Re: bisected: futex regression >= 3.14 - was - Slowdown due to threads bouncing between HT cores

[Mike Galbraith <umgwanakikbuti@gmail.com>]

On Wed, 2014-10-08 at 18:45 +0200, Steinar H. Gunderson wrote: 
> On Wed, Oct 08, 2014 at 06:14:18PM +0200, Thomas Gleixner wrote:
> > It looks far more like an issue with the stocking fish code, but hell
> > with futexes one can never be sure.
> 
> OK, maybe we should move to a more recent Stockfish version first of all;
> the specific benchmark was about that specific binary, but for tracking down
> futex issues we can see if more recent code fixes it (the SMP in this thing
> keeps getting developed).

Mine was self build master, downloaded as a tarball.

> I'm moving to 2ac206e847a04a7de07690dd575c6949e5625115 (current head) of
> https://github.com/mcostalba/Stockfish.git, and building with
> “make -j ARCH=x86-64-bmi2”.
> 
> I still don't see any hangs, but I do see the same behavior of moving around
> between CPUs as the older version exhibited. In a test run (using the given
> test script, just with 28 replaced by 20), I get 18273 kN/sec with default,
> and 21875 kN/sec when using taskset.

Sure, that will likely be workqueues or whatnot running when one of the
chess threads wakes, so select_idle_sibling() shows one of its two
faces.. the butt ugly one.  That's what I wanted to play with on a too
darn big socket box when I discovered the thing didn't want to play.

-Mike