On Mon, Jun 15, 2009 at 12:06:39PM +0200, Kristian Nielsen wrote:
> Stewart Smith <stewart@stripped> writes:
> > Really simple patch to get the CPU count out of an OSX box for automatic
> > --parallel setting for mtr2.
> Actually, in my experience, --parallel gives huge speedups even on a
> single-processor machines. Most of the time seems to be spent in sleep(), not
> in burning CPU :-(.
> So I would think it makes more sense to set parallelism on available memory?
> (Assuming running in ram disk, if not disk trashing becomes the limit under
> higher --parallel values).
> But maybe this is different in Drizzle, haven't run Drizzle tests in a while.
A followup based on our conversation on IRC the other day:
- libeatmydata gives near --mem performance without (as bad) thrashing.
- the speedup for Drizzle with libeatmydata and --parallel is much more
dramatic than for MariaDB.
Should we throw libeatmydata in the tree (or package it and have it
installable) and in mtr2 find it and if we do, load it in mtr2?
I use it now by just LD_PRELOAD and ./test-run.pl - however this creates
issues on (yes, you guessed it) Solaris due to different 32/64bit etc.
So an advantage of throwing it in the tree would be we could build it
with the same damn options on Solaris and just load it for our binaries
(not, e.g. perl).