Skip site navigation (1)Skip section navigation (2)
Date:      Thu, 24 Aug 1995 11:33:40 -0700 (PDT)
From:      "Rodney W. Grimes" <rgrimes@gndrsh.aac.dev.com>
To:        brian@arl.wustl.edu (Brian Gottlieb)
Cc:        freebsd-hardware@freebsd.org
Subject:   Re: Upgrade to my machine
Message-ID:  <199508241833.LAA08175@gndrsh.aac.dev.com>
In-Reply-To: <9508241430.AA00576@beru.wustl.edu> from "Brian Gottlieb" at Aug 24, 95 09:30:54 am

next in thread | previous in thread | raw e-mail | index | archive | help
[CC: trimmed, save a few people dual copies]

> -Vince-  <vince@penzance.econ.yale.edu> (-Vince-) writes:
> 
> -Vince-> On Wed, 23 Aug 1995, Brian Gottlieb wrote:
> >> 
> >> It all depends on what you're doing with it.  In my machine at work I
> >> have a 400 meg drive dedicated to swap.  The circuit synthesis and
> >> simulations we run here need LOTS of memory and swap.  The "big"
> >> machines in our group have 256M of memory and a 1 Gig swap drive.
> 
> -Vince-> 	Hmmm, is there like a way to do well with a big swap and
> -Vince-> like 16 megs of physical memory?  How much physical memory is
> -Vince-> on the machine with 400 meg swap?
> 
> Today is a good day to answer this.  Last night I got a memory
> upgrade.  There is now 192M of ram in the machine (I feel like a kid
> in a playground..."My machine could beat up your machine" ;)

:-), okay I'll play, but there not BSD boxes :-).

> When I originally wrote this, I had 64M RAM.  But the synthesiser
> swapped too much and took way to long to run on here.  With more
> memory, it swapped less and ran much faster.  I don't remember the
> numbers, but it was very significant.  A Sparc 5 with 192M RAM was
> keeping up with a Sparc 10 with 128M.  One factor we probably didn't
> consider was that the 10 may have had a faster disk on it.
> 
> I'm no expert on things, but I think there is a point where you may
> get diminishing returns on having lots of swap.  On the other hand, I
> have 16 Megs RAM in my freebsd machine and 32 megs of swap, and I have
> run out of memory a few times.  More swap would probably help.  But
> perhaps if I got my swap too big, it may become less efficient since I
> could run more programs, but it would spend more time swapping.

Most times when you start to _use_ more than 2x physical memory for swap
you are taking large performance hits.  This may be acceptable depending
on what it is you are doing.  Since I happen to catch you are playing
with what appears to be logic synthisis stuff, the performance hit is the
only way you will ever run this type of stuff.

I have done a far amount of work in supporting Intel in the area of
clusters of HP9000/7XX (they upgrade these things to the fastest boxes
avaliable every 6 months, thus the XX) for the purpose of running Synopsis
(large commercial logic synthesis software).  Typical machine is 512 to
768MB of physical ram, and 1 to 2G of swap.

>From watching the systems with ``Glance'' (an HP performance montioring
tool) we learned that once swap usage hits about 4x physical memory the
synopsis jobs slow to a crawl and it is best to partition the synthesis
run and restart it rather than let it grind away forever.

Note, this is one specific example of large memory large swap configurations,
it works fairly well for the intended purpose.  I would never want to log
into one of these systems and try to do interactive work (well, unless there
was no jobs running, then there the quite NICE to use).  

My general rule of thumb is if the box has an interactive work load on
it (ie, users running a shell) that anytime you hit 2x physical memory
in swap space your paying for it badly and rather than throwing some more
swap space at it you should really consider a memory upgrade.  If the
machine is a non-interactive cruncher (I don't consider ftp an interactive
application, so this applies to wcarchive.cdrom.com) 4x to 6x is pretty
much an upper limit before performance takes such a hit it becomes useless.

> I don't know.  To tell the truth, the 400 Megs of swap on here
> probably rarely gets filled up, since I typically ran my simulations
> elsewhere.  But I know that the Sparc 10 in the office upstairs was
> swapping like mad when I ran stuff on it (I could almost hear the
> drive down here ;).

I have seen the engineers out at Intel consume as much as 1.5G of swap,
but most of the time they are working with a partition of the design and
use about 1G of swap.  They cry wolf very loudly when a 18 hour synthesis
run barfs on an out of memory error, so a close eye is keep on the upper
bounds of swap useage and anytime jobs start getting close to it they
are asked to watch out, and if they can repartition the design, or give
us a day to chuck another 1G swap disk on a machine and target there jobs
to it.

> Hmm...long post.  So the answer is...I have no clue ;)

Its not that you have to clue, it's that this is all black magic and
there are no hard rules to follow. :-).

> brian


-- 
Rod Grimes                                      rgrimes@gndrsh.aac.dev.com
Accurate Automation Company                 Reliable computers for FreeBSD



Want to link to this message? Use this URL: <https://mail-archive.FreeBSD.org/cgi/mid.cgi?199508241833.LAA08175>