Path: utzoo!utgpu!news-server.csri.toronto.edu!bonnie.concordia.ca!thunder.mcrcim.mcgill.edu!snorkelwacker.mit.edu!usc!sdd.hp.com!hplabs!pyramid!moliver
From: moliver@shadow (Mike Oliver)
Newsgroups: comp.arch
Subject: Re: ~8-job "knee" in response curves on Suns (was Re: IBM RS6000)
Message-ID: <142018@pyramid.pyramid.com>
Date: 22 Jan 91 18:55:50 GMT
Sender: daemon@pyramid.pyramid.com
Reply-To: moliver@shadow.pyramid.com (Mike Oliver)
Organization: Pyramid Technology Corp., Mountain View, CA
Lines: 51

In article <3964@skye.ed.ac.uk> richard@aiai.UUCP (Richard Tobin) writes:
>
>Jim Reid sent me results from a 4-processor Sequent symmetry.  The results
>show a remarkably slow decline (or perhaps, the Sun ones are rather fast).

I suspect that the Sun decay is rather fast, but that's an
understandable feature of the memory management philosophy on a machine
built for a small number of users.  Doesn't seem to have hurt their
sales much :-)

>Perhaps this is due to the multiple processors.

More likely due to the behaviour of the cache and MM subsystem.

> [Symmetry results elided.]

Figures for a single-CPU Pyramid MIServer (running a multi-CPU kernel,
so there's some semaphore overhead in here) and for a 4-CPU MIServer,
are:

# processes	single CPU	4 CPU
2:		2000		2000 switches / second
4:		1846		1846 switches / second
6:		1895		1800 switches / second
8:		1846		1846 switches / second
10:		1818		1818 switches / second
20:		1765		1765 switches / second
30:		1731		1748 switches / second
40:		1667		1752 switches / second
50:		1631		1714 switches / second
60:		1579		1706 switches / second
70:		1597		1641 switches / second
80:		1584		1673 switches / second
90:		1561		1657 switches / second
100:		1600		1681 switches / second
120:		1545		1655 switches / second
140:		1530		1606 switches / second
160:		1569		1624 switches / second
180:		1506		1607 switches / second
200:		1540		1594 switches / second

These are on "idle" (not dedicated, but not much else going on)
machines in multi-user mode, using the default number of 100 messages
through the ring of pipes in Richard Tobin's benchmark.  Sending 2000
messages around the ring to establish a steady state improves these
figures uniformly by about 200, or between 10 and 13 per cent.

Cheers, Mike.

moliver@pyramid.com
{allegra,decwrl,hplabs,munnari,sun,utai,uunet}!pyramid!moliver