Some mmap observations compared to Linux 2.6/OpenBSD

[Vivek Pai]

David Schultz wrote:
> 
> Okay, I guess SpecWeb99 is ``real world'' enough for me to justify
> the assertion that there is an mmap() performance problem.  Just
> out of curiosity, how many regions did SpecWeb99 map?
> (i.e. what does 'dd if=/proc/$pid/map bs=64k count=1 | wc -l' give?)

The SpecWeb99 benchmark is self-scaling - the higher request rate you
try to achieve, the more files are included in the data set. It's
entirely up to the server to decide how to go about dealing with the
data, and the Flash server used to keep a cache of mmap'd regions.
Files smaller than 64KB were mapped as a single region, while larger
files were broken into 64KB pieces, basically to avoid exhausting the
address space.

For the kinds of throughputs that we achieve, we're dealing with
about 13,000 files. We probably had a number of regions somewhere in
that ballpark, but I don't have a free machine at the moment to give
you an exact number.

Our new server basically gave up on using mapped files entirely due
to the problems with scaling to larger memory sizes - that's why I'm
pushing for the sendfile changes I mentioned in a different mail.

-Vivek