Subject: Re: Latency vs performance (was Re: Sun Servers take advantage of board local memory?)
Date: Mon, 27 Apr 1998 12:24:21 -0700
From: handleym@apple.com (Maynard Handley)
Organization: Apple Computer
Newsgroups: comp.sys.sun.hardware,comp.sys.sun.misc,comp.unix.solaris,comp.arch,comp.sys.dec,comp.benchmarks

In article <25APR199814272773@gerg.tamu.edu>, carl@gerg.tamu.edu (Carl
Perkins) wrote:

> What if you are looking at it wrong?
>
> }Several vendors offer large SMP machines using the same cpus as
> }in smaller machines (uniprocessors or small SMP's).  In every case,
> }the latency of the larger machines is higher than the latency of
> }the smaller machines, and the penalty is clearly visible in SPECfp95
> }performance.
> }-----------------------------------------------------------------------
> }            big machine                    small machine
> }vendor model latency SPECfp95          model latency SPECfp95
> }DEC    8400   ~400   20.8 @ 612MHz     4100   ~250   25.1 @ 612MHz (1)
> }HP     V2250  500?   24.8 @ 240MHz     K580    ???   28.5 @ 240MHz (2)
> }Sun    UE10K  ~500   11.3 @ 250MHz (3) UE450  ~200   15.0 @ 250MHz
> }-----------------------------------------------------------------------
> }SGI    Origin  290   24.5 @ 250MHz     Origin  290   24.5 @ 250MHz
> }-----------------------------------------------------------------------
> }
> }In summary, DEC loses 17% of uniprocessor SPECfp95 due to the
> }increased latency of the larger system, Sun loses 25%, and HP
> }loses 13%.  The SGI results are in there to show that a correctly
> }designed ccNUMA machine has no uniprocessor performance penalty
> }when running in large configurations.
> }John D. McCalpin, Ph.D.      Server System Architect
>
> In summary, DEC gains 20% of multiprocessor SPECfp95 due to the
> decreased latency of the smaller system, Sun gains 32%, and HP
> gains 15%. The SGI results are there to show that an incorrectly
> designed ccNUMA machine has no uniprocessor performance gain
> when running in small configurations.

And, for those of us simply looking on at this, and with no especial stake
in who's right or wrong, the above numbers, in the absence of the cost
attached to each box are not really useful For example, if SGI's
(according to John) superbly engineered cc-NUMA box costs 1.5x as much as
the comparable DEC box, maybe that superb engineering is not the best use
of resources.

Maynard

--
My opinion only