9fans archive / 1994 / 03 / 25 /    prev next

From: rob@pla... rob@pla...
Subject: No subject
Date: Mon, 28 Mar 1994 10:06:43 -0500

the sun is perhaps a poor choice for this test because its
mmu and caches are pessimal for anything involving
lots of processes.  the cost of flushing caches and mmu
will dominate any operating system performance.
in other words, the hardware is so expensive that
software hardly matters.

i tried the test on our challenge multiprocessors.

10,000 fork/exit/wait.

plan 9: real time 1.32 seconds.
unix: real time 41.1 seconds.

the unix system had more load, but was still
pretty quiet.  system time was 27.9 seconds.

on our seven-year-old SGI power machines,
the test takes 6.26 seconds.  the time is about
4 seconds on a single-processor MIPS magnum
of the same vintage, so the multiprocessor
part isn't very important.

now scott's test was 1,000 forks on a sun; my test
was 10,000.  on our at&t gnots, with 25MHz 68020's
and awful mmu, it takes 75 seconds to do 10,000.
our modern sparcstation-2 machine takes 34 seconds.
(a newer machine than the magnum that takes 4 seconds.)
you see, suns suck.  they really do.  the processors are
borderline OK but the rest of the system hurts.  yesterday,
quite independently, i was doing some tests of large-scale
memory bandwidth and the suns were laughable.

as for compiler names, the scheme we have is simple,
consistent, and parsimonious.  if crypticness is the
complaint, remember you're talking about a C compiler.