John Doe wrote:
> From: Matus UHLAR - fantomas <uhlar_at_fantomas.sk>
>> On 08.12.09 02:41, John Doe wrote:
>>> Yes but, as long as squid does not handle disk crashes gracefully, I am
>>> stuck with RAID...
>> what kind of RAID? for mirrors, you don't need stripe size. Stripes aren't
>> safer than single disks. RAID5 is slow, unless you have hardware RAID.
>
> Hardware RAID (so the OS/squid would not see any disk failure) 10 (or maybe n * RAID 1 to separate the logs/cache_dirs, not sure yet).
> I agree with RAID 1 not needing stripes; but not sure about RAID 10...
RAID is much slower than JBOD.
"hardware RAID" is a RAID implemented in hardware with
battery-backed read and write cache. The type of
hardware with this functionality range from simple
$200 I/O controllers to $ multi-million advanced disk
arrays. The performance varies between acceptable and
amazing.
In case that you have the simple RAID I/O controller,
it is recommended to have as much cache as possible
and to use a stripe size which is larger than the
average object size to reduce I/Os.
Many times, the Squid average object size is 13K so a
stripe size of 16K might work for you.
Note, however, that the average object size varies and you
may calculate it for your site.
Find in cache.log the lines like (when Squid starts)
2009/12/19 06:06:07| Validated 56925 Entries
2009/12/19 06:06:07| store_swap_size = 829416
The average object size for this installation is
829416/56925 or 15K. Since there is a significant
percentage larger than 16K, a stripe size of 32K
is most likely better than a stripe of 16K for this
Squid installation.
In case that you have an EMC DMX the stripe size is
much larger and the performance is amazing due to the
large cache and large number of spindles.
-Marcus
Received on Thu Dec 24 2009 - 12:10:47 MST
This archive was generated by hypermail 2.2.0 : Thu Dec 24 2009 - 12:00:02 MST