[mpich-discuss] Why does MPICH perform busy wait?

Rob Latham robl at mcs.anl.gov
Thu Sep 29 10:08:55 CDT 2016


Your proposal introduces latency.  Data centers want power efficiency, 
but several large classes of applications care about message rate.

you don't need usleeep.  There are several event-based system calls you 
could use: sleect(), poll() -- which does not actually poll in most 
implementations.

Also, there are non-spinning channels you can use: e.g. ch3:sock

It's been something we've been thinking about for a while!
http://trac.mpich.org/projects/mpich/ticket/79


==rob


On 09/29/2016 09:07 AM, Tim Jammer wrote:
> Hi
>
> Within my bachelor thesis I come across that the MPICH Implementation
> implements a busy wait.
> In our application, this leads to a waste of energy as all processes are
> spinning in a barrier while one process performs I/O.
>
> Therefore I want to propse to not wait as busy as it is done now.
> On the test system this saves 50% energy compared to busy wait.
> My basic prototype implementation causes only a negligible additional
> runtime overhead.
>
> My proposal is to sleep between polling if the communication has not
> finished yet.
> Within the first few iterations the implementation should not sleep, to
> guarantee maximum performance.
> But if the communication was not finished right away, the importance of
> saving energy grows.
> Therefore I propose to sleep and increase this sleeping time after every
> polling
> up to a maximum value in order to not waste much additional runtime on
> waiting.
> This will allow the CPU to power down and save energy.
>
> my basic prototype basically works as follows:
>
> int stime=0;
>     while (/* Incoming data does not match the request*/)
>     {
> usleep(stime);
> stime++;
> // do the actual polling...
>
> I can imagine, that the user can choose to activate this behaviour (i.e.
> with an environment variable) if he wants to save energy.
>
> I have attached the section of my thesis which describe it more in detail.
>
>
> My question is: was something like this considered before and are there
> any arguments why this was not implemented.
>
> Regards
> Tim Jammer
>
>
> _______________________________________________
> discuss mailing list     discuss at mpich.org
> To manage subscription options or unsubscribe:
> https://lists.mpich.org/mailman/listinfo/discuss
>



More information about the discuss mailing list