Re: [squid-users] Tweaking Squid for speed, not max requests

On Thursday 30 August 2001 04:44 pm, Steve Snyder wrote:
> > Going for raw speed, I'd probably load the box with memory and run the
> > cache_dir on tmpfs (a dynamically sized RAM disk). Since 'disk' reads
> > would be instant, ufs would be the best choice. Of course, like you
> > mentioned in another branch, that doesn't leave the other CPU with
> > much to do. Maybe you could lock the NIC's interrupt to the second
> > processor.
>
> Why is ufs the best choice for accessing a RAM disk?
>
> Thanks for the response.

The major performance issue with ufs is that it blocks until the
read/write completes, so it can't handle other tasks while it waits for
the disk. With a RAM disk, there's no wait involved.

The other two (diskd, aufs) produce latency from the whole queue and
callback structure, plus context switches of varying severity. Squid can
scale higher using them, but each individual request is slightly slower.

        -- Brian
Received on Thu Aug 30 2001 - 17:03:54 MDT