Some mmap observations compared to Linux 2.6/OpenBSD

[Vivek Pai]

Take a look at Figure 6, page 9 in the following:
http://www.cs.princeton.edu/~yruan/DeBox/debox.pdf

On a 1GHz box with 1GB of memory, we were spending
4-5 milliseconds per mmap call, and that was limiting
the throughput of our server on SpecWeb99.

Figure 9 on page 11 shows that just getting rid of the
mmap/munmap/mincore calls in this server got us a 50%
performance boost on a fairly complicated workload. The
SpecWeb99 workload was modeled after several web sites,
so this might actually be a performance problem in the
real world.

If you look at figure 11, page 12, you'll see that with
various improvements, our server's median latency dropped to
less than 1ms. An mmap time of several milliseconds would
kill that benefit.

-Vivek

David Schultz wrote:
> Your idea of using a size-hashed freelist as well as a
> location-sorted list is appealing in its simplicity.  Though it
> can cause a bit of fragmentation, it gives you constant time
> lookup.  Bonwick's vmem allocator ([1], section 4.4.2 and
> following), apparently works quite well using this principle.
> 
> But regardless of the approach, someone has yet to demonstrate
> that this is actually a performance problem in the real world. ;-)
> 
> [1] http://www.usenix.org/event/usenix01/full_papers/bonwick/
> _______________________________________________
> freebsd-hackers at freebsd.org mailing list
> http://lists.freebsd.org/mailman/listinfo/freebsd-hackers
> To unsubscribe, send any mail to "freebsd-hackers-unsubscribe at freebsd.org"
>