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

On Fri, Aug 31, 2001, Brett Lymn wrote:
> According to Brian:
> >
> >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.
>
> Errrrmmmmm.... not what I understood UFS did. The _metadata_ (file
> create, directory update) is a blocking wait for disk operation sort
> of call but normal reads and writes are done via the buffer cache. So
> if you have enough memory the writes will just go to memory and get
> flushed to disk later, the read cache has read-ahead which is
> controllable so you can bring in disk blocks to fulfill future read
> requests. You can decouple the metadata updates by either using a
> logging file system (which does have downsides - especially on large
> deletes) or log the metadata to another disk which you can do with
> Sun's Solstice Disk Suite just to name one product.

Yes, but you can only complete one file operation at a time.
Thats the issue - the diskd/aufs filesystems can schedule multiple
file operations without blocking squid. UFS can only schedule one,
and you have to wait for it to finish regardless of whether the
OS supplies it from the buffer cache or whether the OS has to
initiate disk IO.

Adrian
Received on Thu Aug 30 2001 - 18:34:36 MDT