Skip site navigation (1)Skip section navigation (2)
Date:      Mon, 15 Feb 1999 12:15:05 -0800
From:      Kent Stewart <kstewart@3-cities.com>
To:        Karl Pielorz <kpielorz@tdx.co.uk>
Cc:        Thomas van Gulick <melkor@Cal040031.student.utwente.nl>, freebsd-questions@FreeBSD.ORG
Subject:   Re: FreeBSD performance
Message-ID:  <36C88049.A423185A@3-cities.com>
References:  <19990215144708.A34665@utumno.student.utwente.nl> <36C83508.B7F45FC8@tdx.co.uk>

next in thread | previous in thread | raw e-mail | index | archive | help


Karl Pielorz wrote:
> 
> Thomas van Gulick wrote:
> >
> > I'm currently evaluating FreeBSD as a replacement of Linux as my primary
> > development OS. I'm using a Dual PentiumPro system with 192 MB for this.
> 
> I have a similar setup, Dual P-Pro 200 with 256Mb of RAM...
> 
> > However, FreeBSD feels a bit 'sluggish' compared to Linux. A bit slower on
> > every task so it seems.
> 
> Hmmm... That's funny, the main reason I switched to FreeBSD was it performed
> better / faster - especially under load  - heavier the better :)
> 
> > Are there any performance tweaks to increase my system's operating speed?
> > I'm currently using the 4-CURRENT, async io and edited the kernel config
> > file for as much speed as possible.
> 
> If your running -Current try switching to softupdates rather than async IO -
> The _only_ area I found Linux used to be quicker than FreeBSD was with
> filesystem throughput, which is apparntly because they run their filesystem as
> the equivalent of 'sync', I'd always been more than willing to pay the price
> of 'async' vs. 'sync' on the damage recovery front...

I don't have anything on FreeBSD but in the past I ran a number of
throughput benchmarks on a Cray XM/P running UniCos (a 64-bit AT&T flavor of
unix). For our mix of jobs, which included externally simulated user
activity and program builds, the benchmarks established that that we could
double throughput by caching the writes and allowing the system to write
behind. The reads were much more important because a job was waiting for
input and any significant delay could cause it to be swapped out. The
swapping increased the load artificially since we were running in excess of
60% cpu utilization. The jobs running at the tail end of the benchmark were
the ones most affected by delays. The HD system was a DD40, which was
capable of 20MB/s continous to each set of stripped disks (4x9.6MB/s HD's
operating in parallel). The I/O subsystem was accessed via either a 100MB/s
or 1000MB/s data channels. The data channel used for the disk I/O was
something only Cray knew.  The writes, in all cases, were not important to
the running jobs and could be delayed with no side effects. The vector
processing capability of the system did not significantly help our mix of
jobs.

The argument of sync versus async on the writes only matter if you have a
flaky system or are doing online database transactions. The delay for the
writes are only a couple of seconds. That is why you used to "sync" disks at
shutdown. The one funny example we had was when someone typed
"sync;sync;sync<cr>" and turned power off. The system was in the middle of
sync'ing when the power disappeared. It was explained to him that each sync
was to have a <cr> and the commands were to be entered by hand. This
provided sufficient delay to clear the buffers. The current set of shutdown
scripts eliminate that timing problem.
-- 
Kent Stewart
Richland, WA

To Unsubscribe: send mail to majordomo@FreeBSD.org
with "unsubscribe freebsd-questions" in the body of the message



Want to link to this message? Use this URL: <https://mail-archive.FreeBSD.org/cgi/mid.cgi?36C88049.A423185A>