Unmapped I/O

[Konstantin Belousov]

On Fri, Feb 15, 2013 at 01:32:11PM +0100, Dag-Erling Sm??rgrav wrote:
> Konstantin Belousov <kostikbel at gmail.com> writes:
> > I consider the current patch ready to be committed into the HEAD.
> > Some elements of the design were discussed with Jeff Roberson.  The
> > patch was tested by Peter Holm, a backport to stable/9 got load
> > testing by Scott Long.  I see an ~30% reduction in the system time on
> > reading large files over UFS/ahci on the 4-core HTT machine.
> 
> That sounds very nice indeed.  Have you observed a similar reduction in
> buildworld times?

I measured the buildworld times with the version of the patch I had
a month ago. That time, I did not see any statistically significant
difference in either times (user, system, total).

On the other hand, the machine I used to measure is netbooted, and
despite both src/ and obj/ being on UFS over ahci, I still see a steady
and significant amount of the NFS RPC going out, in particular, getattr
RPC. From what I saw, this is due to execing /bin/sh and several other
utilities from the base system over NFS. I believe that it induces the
latency that hides any speed up.

From that moment, there were some bug-fixes related to speed, in particular,
sometimes clustered reads forced the mapping for the cluster buffer, even
if not needed, which could also increase the gain.

I probably do not have time to spend to reevaluate the buildworld
timings, and definitely I do not see much reason to do it on my hardware.
If somebody wants to do the benchmarking, please.
-------------- next part --------------
A non-text attachment was scrubbed...
Name: not available
Type: application/pgp-signature
Size: 834 bytes
Desc: not available
URL: <http://lists.freebsd.org/pipermail/freebsd-arch/attachments/20130215/77251bf5/attachment.sig>