Hi there,
I just want to back Henrik's opinion. I think that tuning/optimizing
Squid disk I/O based on traditional "regular file system" paradigms is wrong.
A design must account for such caching specifics as high tolerance to "lost"
data and uneven QoS demands for reads and writes.
A cyclic-like cache with, perhaps, a relatively small "protected" area for
"popular" objects sounds like a reasonable idea worth investigating.
On a side note, elevator-based optimizations are key to a good I/O
performance. Seek and rotation are the bottleneck when it comes to raw disk
I/O. There is a question on how much control over raw disk request order,
raw disk blocks dataplacement, disk caching policies, etc. we have. However,
I believe that an "elevator+" that works 75% of the time is still
significantly better than random I/O.
$0.02,
Alex.
> chunked cyclical multilevel feedback file system
When a [spare time] hacker starts talking like a first-year Ph.D. student
deciding on a thesis title, now that's scary! ;-)
Received on Tue Jul 29 2003 - 13:15:56 MDT
This archive was generated by hypermail pre-2.1.9 : Tue Dec 09 2003 - 16:12:02 MST