I did one "chapter" of it on Jeff's blog and probably should complete it. Definitely need to update the FAQ for the new options.
Sadly, outside of that and the mpirun man page, there isn't much available yet. I'm woefully far behind on it. On Feb 26, 2014, at 4:47 PM, Saliya Ekanayake <esal...@gmail.com> wrote: > Thank you Ralph, this is very insightful and I think I can better understand > performance of our application. > > If I may ask, is there a document describing this affinity options? I've been > looking at tuning FAQ and Jeff's blog posts. > > Thank you, > Saliya > > > On Wed, Feb 26, 2014 at 7:34 PM, Ralph Castain <r...@open-mpi.org> wrote: > > On Feb 26, 2014, at 4:29 PM, Saliya Ekanayake <esal...@gmail.com> wrote: > >> I see, so if I understand correctly, the best scenario for threads would be >> to bind 2 procs to sockets as --map-by socket:pe=4 and use 4 threads in each >> proc. > > Yes, that would be the best solution. If you have 4 cores in each socket, > then just bind each proc to the socket: > > --map-by socket --bind-to socket > > If you want to put one proc on each socket by itself, then do > > --map-by ppr:1:socket --bind-to socket > > >> >> Also, as you've mentioned binding threads to get memory locality, I guess >> this has to be done at application level and not an option in OMPI > > Sadly yes - the problem is that MPI lacks an init call for each thread, and > so we don't see the threads being started. You can use hwloc to bind each > thread, but it has to be done in the app itself. > >> >> Thank you, >> Saliya >> >> >> On Wed, Feb 26, 2014 at 4:50 PM, Ralph Castain <r...@open-mpi.org> wrote: >> Sorry, had to run some errands. >> >> On Feb 26, 2014, at 1:03 PM, Saliya Ekanayake <esal...@gmail.com> wrote: >> >>> Is it possible to bind to cores of multiple sockets? Say I have a machine >>> with 2 sockets each with 4 cores and if I run 8 threads with 1 proc can I >>> utilize all 8 cores for 8 threads? >> >> In that scenario, you won't get any benefit from binding as we only bind at >> the proc level (and binding to the entire node does nothing). You might want >> to bind your threads, however, as otherwise the threads will not necessarily >> execute local to any memory they malloc. >> >>> >>> Thank you for speedy replies >>> >>> Saliya >>> >>> >>> On Wed, Feb 26, 2014 at 3:21 PM, Ralph Castain <r...@open-mpi.org> wrote: >>> >>> On Feb 26, 2014, at 12:17 PM, Saliya Ekanayake <esal...@gmail.com> wrote: >>> >>>> I have a followup question on this. In our application we have parallel >>>> for loops similar to OMP parallel for. I noticed that in order to gain >>>> speedup with threads I've to set --bind-to none, otherwise multiple >>>> threads will bind to same core giving no increase in performance. For >>>> example, I get following (attached) performance for a simple 3point >>>> stencil computation run with T threads on 1 MPI process on 1 node (Tx1x1). >>>> >>>> My understanding is even when there are multiple procs per node we should >>>> use --bind-to none in order to get performance with threads. Is this >>>> correct? Also, what's the disadvantage of not using --bind-to core? >>> >>> Your best performance with threads comes when you bind each process to >>> multiple cores. Binding helps performance by ensuring your memory is always >>> local, and provides some optimized scheduling benefits. You can bind to >>> multiple cores by adding the qualifier "pe=N" to your mapping definition, >>> like this: >>> >>> mpirun --map-by socket:pe=4 .... >>> >>> The above example will map processes by socket, and bind each process to 4 >>> cores. >>> >>> HTH >>> Ralph >>> >>>> >>>> Thank you, >>>> Saliya >>>> >>>> >>>> On Wed, Feb 26, 2014 at 11:01 AM, Saliya Ekanayake <esal...@gmail.com> >>>> wrote: >>>> Thank you Ralph, I'll check this. >>>> >>>> >>>> On Wed, Feb 26, 2014 at 10:04 AM, Ralph Castain <r...@open-mpi.org> wrote: >>>> It means that OMPI didn't get built against libnuma, and so we can't >>>> ensure that memory is being bound local to the proc binding. Check to see >>>> if numactl and numactl-devel are installed, or you can turn off the >>>> warning using "-mca hwloc_base_mem_bind_failure_action silent" >>>> >>>> >>>> On Feb 25, 2014, at 10:32 PM, Saliya Ekanayake <esal...@gmail.com> wrote: >>>> >>>>> Hi, >>>>> >>>>> I tried to run an MPI Java program with --bind-to core. I receive the >>>>> following warning and wonder how to fix this. >>>>> >>>>> >>>>> WARNING: a request was made to bind a process. While the system >>>>> supports binding the process itself, at least one node does NOT >>>>> support binding memory to the process location. >>>>> >>>>> Node: 192.168.0.19 >>>>> >>>>> This is a warning only; your job will continue, though performance may >>>>> be degraded. >>>>> >>>>> >>>>> Thank you, >>>>> Saliya >>>>> >>>>> -- >>>>> Saliya Ekanayake esal...@gmail.com >>>>> Cell 812-391-4914 Home 812-961-6383 >>>>> http://saliya.org >>>>> _______________________________________________ >>>>> users mailing list >>>>> us...@open-mpi.org >>>>> http://www.open-mpi.org/mailman/listinfo.cgi/users >>>> >>>> >>>> _______________________________________________ >>>> users mailing list >>>> us...@open-mpi.org >>>> http://www.open-mpi.org/mailman/listinfo.cgi/users >>>> >>>> >>>> >>>> -- >>>> Saliya Ekanayake esal...@gmail.com >>>> Cell 812-391-4914 Home 812-961-6383 >>>> http://saliya.org >>>> >>>> >>>> >>>> -- >>>> Saliya Ekanayake esal...@gmail.com >>>> Cell 812-391-4914 Home 812-961-6383 >>>> http://saliya.org >>>> <3pointstencil.png>_______________________________________________ >>>> >>>> users mailing list >>>> us...@open-mpi.org >>>> http://www.open-mpi.org/mailman/listinfo.cgi/users >>> >>> >>> _______________________________________________ >>> users mailing list >>> us...@open-mpi.org >>> http://www.open-mpi.org/mailman/listinfo.cgi/users >>> >>> >>> >>> -- >>> Saliya Ekanayake esal...@gmail.com >>> Cell 812-391-4914 Home 812-961-6383 >>> http://saliya.org >>> _______________________________________________ >>> users mailing list >>> us...@open-mpi.org >>> http://www.open-mpi.org/mailman/listinfo.cgi/users >> >> >> _______________________________________________ >> users mailing list >> us...@open-mpi.org >> http://www.open-mpi.org/mailman/listinfo.cgi/users >> >> >> >> -- >> Saliya Ekanayake esal...@gmail.com >> Cell 812-391-4914 Home 812-961-6383 >> http://saliya.org >> _______________________________________________ >> users mailing list >> us...@open-mpi.org >> http://www.open-mpi.org/mailman/listinfo.cgi/users > > > _______________________________________________ > users mailing list > us...@open-mpi.org > http://www.open-mpi.org/mailman/listinfo.cgi/users > > > > -- > Saliya Ekanayake esal...@gmail.com > Cell 812-391-4914 Home 812-961-6383 > http://saliya.org > _______________________________________________ > users mailing list > us...@open-mpi.org > http://www.open-mpi.org/mailman/listinfo.cgi/users
