RE: [Fwd: Optimizing Squid performance]

I'll bow out on the tweaking details - I'm not across all the details
enough to be of assistance, but I'm glad my raid 5 explanation made
sense.

Rob

> -----Original Message-----
> From: Athanasios Kouroussis [mailto:akouroussis@unifon.com.ar]
> Sent: Tuesday, 8 August 2000 8:41 AM
> To: Robert Collins
> Cc: Duane Wessels; squid-users@ircache.net
> Subject: Re: [Fwd: Optimizing Squid performance]
>
>
> Thanks for the illustration on RAID 5. That resolved as a not
> to do, what
> else should I tweak to achive my objective?
>
> Athanasios Kourousis
> Unifon - T.C.P. S.A.
> akouroussis@unifon.com.ar
>
> Robert Collins wrote:
>
> > > -----Original Message-----
> > > From: Athanasios Kouroussis [mailto:akouroussis@unifon.com.ar]
> > > Sent: Tuesday, 8 August 2000 8:10 AM
> > > To: Duane Wessels
> > > Cc: squid-users@ircache.net
> > > Subject: Re: [Fwd: Optimizing Squid performance]
> > >
> > >
> > > Thanks for the prompt reply:
> > >
> > > 1) Top or free report 3 of 512Mb of physical memory free.
> > > Only 40Mb of swap
> > > space used.
> > > 2) Why do not use raid if it controlled by hardware. I can
> > > understand not
> > > using software RAID or RAID 5 with a poor controller but the
> > > RAID controller
> > > of this box is a Smart-2 3200. Anyway, what do you
> > > recommend. How should I
> > > set up the partitions (size, fs, etc) for max perf. Any
> > > other, squid related
> > > parameters I should tweek?
> >
> > RAID 5 incurs a read-write overhead regardless of software
> or hardware
> > implementation. It can be up to 1/4 the speed of a striped no-parity
> > configuration.
> >
> > The reason is that for every stripe, a write anywhere to the stripe
> > requires:
> > A) read the stripe block BLOCK-A(up to 64K on most controllers) that
> > contains logical blocks that are being written to.
> > B)read the parity stripe block for that stripe. (same size as A)
> > C) remove the old BLOCK-A from the parity value
> > D) add the new BLOCK-A to the parity value
> > E) write the new BLOCK-A to disk
> > F) write the new parity block to disk.
> >
> > As you can see a single write of 2K (say a small.gif) will require 2
> > head reads (1 each on two of the array disks) and 2 writes
> (back to the
> > same sector).
> >
> > Whereas mirroring or striping no-parity require only two writes
> > (mirroring) or 1 write (no parity striping).
> >
> > So we won't use RAID 5 for high-performance applications.
> >
> > For cache data where you probably don't care about the data itself,
> > having multiple independant disks means that if a failure
> occurs, squid
> > can stop using that disk rather than you losing a full
> stripe, and you
> > don't need to pay the 100% overhead of mirroring.
> >
> > Rob
> >
> > > 3) May that is the reason.. I will check it and let you know.
> > >
> > > Thanks,
> > >
> > > Athanasios Kouroussis
> > > Unifon - T.C.P. S.A.
> > > akourousssis@unifon.com.ar
> > >
>
>
Received on Mon Aug 07 2000 - 18:55:12 MDT