[OE-core] [PATCH 0/3] Pseudo performance changes...

[Richard Purdie]

On Sat, 2013-02-16 at 20:23 -0600, Peter Seebach wrote:
> Unlike most of my submissions, this isn't patches against oe-core; rather,
> it's patches against pseudo, and if I can get some confirmation that they do
> what I think they do, and some review, I'm planning to make this into
> pseudo 1.5, and send a patch "soonish" to merge that into oe-core.
> 
> What this does: Fix a number of build performance issues. By far the
> largest change is actually not so much a problem with pseudo as a problem
> that pseudo can solve by brute force. Packaging systems (at least RPM and
> SMART) do a lot of fsync() and fdatasync() calls. That usually implies
> flushing EVERYTHING that's been written, not just one specific file. And
> that, in turn, results in a severe performance hit.
> 
> So, for instance, on one of my test workstations, this moves a do_rootfs
> with about 1200 RPMs from about 22 minutes to about 4.5. Yeah.
> 
> The other changes aren't as dramatic for that case, but have very significant
> performance impact for at least some workloads. The first is switching to
> using an in-memory database for the files database, dumping it to disk only
> when the pseudo daemon is idle or shutting down. This doesn't produce huge
> benefits in all cases, but for workloads with a lot of parallelism, it can
> produce a very noticeable reduction in how much pseudo slows things down.
> 
> The second is a fairly major protocol change. In short, with this patch,
> pseudo clients only wait for a server response when they need information
> from the server in order to continue. That's OP_FSTAT, OP_STAT,
> OP_MAY_UNLINK, and OP_MKNOD. Everything else just silently assumes that
> it probably succeeded.
> 
> How much does this matter? Between the protocol change and the memory
> DB, a trivial unpack of a tarball (lots of writes to the database, very
> few reads) can be about 4x faster. Removing stuff isn't much faster, but
> it might be a bit faster.
> 
> This is most noticeable, by far, when running more than one build, or
> when running builds while doing other things. It has a much smaller effect
> on builds with no shared state (compile time still dominates that), but
> even there I'm seeing decreases from ~83 minutes to ~64 from just the
> fsync and memory changes. Still waiting on my real test case (multiple
> simultaneous builds which need compiles) completing.

All sounds good to me. I've set the autobuilder away to test this set of
changes.

Cheers,

Richard