
nrcb> Pallas benchmark with mpi_zerocopy=on option now gives 231Mbytes/s
nrcb> on Athlon MPX systems (using scrun -nodes=2x1 ), which is *EXCELLENT*.
nrcb> 
nrcb> Can Shinji or someone explain why mpi_zerocopy is off be default, but in 
nrcb> MPICH-GM it is on ?
nrcb> What are the advantages/disadvantages ?

Because MPICH/SCore supports not only Myrinet but also Ethernet and
other networks, however, MPICH/GM supports only Myrinet.

It depends on hardware combinations whether zero copy communication
takes effect or not in the point of communication performance.

In the combination of Myrinet 2000 and 64bit PCI bus, zero-copy
communication takes effects. However, 100Base/T Ethernet, zero-copy
communication does not takes effects.

Anyway, providing some default options for each network is good idea
for users to use. We will try to introduce the options.

Shinji.
------
Shinji Sumimoto, Fujitsu Labs