Stewart Forster wrote:
> I've done some research on this. Squid would work well with a 4K
> frag/ 8K chunk filesystem. Basically we want to write a stripped
> down version of UFS that allows indexing by inode alone. I have a
> design here at Connect that works out to an average of 2.5 disk
> accesses per new object pulled in, and 1.7 disk accesses per cache
> hit, based on a sample taken from our live caches.
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.
I am not convinced that using a user-space FS is the right direction
for Squid. If you solely look at performance then it is but if you
add complexity, flexibility and crash recovery then it is not so
clear anymore.
What we rather need is modularity and documented module API's to
make it possible to switch parts of the design to another
implementaion, like replacing the disk store manager with a custom
filesystem model or whatever.
/Henrik
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