RE: [squid-users] anyone compared comparison of reiserfs vs ext3 on linux recently ?

An informal observation: I think cache size would likely be a factor too,
but likely in the extreme case only. I was running another application that
stores many small files districtued across many nested directories (bushy
filesystem organization, similar to Squid, but likely a more extreme case)
and found that with high hundreds of thousands or millions of files, the
system (Dual XeonDP 2.0, 4GB Ram, RAID 10 across 8 10k U160 disks) slowed to
the point the system was largely unresponsive when accessing anything in
that filesystem (there were 1.2 million files before I decided I needed to
try ReiserFS instead, though I haven't gotten to the point of comparing at
this scale yet). With this many files, accessing metadata seemed
problematic. Even doing an ls in the root of this fs was painful. It may
well be though, that at even the most reasonable maximum size you would ever
see a cache, ext3 may scale just fine.

By way of background reading, I have found Von Hagen's _Linux Filesytems_
book a rather good, fairly up-to-date read, at least perhaps a good guide to
making educated guesses in this case.

Sean

-----Original Message-----
From: Joe Cooper [mailto:joe@swelltech.com]
Sent: Thursday, January 02, 2003 12:45 PM
To: Mika Aleksandroff
Cc: squid-users@squid-cache.org
Subject: Re: [squid-users] anyone compared comparison of reiserfs vs
ext3 on linux recently ?

Mika Aleksandroff wrote:
> On Thu January 2 2003 21:04, Mike wrote:
>
>>Can anyone proffer an opinion about the merits of using reiserfs over
>>default linux filesystems ext3 for a squid server ?
>
>
> Joe says reiserfs rules. Joe rules. Trust Joe. :-)
>
> http://list.cineca.it/cgi-bin/wa?A2=ind0203&L=squid&P=R40022

Thanks for the vote of confidence, Mika. But I will temper it by noting
that Duane obtained different results in some benchmarks a few months
ago, showing that ReiserFS was slower on his hardware than ext2 or ext3.

I suspect CPU:disk ratio to be the determining factor, because ReiserFS
is harder on the CPU than ext2 or ext3. Duanes test machines had plenty
of disks (6 x 10k RPM, I think), but not a lot of CPU horsepower (2 x
500MHz, maybe?), whereas my machines are more likely to have fewer disks
(always 3 or fewer, as that's all our 1U chassis will support--I don't
keep any 2U systems on hand, as they are too expensive to stock) and
more CPU (always more than 1GHz these days). So, if CPU is your
limiting factor, ext2 or ext3 might be a better option. AUFS is also
very CPU hungry, so it will behave differently on a slower CPU with more
disks.

I still choose ReiserFS for all of my cache_dirs, because I know it is
fast and extremely reliable for Squid workloads, and my experiences with
ext3 for Squid have always been negative. All of the benchmarks I've
done on ext3 have ended in kernel oopses, though I assume those problems
have been fixed by now. My tests were done on kernel 2.4.9 (plus all of
the Red Hat patches at RPM release -31 and a few others), so I'm sure a
year+ of development has fixed those issues. ext2 is not an option for
me, as the fsck time and flakiness of an unjournalled FS would be
extremely problematic for our clients.

So, as always, if you /must/ know for sure which is faster, then you'll
need to test it on your own hardware to see. Most folks will see no
difference one way or another, as modern hardware is extremely oversized
for most Squid installations. You simply can't buy or build a Squid
machine with modern components that is capable of supporting less than
two full T1 links with plenty of headroom to spare.

-- 
Joe Cooper <joe@swelltech.com>
Web caching appliances and support.
http://www.swelltech.com