Validation of Dimemas communication model for MPI collective operations Sergi Girona, Jesús Labarta, Rosa M. Badia European Center for Parallelism of Barcelona Departament d´Arquitectura de Computadors Technical University of Catalonia Barcelona, Spain Dimemas Dimemas � Application performance analysis tool for message passing programs � In development since 1992 � On a workstation � Dimemas currently distributed by CEPBA Sergi Girona, EuroPVM/MPI’2000 1
Tuning Methodology Tuning Methodology MP library MP library Tracing Tracing Sequential Sequential Message machine Message machine - - MPI MPI Passing Passing - PVM - PVM Tracing Tracing Code Code - etc... etc... facilities facilities - Parallel machine Parallel machine Dimemas Dimemas Trace Trace Code Code File File modification modification DIMEMAS DIMEMAS Paraver Paraver Visualization Visualization Trace Trace File File Visualization Visualization and analysis and analysis Simulation Simulation Parameters Parameters modification modification Sergi Girona, EuroPVM/MPI’2000 Tracefile Tracefile � Characterizes application � Sequence of resource demands for each task � Sequence of events: communication � Application model Sergi Girona, EuroPVM/MPI’2000 2
Simulated Architecture Simulated Architecture � “Abstract” architecture � Simple/General � Network of SMPs � Fast simulation � Key factors influencing performance � Abstract interconnect � Local/remote latency/BW � Injection mechanism (#links, half/full duplex) � Bisection BW, contention B L L L CPU CPU CPU Local Local Local CPU CPU CPU Memory Memory Memory CPU CPU CPU Sergi Girona, EuroPVM/MPI’2000 System System � Process to processor mapping � Multiprogramming � Tasks sharing node � Different applications B L L L CPU CPU CPU Local Local Local CPU CPU CPU Memory Memory Memory CPU CPU CPU Sergi Girona, EuroPVM/MPI’2000 3
Point to Point Communication Point to Point Communication Size = + T Latency Bandwidth � Latency t a t b t c sender � Bandwidth T 0 t d � Resource contention t e t b t f t g receiver T 1 Sergi Girona, EuroPVM/MPI’2000 Collective Communication Model Collective Communication Model � Barrier � Fan-in/fan-out phases � Size of message � Null/Const/Lin/Log Processor time Block time Comm. time Sergi Girona, EuroPVM/MPI’2000 4
Collective Communication Model Collective Communication Model � Communication time Size = + ∗ Time Latency MODEL_FACT OR Bandwidth � Model factor Model Factor Null 0 Constant 1 Linear P log P C = ∑ 2 = Nsteps steps , steps Logarithmic i i B i = 1 Sergi Girona, EuroPVM/MPI’2000 Parameters Acquisition Parameters Acquisition � Execution of PBM on SGI Origin � Dedicated: execution time � Shared: traces for Dimemas � Compute latency, bandwidth, links, buses, phases, ... � ST-ORM http://www.cepba.upc.es/ST-ORM � Objective: Predicted time with less than 10% error Dedicated D i Parameters >10% m Predicted e m a s Sergi Girona, EuroPVM/MPI’2000 5
System Characterization System Characterization 200 160 Bandwidth 120 80 40 0 t g r v e l l r s e g e e A e n v r c a n c r c h e i g i e u o r C P o u e t h r n a d t a g P d t R t l l B a t g a e A B n g e a h d l g r i n R c l c P s n A l l x i l l A P _ e A E e S c u < 10% error d e R 100 regions 75 Latency 50 25 0 l r t g g e r e A l e s e v e v n n c r c i a g r c h o r P i o u e e u r C n e t t a P d a h d l a g t R A l B B g e t g t e h n a d a n R l r c P i c n l g i s A l l x P e l A _ A l E S e c u d e R Sergi Girona, EuroPVM/MPI’2000 System Characterization System Characterization IN OUT Operation Model Size Model Size Barrier LIN MAX LIN MAX Bcast LOG MAX NULL Gather LOG MEAN NULL Gatherv LOG MEAN NULL � Latency = 25 µ µ seconds Scatter NULL LOG MEAN � Bandwidth = 87.5 MB/s Scatterv NULL LOG MEAN � 1 HD link per node Allgather LOG MEAN LOG MEAN Allgatherv LOG MEAN LOG MEAN Alltoall LOG MEAN LOG MAX Alltoallv LOG MEAN LOG MAX Reduce LOG 2MAX NULL Allreduce LOG 2MAX LOG MAX Reduce_Scatter LOG 2MAX LOG MIN Scan LOG MAX LOG MAX Sergi Girona, EuroPVM/MPI’2000 6
Influence of Buses Influence of Buses SendRecv Exchange 90 170 70 130 90 50 30 50 10 10 1 3 5 7 9 11 13 15 1 3 5 7 9 11 13 15 Reduce_scatter Allgather 150 1200 120 950 90 700 60 450 30 200 1 3 5 7 9 11 13 15 1 3 5 7 9 11 13 15 Sergi Girona, EuroPVM/MPI’2000 Validation Validation � NAS benchmarks � Classes W, A � Size: 8/9, 16, 25/32 40% 30% 20% 10% 0% -10% -20% BT CG FFT IS LU MG SP Sergi Girona, EuroPVM/MPI’2000 7
Conclusions Conclusions � Simple but accurate formulation for collective communication � Methodology for model validation � Dimemas is a valid tool for performance analysis of message passing programs, parallel machines and message passing libraries � Future: RMA and I/O operations pending for validation Sergi Girona, EuroPVM/MPI’2000 8
Recommend
More recommend
