Forward new messages from IBM: * If you set PAMID_COLLECTIVES_MEMORY_OPTIMIZED=1 then the pami optimized collectives are disabled for that comm and the geometry / collective network is not created at comm creation time if the block is irregular, so the comm dup/split performance is good. So for now the user can go ahead and use it, and if there are problems with it we'll need to decide then if we will support it or look at what code changes would be necessary to address the underlying issue. * This whole 'power-of-2' thing was basically too loose a definition for an irregular block, for Blue Gene /Q at the moment there is no crystal-clear formal definition, but a tighter one would be a block that is fully populated with every node in all 5 dimensions, with the same number of MPI ranks on each node --- there are still some exceptions and caveats beyond this that can only be answered by looking at the code, but this is the tightest and simplest definition I can come up with at the moment. On Mon, Sep 22, 2014 at 10:51 AM, Sam Williams <[email protected]> wrote:
All of Mira (48k) is technically not a power of two.
I tried the pami option at one point. I don't recall it giving better performance
- Sam
On Sep 22, 2014, at 8:39 AM, Junchao Zhang <[email protected]> wrote:
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 <[email protected]> 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 <[email protected]> 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