Re: vm.bufcache and git

[Adam Hamsik <haaaad%gmail.com@localhost>]

On May,Wednesday 26 2010, at 7:42 PM, Greg Troxel wrote:

> 
> Increasing kern.maxvnodes indeed helps; obviously that's about not
> having to recycle vnodes per stat.  Since the basic issue is that stat
> latency is greater for 180K files than a few, normal microbenchmarks
> don't help.
> 
> I wrote a short program to run stat on many files, and to loop on
> subsets.  It outputs data in xplot format, plotting stat latency in us
> vs log2 of number of files.  To use it, create FILES containing one
> filename per line (I used "find . -type f | egrep -v /.git", and yes I
> know I should have quoted the . ...) and then
> 
>  statbench < FILES
> 
> On a system with big enough bufcache and maxvnodes, I found stat to take
> about 4 us up to about 32 files, and then it starts to grow, reaching
> 9-10 us around 4096-8192 files, and then it remains relatively constant.
> 

How many times have you repeated this test. Maybe there was some external 
influence (cron, other program etc.. ) who mixed with your results.

Regards

Adam.