Re: [squid-users] throughput limitation from cache

Hi Peter,

Also sprach Peter Smith <peter.smith@utsouthwestern.edu> (Thu, 12 Jan
2006 18:45:49 -0600):
> Peter Smith wrote:
> > Richard Mittendorfer wrote:
> >> When downloading a cached file from the local squid, I just get
> >about > 250 - 280kB/s. Even on localhost. Is this a limitation with
> >diskd > serving files from cache or some intern limit? I also tried
> >aufs, but > didn't get a better rate. I found a thread here about
> >this, but it got > more into a diskd/aufs discussion :) and didn't
> >provide a solution or > explains it. It doesn't look like it's
> >disk/system related, > vmstat/top doesn't show
> >> high load.
> >>
> >> linux 2.6 celeron 600MHz
> >> squid-2.5.STABLE12 Debian GNU/Linux
> >> 2 x cache_dir diskd 1G 8 128 on 2 x SCSI reiserfs
> >> 32M cache_mem, acls, .. need more info?
> >
> > I too have experience this. I believe this is because Squid has
> > built-in safeguards to keep users from flooding the disk and to even
> > the load across all disks for all users. Another reason is Squid
> > handles all requests in sequential order--client requests, server
> > requests, and memory/disk cache reads/writes.. It is like a
> > virtualized network handler or VM. It really doesn't want one client
> > to kill all others..

That would explain it.

It's even if I'm the only client and it's one big file that's retrieved,
so it must be some kind of internal limit. I have to look into the
source, maybe I can find it hardcoded somewhere. 256kB/s looks so
artificial ;)

In this particulae case I use squid to serve bigger cached objects
(.deb's, kinda apt-cache) to a few clients. It's not horrible slow --
I'm fine with 260kB/s.

I was just wondering what happens if you're on a faster line. Wouldn't
squid slow things down too much then?

> Btw, it still does have a lot to do with diskd/aufs... Read here
> http://squid-docs.sourceforge.net/latest/html/x220.html#AEN318

Had a look at it. Doesn't look like debian's squid is compiled with
async-io. Have to try if I got time for it. However, I'm quite sure it
won't help here. Thx for the docs link - didn't know it.

..hmm - <coffee> - sure, debian's is async-io. Must be. aufs _is_
compiled in: --enable-storeio=ufs,aufs,diskd,null

> Peter

thx a lot, ritch
Received on Thu Jan 12 2006 - 18:49:40 MST