Subject: Re: SMP/flogging a dead horse
To: Erik E. Fair <email@example.com>
From: nm <firstname.lastname@example.org>
Date: 08/29/1998 12:32:16
At 09:20 AM 8/29/98 -0700, you wrote:
>There is a marvelous book I recommend:
> In Search of Clusters, 2nd Ed.
> Gregory Pfister
>It discusses multiprocessor configurations extensively, in an
>In an MP configuration (kernel runs on only one processor), you
>will bottleneck on the processor running the kernel. How soon that
>happens depends on what type of job mix (computing load) you have.
>If you're I/O intensive, MP may be a net-lose, because switching
>processors has some cost in L1/L2 cache coherency and such. If you're
>*totally* CPU-bound (e.g. cracking RC5-64 keys, SETI@Home, GIMPS),
>then MP should win quite nicely.
Does having an onchip L2 (like the ppro) make the cache hit less of
an issue than on say a pentium chip?
>The trick is getting your job mix to spend more than 50% of its
>time in user mode, computing, instead of making system calls. The
>more processors you have, the more CPU-bound jobs you need to take
>full advantage of the system.
>There is a short paper about the very first MP UNIX configuration
>(a two-headed VAX-11/780) available at
>George Goble also lights his BBQ grill with LOX to make sure that
>it takes minimum time to be ready to grill those hot dogs and
>It's quite a sight.
> Erik E. Fair email@example.com