Sam, 

   I had some updates from IBM last week. They reproduced the problem and found it only happens when the number of MPI ranks is non-power-of-2.  Their advice is that since the IBM BG/Q optimized collectives themselves are mostly designed only to be helpful for blocks with power-of-2 geometries,  you can try in your program to see if subsequent collective calls with PAMID_COLLECTIVES=1 are actually faster than PAMID_COLLECTIVES=0 on comms with a non-power-of-2 geometry. If the answer is no, then you can just run with PAMID_COLLECTIVES=0 and avoid the dup/split performance issue. Otherwise, IBM may prioritize this ticket.

  Thanks.

--Junchao Zhang

On Thu, Jul 3, 2014 at 4:41 PM, Junchao Zhang <jczhang@mcs.anl.gov> wrote:

Hi, Sam,  

  I wrote micro-benchmarks for MPI_Comm_split/dup. My profiling results suggested the problem lies in a IBM PAMI library call, PAMI_Geometry_create_taskrange().  Unfortunately, I don't have access to the PAMI source code and don't know why. I reported it to IBM and hope IBM will fix it.
  Alternatively, you can set an environment variable PAMID_COLLECTIVES=0 to disables pami collectives. My tests showed it at least fixed the scalability problem of Comm_split and Comm_dup.

  Also through profiling, I found the qsort() called in MPICH code is actually using the merge sort algorithm in Mira's libc library.
  

   

--Junchao Zhang

On Sat, May 17, 2014 at 9:06 AM, Sam Williams <swwilliams@lbl.gov> wrote:

I've been conducting scaling experiments on the Mira (Blue Gene/Q) and K (Sparc) supercomputers.  I've noticed that the time required for MPI_Comm_split and MPI_Comm_dup can grow quickly with scale (~P^2).  As such, its performance eventually becomes a bottleneck.  That is, although the benefit of using a subcommunicator is huge (multigrid solves are weak-scalable), the penalty of creating one (multigrid build time) is also huge.

For example, when scaling from 1 to 46K nodes (= cubes of integers) on Mira, the time (in seconds) required to build a MG solver (including a subcommunicator) scales as
222335.output:   Total time in MGBuild      0.056704
222336.output:   Total time in MGBuild      0.060834
222348.output:   Total time in MGBuild      0.064782
222349.output:   Total time in MGBuild      0.090229
222350.output:   Total time in MGBuild      0.075280
222351.output:   Total time in MGBuild      0.091852
222352.output:   Total time in MGBuild      0.137299
222411.output:   Total time in MGBuild      0.301552
222413.output:   Total time in MGBuild      0.606444
222415.output:   Total time in MGBuild      0.745272
222417.output:   Total time in MGBuild      0.779757
222418.output:   Total time in MGBuild      4.671838
222419.output:   Total time in MGBuild     15.123162
222420.output:   Total time in MGBuild     33.875626
222421.output:   Total time in MGBuild     49.494547
222422.output:   Total time in MGBuild    151.329026

If I disable the call to MPI_Comm_Split, my time scales as
224982.output:   Total time in MGBuild      0.050143
224983.output:   Total time in MGBuild      0.052607
224988.output:   Total time in MGBuild      0.050697
224989.output:   Total time in MGBuild      0.078343
224990.output:   Total time in MGBuild      0.054634
224991.output:   Total time in MGBuild      0.052158
224992.output:   Total time in MGBuild      0.060286
225008.output:   Total time in MGBuild      0.062925
225009.output:   Total time in MGBuild      0.097357
225010.output:   Total time in MGBuild      0.061807
225011.output:   Total time in MGBuild      0.076617
225012.output:   Total time in MGBuild      0.099683
225013.output:   Total time in MGBuild      0.125580
225014.output:   Total time in MGBuild      0.190711
225016.output:   Total time in MGBuild      0.218329
225017.output:   Total time in MGBuild      0.282081

Although I didn't directly measure it, this suggests the time for MPI_Comm_Split is growing roughly quadratically with process concurrency.




I see the same effect on the K machine (8...64K nodes) where the code uses comm_split/dup in conjunction:
run00008_7_1.sh.o2412931:   Total time in MGBuild      0.026458 seconds
run00064_7_1.sh.o2415876:   Total time in MGBuild      0.039121 seconds
run00512_7_1.sh.o2415877:   Total time in MGBuild      0.086800 seconds
run01000_7_1.sh.o2414496:   Total time in MGBuild      0.129764 seconds
run01728_7_1.sh.o2415878:   Total time in MGBuild      0.224576 seconds
run04096_7_1.sh.o2415880:   Total time in MGBuild      0.738979 seconds
run08000_7_1.sh.o2414504:   Total time in MGBuild      2.123800 seconds
run13824_7_1.sh.o2415881:   Total time in MGBuild      6.276573 seconds
run21952_7_1.sh.o2415882:   Total time in MGBuild     13.634200 seconds
run32768_7_1.sh.o2415884:   Total time in MGBuild     36.508670 seconds
run46656_7_1.sh.o2415874:   Total time in MGBuild     58.668228 seconds
run64000_7_1.sh.o2415875:   Total time in MGBuild    117.322217 seconds


A glance at the implementation on Mira (I don't know if the implementation on K is stock) suggests it should be using qsort to sort based on keys.  Unfortunately, qsort is not performance robust like heap/merge sort.  If one were to be productive and call comm_split like...
MPI_Comm_split(...,mycolor,myrank,...)
then one runs the risk that the keys are presorted.  This hits the worst case computational complexity for qsort... O(P^2).  Demanding programmers avoid sending sorted keys seems unreasonable.


I should note, I see a similar lack of scaling with MPI_Comm_dup on the K machine.  Unfortunately, my BGQ data used an earlier version of the code that did not use comm_dup.  As such, I can’t definitively say that it is a problem on that machine as well.

Thus, I'm asking for scalable implementations of comm_split/dup using merge/heap sort whose worst case complexity is still PlogP to be prioritized in the next update.


thanks
_______________________________________________
To manage subscription options or unsubscribe:
https://lists.mpich.org/mailman/listinfo/devel