On Sat, 22 Aug 1998, Henrik Nordstrom wrote:
> Another thing that may be worth checking out is what the optimal
> disk I/O block size is. The 8K block size currently used is
> probably a bit small as most objects are larger, requiring
> at least two read/write operations.
Not exactly. Most (more than 50%) objects are smaller than 4 KB. Most
swapouts (more than 80%) take just one disk IO from Squid point of view. At
least, this was true for a dozen of [very different] caches we profiled less
than a year ago. I have no reasons to believe that the situation changed
significantly since then.
However, it takes more than one IO (some say up to seven!) for the FS to
complete the swapout request. The overhead comes from inode/file
manipulations that are transparent to Squid (and can be eliminated/reduced by
SquidFS).
Alex.
Received on Tue Jul 29 2003 - 13:15:52 MDT
This archive was generated by hypermail pre-2.1.9 : Tue Dec 09 2003 - 16:11:52 MST