A COMPARISON MPI vs POSIX Threads
Overview MPI allows you to run multiple processes on 1 host How would running MPI on 1 host compare with POSIX thread solution? Attempting to compare MPI vs POSIX run times Hardware Dual 6 Core (2 threads per core) 12 logical Intel Xeon CPU E5 – 2667 (show schematic) 2.96 GHz 15 MB L3 Cache All code / output / analysis available here:
Specifics Going to compare runtimes of code in MPI vs code written using POSIX threads and shared memory Try to make the code as similar as possible so we’re comparing apples with oranges and not apples with monkeys Since we are on 1 machine the BUS is doing all the com traffic, that should make the POSIX and MPI versions similar (ie. The network doesn’t get involved) Only makes sense with 1 machine Set up test bed Try each step individually, check results, then automate Use Matrix Matrix multiply code we developed over the semester Everyone is familiar with the code and can make observations Use square matrices Vary Matrix sizes from 500 -> 10,000 elements square (plus a couple of big ones) Matrix A will be filled with 1-n Left to Right and Top Down Matrix B will be the identity matrix Can then check our results easily as A*B = A when B = identity matrix Ran all processes ie. compile / output result / parsing many times and checked before writing final scripts to do the processing
Matrix Sizes MATRIX SIZENUM ELEMENTSLOOP CALCULATIONS N multiplies N-1 Adds E E E E E E E+12 Third Column: Just the number of calculations inside the loop for calculating the matrix elements
Specifics cont. About the runs For each MATRIX size (500 -> 3000,4000, 5000, 6000,7000,8000,9000,10000) Vary thread count 2-12 (POSIX) Vary Processes 2-12 (MPI) Run 10 trials of each and take average (machine mostly idle when not running tests, but want to smooth spikes in run times caused by the system doing routine tasks) Make observations about anomalies in the run times where appropriate Caveats All initial runs with no optimization for testing, but hey this is a class about performance Second set of runs with optimization turned on –O1 ( note: -O2 & -O3 made no appreciable difference) First level optimization made a huge difference > 3 x improvement GNU Optimization explanation can be found here: Built with just the –O1 flags to see if I could catch the “one” making the most difference (nope) (code isn’t that complicated) Not all optimizations are flag controlled Regardless of whether the code is written in the most efficient fashion (and it’s not) because of the similarity we can make some runs and observations Oh No moment ** Huge improvement in performance with optimized code, why? What if the improvement in performance ( from compiler optimization) was due to the identity matrix? Came back and made matrix B non Identity, same performance. Whew. I now Believe the main performance improvement came from loop unrolling. Maybe the compiler found a clever way to increase the speed because of the simple math and it’s not really doing all the calculations I thought it was? Came back and made matrix B non Identity, same performance. Whew. Ready to make the runs
Discussion Please chime in as questions come up. Process Explanation: (After initial testing and verification) Attempted a 25,000 x 25,000 matrix Compiler error for MPI (exceeded MPI_Bcast 2 GB limit on matrices) Not an issue for POSIX threads (until you run out of memory on the machine) swap Settled on 12 Processes / Threads because of the number of cores available Do you get enhanced or degraded performance by exceeding that number? Example of process space / top output (10,000 x 10,000) Early testing, before runs started. Pre Optimization
Time Comparison (Boring)
Time Comparison (still boring…) In all these cases time for 5,4, 3, 2 processes much longer than 6 so left of for comparison MPI Doesn’t “catch” back up till 11 processes POSIX Doesn’t “catch” back up till 9 processes
MPI Time Curve
POSIX Time Curve
POSIX Threads Vs MPI Processes Run Times Matrix Sizes 4000x4000 – 10,000 x 10,000
POSIX Threads 1500 x 1500 – 2500x2500
1600 x 1600 case Straight C runs long enough to see top output (here I can see the memory usage) threaded,MPI, and non mp code share same basic structure for calculating “C” Matrix Suspect some kind of boundary issue here, possibly “false sharing”? Process fits entirely in shared L3 cache 15 MB x 2 = 30MB Do same number of calculations but make initial array allocations larger (shown below) ~/SUNY]$ foreach NUM_TRIALS ( ) foreach?./a.out foreach? End Matrices (1600x1600) Size Allocated (1600 x 1600) : Run Time secs Matrices (1600x1600) Size Allocated (1600 x 1600) : Run Time secs Matrices (1600x1600) Size Allocated (1600 x 1600) : Run Time secs Matrices (1600x1600) Size Allocated (1600 x 1600) : Run Time secs Matrices (1600x1600) Size Allocated (1600 x 1600) : Run Time secs ~/SUNY]$ foreach NUM_TRIALS ( ) foreach?./a.out foreach? End Matrices (1600x1600) Size Allocated (1601 x 1601) : Run Time secs Matrices (1600x1600) Size Allocated (1601 x 1601) : Run Time secs Matrices (1600x1600) Size Allocated (1601 x 1601) : Run Time secs Matrices (1600x1600) Size Allocated (1601 x 1601) : Run Time secs Matrices (1600x1600) Size Allocated (1601 x 1601) : Run Time secs ~/SUNY]$
Future Directions POSIX Threads with Network memory? (NFS) Combo MPI and POSIX Threads? MPI to multiple machines, then POSIX threads ? POSIX threads that launch MPI ? Couldn’t get MPE running with MPIch (would like to re-investigate why) Investigate optimization techniques Did the compiler figure out how to reduce run times because of the simple matrix multiplies? <- NO Rerun with non-identity B matrix and compare times <- DONE Try different languages ie CHAPEL Try different algorithms Want to add OpenMP to the mix Found this paper on OpenMP vs direct POSIX programming (similar tests) For < 6 processes look at thread_affinity and assignment of threads to a physical processor