Re: squid-2.4stable2 ?

Adrian Chadd wrote:

>
> the diskio branch has the COSS code from HEAD, the sdirno/sfileno types
> modified to be 2gb-happy (don't go over a dirsize of 2000 just yet..)
> and some zero-copy store client patches which I'm going to need to
> develop to get some efficient read code in there.
>
> Don't try to talk to a raw disk just yet either. It will work in theory,
> but it might run like _crap_, if it runs at all.
>
> Good luck!

Ok, about 20 minutes into a run at 40 reqs/sec (single disk, 450MHz CPU,
512MB RAM), disk hits are at 1-3% of overall and 97-99% memory, and
overall hits are at about 3% out of a possible 55%. A run at 70 was
started but didn't make it. SWAPFAIL_MISS is the access.log entry de
jeur here, so I guess we're shedding load voluntarily?

Seems very solid however. No crashes, and no errors being reported, so
I'll call that good work!

The disk file is growing at a nice clip, so I guess it's writing out
stuff, but not reading. We haven't reached the end of the file yet
(1600 MB, we're at 1300 now), so don't know how it will make the turn.
Anyway, looks like we're writing to the disk but not reading at all,
once it's under a bit of load.

I think the load shedding might be a little aggressive here also, since
we have a little CPU to burn, and certainly more disk throughput than
we're using (aufs can do 70 reqs/sec on the same machine). The CPU is
sitting at around 20-30% idle, but I'm seeing a pretty big interrupt
spike on each bytes out burst (shoots from ~1200 up to ~10000 with CPU
idle dropping to <10% during on flush). Definitely need to balance the
shedding in favor of reads, rather than writes.

Will look deeper tomorrow...

Nice work, so far, Adrian. Looks pretty neat.
                                   --
                      Joe Cooper <joe@swelltech.com>
                  Affordable Web Caching Proxy Appliances
                         http://www.swelltech.com
Received on Mon Aug 20 2001 - 00:35:27 MDT