inn performance

[Mark Hittinger]

> In general am I going to see better performance with, say, 128 2GB - 1
> buffers and no largefile support or, say, 16 16GB buffers with largefile
> support in the latest betas?

For me (solaris 2.6/veritas) cnfs artwrites in both non-largefile and
largefile environments tend to hover at around 10% of innd time.

The area where I think we need work is on the overview side, where the
different techniques still seem to hover around 30% of innd time.  I
think there is more juice to be squeezed from the overview side of things
than the cnfs side.

In largefile mode I have noticed some performance degradation in artwrites 
beyond the triple indirect block - I need to write some modifications for the 
cnfs code to allow large raw partitions under solaris (reads/writes need to be
in 512 byte blocks on the largefile raw devices).

The other side of that one is that we loose some file system caching effect
for the readers that would use those raw devices.

In any event overview is where you are going to see your largest performance
hit and/or benefit so thats where you should concentrate your attention for
now.

Later

Mark Hittinger
Earthlink!Mindspring!Netcom!Dallas
msh at freebsd.org
bugs at netcom.com