New Results

From: John, dtn 381-0378 24-May-1996 1654 (henning@perfom.ENET.dec.com)
Date: Fri May 24 1996 - 16:26:00 CDT


Hi John,

In the next messages I'll be sending you some new results using 16, 24,
and 32 cpus on an AlphaServer 8400 TruCluster.

This method of interconnect is looser than traditional SMP, but tighter
than traditional MPP. In the first message after this one, I'll send
you an excerpt from the "Customer Update" article on TruClusters.

I would suggest that you call the result something like

    DEC 8400tc 5/350

in your table, but of course the actual abbreviation is up to you! I'm
sure you don't have space for

    Digital AlphaServer 8400 5/350 with TruCluster and Memory Channel

The starting point for the benchmark was a recent stream_d.f from your
web site. Jonathan Harris of our High Performance Computing Group
modified the source code to use F90 and to add the needed directives.
I'm sure you'll want to diff the sources, which will be in one of the
following messages. I did a diff and his modifications sure seemed to
make sense to me.

The benchmark was compiled using the Digital Fortran 90 compiler V4 with
the command:

     F90 -wsf <n> -fast -non_shared -O5 -tune ev4

The "-wsf" switch is a legacy of the days of workstation farms, but does
NOT indicate the cluster interconnect used. In fact, it uses the memory
channel (and was measurably faster because of it).

The timer was F90 system clock intrinsic

      call system_clock(count = clock_count, count_rate = clock_rate)
      second = (dble(clock_count) / dble(clock_rate))

Have a great Memorial Day weekend!

    /John Henning
     CSD Performance Group
     Digital Equipment Corporation
     henning@zko.dec.com



This archive was generated by hypermail 2b29 : Tue Apr 18 2000 - 05:23:06 CDT