JAMSTECR, October MPI Message Passing Interface JAMSTEC NEC SX- IBM RS /SP PC MPI MPI_SENDRECV SX- SP PCC MPI MPI, Performance of MPI on parallel comp

Size: px

Start display at page:

Download "JAMSTECR, October MPI Message Passing Interface JAMSTEC NEC SX- IBM RS /SP PC MPI MPI_SENDRECV SX- SP PCC MPI MPI, Performance of MPI on parallel comp"

まさとしかに
4 years ago
Views:

1 JAMSTECR, October MPI Message Passing Interface JAMSTECNEC SX- IBM RS/SPPC MPI MPI_SENDRECVSX- SP PCC MPI MPI, Performance of MPI on parallel computers in JAMSTEC Hideaki SAITO Kazushi FURUTA Jun NAOI MPI Message Passing Interface is the interface agreement of communication library which aimed at carrying out parallel processing. Vendors implement MPI adapted to the parallel computers with which they deal. This research shows MPI library's characteristics, performance for Himeno Benchmark Test, scalability and efficiency in parallel computing for test programs with NEC SX- /, IBM RS/SP in JAMSTEC and PC cluster parallel computer of our own assembling. Some instructive information for efficient parallelization is obtained by evaluating scale and structure of a serial processing program and its estimating amount of computation and communication based on this store of knowledge.( ) On those platforms, MPI_SENDRECV that can transfer data at one call indicates the highest bandwidth than any other one-to-one routines of MPI. ( ) SX- is especially effective for executing the programs that need to transfer big data and can be vectorized. ( ) SP shows a good performance in executing the programs that range from small-scale to middle-scale. ( ) PCC at the present is suitable for processes that need relatively small datatransferring. (i.e. the coarse-grained process) Along the above information, we are planning to parallelize an Ocean Global Circulation Model program with MPI and to carry out the realistic computation. Key Words : MPI, Parallelization, Scalability, Parallel efficiency Computer and Information Department, Japan Marine Science and Technology Center

2 MPI Message Passing Interface MPI MPIPVM MPI Forum MPI MPP Workstation Cluster etc MPI NEC SX- / SX- IBM RS/ SP SP MPI MPI PC PCPCC MPICH MPISX-SP PCC MPI MPI MPI MPI MPI PE PE PE PE SX- CPU MMU Fortran/SX Fortran/SX MPI SX- CPU MMU. GB/sec CPU-MMU CPU GB/sec CPU CPU PC Ethernetip SPUS MB/ sec MPIIBM MPE IBM MPL cluster JAMSTECR,

3 SPSP PC MPI_WAIT MPI_RSEND vs MPI_RECV MPI_SEND vs MPI_RECV MPI_SENDRECV MPI_SSEND vs MPI_RECV MPI_BSEND vs MPI_RECV MPI_ISEND vs MPI_RECV ByteRANK MPI MPI kbmb SX-SPMPI_SSEND MPI_BSEND PCCMPI_SSEND MPI_BSEND SX- SPPCC RANK MPI SX- MPI_ SENDRECV, MPI_SEND vs MPI_RECV kb MPI_SEND vs MPI_RECV MPI_SENDRECV SPMPI_SENDRECV SP Table Summary of each platform SPEC SX- CPU GB CPUGFLOPS 1CPU GB/sec OS: SUPER-UX R : FORTRAN90/SX version MPI: MPI/SX version SP POWER SuperChip MHz 1 MB CPUMHz MB/sec OS: AIX version : HPF version MPI: MPI IBM version PCC PentiumIII MHz MHz MB/sec OS:RedHat Linux kernel : g version MPI: MPICH version JAMSTECR,

4 MPISX- Fig. Bandwidth with MPI :Total SX- MPISP Fig. Bandwidth with MPI :Single SP MPISX- Fig. Bandwidth with MPI :Single SX- MPIPCC Fig. Bandwidth with MPI :Total PCC MPISP Fig. Bandwidth with MPI :Total SP MPIPCC Fig. Bandwidth with MPI :Single PCC PCC MPI_SENDRECV MPI_SEND vs MPI_RECV MPI_ISEND vs MPI_RECV. MB/sec. MB/sec PE PE SX- MMU SX-SP SX-SPMB JAMSTECR,

5 PCC PCCBase-TX Ethernet ipsp US US User SpaceSP SP Ethernet PCC SP SP PCC SX- MPI_SENDRECV MPI_SENDRECV FLOPS Fortran MPI MPI_SENDRECV SByte, M Byte, LByte SX-SP PCC MPI N/A I/O MPISX- Fig. Himeno Benchmark Test with MPI SX- PoissonJacobi Poisson MPI MPI SX-SP PCCS M L MPISP Fig. Himeno Benchmark Test with MPI SP MPIPCC Fig. Himeno Benchmark Test with MPI PCC JAMSTECR,

6 Table Himeno Benchmark Test with sequential processing SS S M L SX- SP N/A N/A PCC N/A SX-SP PCC MPI SP PCC SX- MPI % SX- SP PCCSP PCC PC MPI MPI MPI SP PCC SP PCC MPI PCC MPI SP PCC SX- MPI % SX- SX- MPIn Tn Sn = T Tn Pn = Tn T Sn n Pn MonteCarlo JAMSTECR,

7 MonteCarlo x y /,,,= - MPI MPISX-,,,, SP,,, PCC,, / +x, MonteCarlo,,, = - MPI MPISX-,,,,SP,,, PCC,, ø t =c ø x + ø y, ø = ø(x, y,t), c = const. ADIJacobi MPI MPISX-,,,,SP,,, PCC,, SX-,,, SP PCC,, ø, ø = ø(x, y) x + ø =0 y SOR Red-Black MPISX-,,, SP,,, PCC,, SX-,,,, SP PCC,, Navier-Stokes xu = u x, y, tyv = v x, y, t p = p x, yre = const. u x + v y =0 u t +u u x +v u p =- y x + Re v t +u v x +v v p = y y + Re u + u x y v + v x y Poisson MAC Marker and Cell Red-BlackSOR Re= MPI SX-,,,,SP,,, PCC,,, MPI_SENDRECV Sn Pn SX-SP PCC SX-. SP SP OS PCC SX-SP PCC SX- Trade-offCPUSP PCC JAMSTECR,

8 SX- Fig. Scalability for calculation SX- SX- Fig. Parallel Efficiency for calculation SX- SP Fig. Scalability for calculation SP SP Fig. Parallel Efficiency for calculation SP PCC Fig. Scalability for calculation PCC PCC Fig. Parallel Efficiency for calculation PCC SX-SP PCC SX- SP. PCC SP SX-SP PCC SX- SPSX- PCC.. MonteCarlo SX-SP PCC. JacobiGauss-Seidel SOR JAMSTECR,

9 SX- Fig. Scalability for solving Diffusion Eq SX- SX- Fig. Parallel Efficiency for solving Diffusion Eq SX- SP Fig. Scalability for solving Diffusion Eq SP SP Fig. Parallel Efficiency for solving Diffusion Eq SP PCC Fig. Scalability for solving Diffusion Eq PCC PCC Fig. Parallel Efficiency for solving Diffusion Eq PCC Jacobi Gauss-Seidel SOR Jacobi Gauss-Seidel Red-Black Jacobi /..Jacobi....Jacobi SOR.. Red-Black..... MPI CPU MPI JAMSTECR,

10 LaplaceSX- Fig. Scalability for solving Laplace Eq SX- LaplaceSX- Fig. Parallel Efficiency for solving Laplace Eq SX- LaplaceSP Fig. Scalability for solving Laplace Eq SP LaplaceSP Fig. Parallel Efficiency for solving Laplace Eq SP LaplacePCC Fig. Scalability for solving Laplace Eq PCC LaplacePCC Fig. Parallel Efficiency for solving Laplace Eq PCC MPI MPI SX- MPI_SENDRECV SP SX- PCC MPI_SENDRECV OS JAMSTECR,

11 SX- Fig. Scalability for solving Fluid Eqs SX- SX- Fig. Parallel Efficiency for solving Fluid Eqs SX- SP Fig. Scalability for solving Fluid Eqs SP SP Fig. Parallel Efficiency for solving Fluid Eqs SP PCC Fig. Scalability for solving Fluid Eqs PCC PCC Fig. Parallel Efficiency for solving Fluid Eqs PCC PCC. MPI_SENDRECV MPI_SENDRECV MPI_SEND SX- JAMSTECR,

12 JAMSTECR, MB MB SP CPU PCC PCC MPI MPI PCC PC Base-TX Ethernet SP MPI MPI

卒業論文

卒業論文 PC OpenMP SCore PC OpenMP PC PC PC Myrinet PC PC 1 OpenMP 2 1 3 3 PC 8 OpenMP 11 15 15 16 16 18 19 19 19 20 20 21 21 23 26 29 30 31 32 33 4 5 6 7 SCore 9 PC 10 OpenMP 14 16 17 10 17 11 19 12 19 13 20 1421