Re: [squid-users] Veritas filesystems and Squid

From: Robert Collins <robert.collins@dont-contact.us>
Date: 18 Aug 2001 11:42:38 +1000

On 17 Aug 2001 18:53:59 -0500, Joe Cooper wrote:
> Wojciech Puchar wrote:
>
> >>Also, in reading the documentation on the web, I saw somewhere
> >>recommended that RAID5 *not* be used as the target disk(s) for
> >>the cache. In trying to maintain uptime, should a disk go bad,
> >>using RAID5 would give us a WIN situation. What are the reasons
> >>to NOT use RAID5, and do you think they outweigh keeping a
> >>
> >
> > RAID5 have sloooow writes.
>
> Wrong. RAID5 actually has quite fast writes (under most workloads, they
> are faster than single disk writes

But_ slower_ than striped-no-parity (RAID 0) for the same N disks. A
raid 5 write (say 4k with a stripe with of 64kb) involves
read the parity stripe
read the data stripe the 4 k is going into
backout the current data stripe from the parity
'write' the 4k to the data stripe
add the data stripe to the parity
write the data stripe to disk
write the parity stripe to disk.

which is, in physical I/O's - count em - 2 reads (128Kb total), and 2
writes (128Kb total).
A raid 0 write requires...
write the 4kb to disk.

> --though not for Squid). It's the
> reads that are a problem for Squid. Seeks are the primary concern for
> Squid workloads, and because RAID5 divides data across multiple disks,
> seeks have to be performed up to N times for each object (where N is the
> number of disks in the array).

Good point - most cache objects would be under 64kb though right? in
which case a large cluster size would minimize this effect. Also all
good controllers independently seek the disks, so you'll just have a lot
of nice elevator seeks going on - and multiplexing your I/O requests
becomes very important. Aufs a must.

> Writes under a Squid workload on a RAID5 array would not be
> significantly slower--and probably would be faster on a high end RAID
> controller with a nice chunk of NVRAM.

The best use for that high end raid controller for squid is as JBOD
disks. That way squid can simply have a cache_dir taken off line if when
a particular physical disk dies. Writing on that should be up to 4 times
faster than with RAID 5 on the same setup.
allocating a couple of disks for a RAID 1 system and log disk would be
useful too.

> This has all been discussed at length in the past here...and well worth
> searching for, for those who don't understand and would like to.

Absolutely!
Rob
Received on Fri Aug 17 2001 - 19:42:23 MDT

This archive was generated by hypermail pre-2.1.9 : Tue Dec 09 2003 - 17:01:43 MST