Re: Slowdown due to threads bouncing between HT cores

[Marc Burkhardt <marc@osknowledge.org>]

* Steinar H. Gunderson <sgunderson@bigfoot.com> [2014-10-03 21:44:29 +0200]:

Hi Steinar,

I had a question a while ago: https://lkml.org/lkml/2014/5/25/57

I compiled mpv which is quite small to compile but has a pretty predictable time
spent in the linking phase of the build.

I, however, did not get an answer to this quastion as of today.

As I understand your mail, you problem is quite similar, isn't it?

Thanks,
Marc

> Hi,
> 
> I did a chess benchmark of my new machine (2x E5-2650v3, so 20x2.3GHz
> Haswell-EP), and it performed a bit worse than comparable Windows setups.
> It looks like the scheduler somehow doesn't perform as well with
> hyperthreading; HT is on in the BIOS, but I'm only using 20 threads
> (chess scales sublinearly, so using all 40 usually isn't a good idea),
> so really, the threads should just get one core each and that's it.
> It looks like they are bouncing between cores, reducing overall performance
> by ~20% for some reason. (The machine is otherwise generally idle.)
> 
> First some details to reproduce more easily. Kernel version is 3.16.3, 64-bit
> x86, Debian stable (so gcc 4.7.2). The benchmark binary is a chess engine
> knows as Stockfish; this is the compile I used (because that's what everyone
> else is benchmarking with):
> 
>   http://abrok.eu/stockfish/builds/dbd6156fceaf9bec8e9ff14f99c325c36b284079/linux64modernsse/stockfish_13111907_x64_modern_sse42
> 
> Stockfish is GPL, so the source is readily available if you should need it.
> 
> The benchmark is run with by just running the binary, then giving it these
> commands one by one:
> 
> uci
> setoption name Threads value 20
> setoption name Hash value 1024
> position fen rnbq1rk1/pppnbppp/4p3/3pP1B1/3P3P/2N5/PPP2PP1/R2QKBNR w KQ – 0 7
> go wtime 7200000 winc 30000 btime 7200000 binc 30000
> 
> After ~3 minutes, it will output “bestmove d1g4 ponder f8e8”. A few lines
> above that, you'll see a line with something similar to “nps 13266463”.
> That's nodes per second, and you want it to be higher.
> 
> So, benchmark:
> 
>  - Default: 13266 kN/sec
>  - Change from ondemand to performance on all cores: 14600 kN/sec
>  - taskset -c 0-19 (locking affinity to only one set of hyperthreads):
>    17512 kN/sec
> 
> There is some local variation, but it's typically within a few percent.
> Does anyone know what's going on? I have CONFIG_SCHED_SMT=y and
> CONFIG_SCHED_MC=y.
> 
> /* Steinar */
> -- 
> Homepage: http://www.sesse.net/
> --
> To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
> the body of a message to majordomo@vger.kernel.org
> More majordomo info at  http://vger.kernel.org/majordomo-info.html
> Please read the FAQ at  http://www.tux.org/lkml/
> 

-- 
Marc Burkhardt