Re: Process size..

Indeed. My metrics come from Linux, exclusively. From what I hear,
they're not invalid in some BSD's either. For Solaris it may be that the
algorithms for cache_mem actually _do_ perform better than the
file-system buffering algorithms. Whole different ball game out there.

D

Jason Wang wrote:
>
> Hi there. I've been following this thread with some interest, as it
> seems
> to go contrary to what has worked best for me.
>
> I am running Squid 2.2_STABLE5 on a dual-processor Ultra60 with 1GB of
> memory. I tried different cache_mem settings, and the one that has given
> me the greatest performance (ie. lowest avg transaction time and highest
> cache hit rate) was a setting of "cache_mem 384 MB". It swaps about
> 100MB
> to HDD, but that doesn't appear to hurt performance too much. Here is
> how
> the system is configured:
>
> Squid 2.2_STABLE5 on Sun Ultra 60 with 1GB RAM and 2x 450MHz CPU's
> running Solaris 2.7 w/ latest Sun patches
> cache_mem 384 MB
> cache_dir /cache1 10000 48 256 ( Cache partitions are on )
> cache_dir /cache2 10000 48 256 ( separate hard drives, no )
> ( RAID. )
>
> After building up about 15GB of cache (on disk), the performance now
> looks
> like this:
>
> Requests per day: ~1 million
> Uniq hosts per day: ~5000
> Cache hit rate: 49.1 %
> Avg transaction time - cache hit: 660 ms
> Avg transaction time - cache miss: 960 ms
>
> With smaller cache_mem settings, the cache hit rate went down to about
> 42%,
> and the avg transaction times for hits and misses went up to 900ms and
> 1000ms
> respectively. I actually have two of these machines which I test out
> different
> settings on. So far, the line of diminishing return for me has been a
> cache_mem
> setting of 384MB. The response times may also be affected by the URL
> filtering
> software running on the machine, but that should affect both machines
> equally.
>
> Can you recommend any settings for this environment? With the default
> cache_mem
> setting at 16MB, the system responded very slowly (upwards of 1500ms
> when it did
> hit cache, vs about 1200 ms when it went direct), and the hit rate was
> pretty
> lousy (about 30%). If not, can you help me understand why I am getting
> these
> contrary results.
>
> Thanks,
> Jason
>
> Dancer wrote:
> >
> > Marc-Adrian Napoli wrote:
> > >
> > > Dave,
> > >
> > > > > From: Marc-Adrian Napoli [SMTP:marcadrian@cia.com.au]
> > > > >
> > > > > cache_mem 500 MB
> > > > >
> > > > The normal reccomendation is to leave this at the
> > > > default 8MB; what's your reasoning in using such
> > > > a large value?
> > >
> > > 1GB of RAM.
> > >
> > > Is that a good enough reason? (honestly).
> >
> > Nope. Squid's memory cache is probably not as efficient as your
> > filesystem's buffer-cache. For linux boxen, at least, I've narrowed the
> > optimal size to between 5 and 15MB. Got to find time to narrow it
> > further, if possible, and to check a few more kernels.
> >
> > D
Received on Thu Nov 11 1999 - 18:21:53 MST