基于MPI的并行計算可視化的研究

1、基于MPI的并行計算可視化的研究 中國地質(zhì)大學（北京）碩士學位論文基于MPI的并行計算可視化的研究姓名：張帆申請學位級別：碩士專業(yè)：計算機應用技術(shù)指導教師：管建和200705011? MPICPU CPUC#? CUP3 ABSTRACTThe technology of modern computer has stimulated the rapid growth of the computational science and the science demand of high performance of computingBut the limit of speed and tech

2、nology has embarrassed the PC to be used widelySingle processing machine can far satisfy the demand of computing resource that the challenging and large-scale computing subject need in many modern fields. One popular and effective method is to use parallel computing to improve parallel computing cap

3、abilityWith the development of the parallel technology and application, the performance analysis and visualization of parallel program is one of the most important parts of parallel computing. In real applications, the gap between the applications real performance and the machines peak performance i

4、s widening. The performance analysis and visualization tool is needed to solve this problem and find out performance bottlenecks quickly by intuitionistically graphics. Then users can develop high performance programs that execute efficiently on high performance computerThere are some internal and e

5、xternal correlative studies. Based on these researches, this paper presents a performance analysis and visualization system frame and brings forward the method of after event to build the frame model. A parallel program performance analysis system based on the frame model is developed to analyze and

6、 visualize performance of parallel programs. The system is used to analyze and visualize parallel programs that use MPI as the parallel support environment. It has good portability and is easy to useThis system mainly analyzes and visualizes the speedup, efficiency and percentage of using CPU. Accor

7、ding to the speedup and efficiency theory computation formula and a performance analysis celebrating method and a practical method to evaluate program execution time and read the percentage of using CPU in the process of parallel computing, we use the program language of C# to design the graphical i

8、nterface of system to visualize performance data. It provides various difference views. It helps to know the status of program running and program communication. It gives references to find out performance bottlenecks and advices to optimize programsTo validate the applicability and validity of the

9、system, the parallel program 4 performance of the Magnetotelluric program using finite element analysis parallel computing is experimented and analyzed by using the analysis and visualization system. It succeeds to analyze the performance, collect the data of program execution time and communication

10、 time and the percentage of using CPU. Then analyze and visualize the performance. Users can find out performance bottlenecks easily. It achieves the purpose of improve program performance Keywords Parallel Computing, Parallel Program, Performance Analysis and Evaluation, Performance Visualization,

11、Finite Elements2_ 1? Parallel Computing? High Performance ComputingSuper Computing1CPU1.1 12Cache32 4? Microkernel5 Prototyping6? I/O? MPPScalable Parallel Computers?PC10Mbps 100Mbps 155Mbps 622Mbps? 1Gbps Murate? PVMPIPVM? MPI PVMPI? 80? MPPPVM DPVMPVM? PRC 31.2 1 2 3? CPU 1 243570 90 2.1.1?211 2 3

12、 4567 4 1 2342 3 48 2 345? I/O host9 2.2 12 3 MPI?52.2.1 HPF40Fortran?HPFF? 1992?31993? 3? FortranFortranHPF Fortran10? SIMD MIMD? HPF1?2 3?4? Fortran? Fortran955CC+? MPI? HPFHPFHPFHPFHPF HPF FortranHPFOpenMPHPF OpenMP DMP? HPF HPF HPF HPF? IBM HP SunHPF PGI? PGHPF APR? Forge xH

13、PF2.2.2 11 PBLASParallel Basic Linear Algebra SubroutinesLAPACK? ScaLAPACK? LAPACKSMPScaLAPACK DMPPBLAS LAPACK ScaLAPACK? PETSc PETSc MPIMATLAB MATLABRTExpress2.2.3? Fortran?C? SMPAPR? Forge Fortran77? SMPDMP? SMP 122.3 MPIMPI? 1994? 5? MPI? MPI2.3.1 MPI?MPI MPIMPI1 MPI?MPIFORTRAN+MPI? C+MPI

14、MPIFORTRAN77/C/Fortran90/C+?/ 2 MPI?MPI?MPI?MPI 3 MPIe MPI?1?MPIMPI?MPI MPI MPI1 234? C Fortran7756? PVM N X E xpress p4? 7 8MPI MPI? MPI 1995? 6?MPI?MPI1.1?MPIMPI? I/OMPI?MPI? 1997 7?MPIMPIMPI-2MPIMPI-1 MPI-2I/O MPI?MPIMPI14? FORTRAN C Fortran? C?MPI-IMPI? Fortran77? C? Fort

15、ran77? C? MPI-1?MPI?MPI-2? Fortran77? CFortran90? C+MPI-22.3.2 MPIMessage?Write? Read? Read?BlockingNon-Blocking exHandleexHandle2.4VisualizationX 162.4.1 1998? /.ormickVisualization in Scientific Computing Visc?Visualization in Scientific Computing Visc2.4.2 12 1734 519?493.1 p?/3-1 pTsTs /

16、p p Ts/ p? p? , pn? p? nn? n? , pnk ,/,pnkpnnnnpn + ? 3-2 100%, pnp20n ,/,pnkpnnpnnpn + ? ,pnpknnpnnpn + ? 3-3? 1,0 pn3.2 Amdahl?3-2? 0, pnk pnnnnpnkpnnnnpn/,/, ? + 3-4 f? / nnnf ? + pnnnnpn/, ? + pnnnnpn/, ? +pfnnfnpn/1/1/, ?+ pffpn/1/11/1, ?+? pffpn/11, ?+? 3-5 Amdahl? f? Amdahl? , pn1/fAmdahl, pn

17、kAmdahl?4 213.3 Gustafson-BarsisGustafson-Barsis AmdahlGustafson-Barsis Amdahl? Amdahlf Gustafson-Barsis s4? s 1-spnnns/?+ pnnpns/1 ?+ spnnn / ? + pspnnn 1/ ?+ ?pnnnnpn/, ? +pnnpsspnnpn/1/, ? +?+ psspn 1, ?+? spppn 1, ?+? 3-6Gustafson-Barsiscpu? cpu3.4 Karp-FlattAmdahlGustafon-Barsis? , pnk Karp-Fla

18、ttP,/, pnkpnnpnT + ?22 pnnnT /1, ? +e 1,/, nTpnkne + enTpnkn 1, + penTenTpnT /11,1, ?+? , pn ,/1, pnTnT,1, pnTnT pepnTepnTpnT /1, ?+ pee /11 ?+? pee /1/1 ?+? pepe /1/1 ?+? ppe /1/11/1 +? ppe/11/1/1? 3-7? , pnk3.5?Amdahl?Gustafon-Barsis?3-2,pnpknnpnnppn + ? 23,1,pnpknpnnnnppn +?+ ? 3-8 ,0 pnT? p-1p,1

19、,0 pnpknppnT +? ,0 pnT3-8 ,0 pnnnnnppnT+?p ,0 pnnnnnpnT+ ? ,11,0nnpnpnT?+Tn 11,/,11,0 nTpnpnT+? ,11,0pnpnnTpnT? ,1,1, 0 pnpnpnnT T?3-9 ,1,pnpn?3-9 ,1, 0 pnCnT T34.1.11Data Generation2Data Display 3? Data Analysis and User Interation? 4-1 264.1.2 Statistics File?Trace File? 4-1 27 4.1.3 4.1.3

20、.1 Hardware View1Curve2Cache Miss3 Volume? Frequency284Topology? Software View 1Scheduling Transport Layer? TCP/IP 2 29 Communication Cost Synchronous Cost4.1.4 12? Multiple View3430? Paradyn225Visual Explosion? Color SpectrumXgecko? Color Scale? Load Percentage 23ZoomingFiltering314.21 Parad

21、yn24 ParadynParadyn W3?on-the-flyW3Why?WhereWhen2 ParaGraph25 ParaGraph? ParaGraph Portable Tnstrumented Communication Library Paragrah 3 PAPI26 PAPIsubstratePAPlPAPI?Perfometerperfometer 324 VAMPIR27Visualization and Analysis of MPI ResourceMPIMPIVAMPIRTRACE? MPIMPI? Vampirtrace MPI Vampir? VAMPIR5

22、 MPE28MPE? mpich? log? MPEMPE? MPI? CLOG? SLOG CLOG SLOG? logfile? mpiP29 TAU30SvPablo31DEEP/MPI32Pablo Performance Analysis Tools32XPVM33XMPI34? IBM? HPM Hardware Performance Monitor35Intel? Vtune36 VIMP37?ACI? THPTiiVENUS384.31 2? 33 Cache?3 4 5PVM 6 2000?39? ?34?40?metatools41?42? AVS apE

23、 Data Explore?434.4.2?/? Delbert Hart? Eileen Kraemer?44 354.4.3 4.4.4? 1212 3364 564.4.51 2 3? Delbert Hart?44? J.Muthukumarasamy454.5? NOW37395.1? I/OI/O MPI MPI?CPU 5-140? 5-1 5-2 MPI C C+ Fortran? Windows5-25.2? MPI5.2.1PCMPIMPICH2 LinuxWindows MPICH2 MPI41? WINNTMPICH2 MPICH2MPICH21? WINNTWINDO

24、WS 4 wmpiconfig.exe,Scan Hosts, MPICH25 6 wmpiexec.exe,?Execute? PC5.2.2MPInet? C#? MPI MPI MPI.H C FortranC+MPIMPI? C#? GDI+ C#? C#? C#? GDI+? GDI? GDI2042C#? MicrosoftNET? C+ 20C+C+ C+? Visual Basic? C# C#?.NET CLRC? C+ MicrosoftNET C#C#? C+C+ C#? C#C#C+?C#C+ C#? C+C#C+.NET? C#C#?.NET?.NET

25、 FrameworkNETVB.NETCLRNETC#.NET FrameworkNETC#.NET FrameworkC#NET Framework 43GDI+? Windows XPC+GDI+ GDI Windows XP GDI GDI+ GDI+? Windows GDI? GDI+?,GDI+?,GDI+ GDI+1Gradient BrushesGDI+?,?shapes,paths,?regions,?shapes 2? Cardinal SplinesGDIGDI 3? Persistent Path Objects? GDI DC? GDI+Graphics? Graph

26、ics4Transformations Matrix ObjectGDI+? PATH? Transform445 Scalable RegionsGDI+? regions? GDI,? GDIRegions Regions GDI+Regions,6 Alpha BlendingGDI+ Alpha Blending?,alpha 7 GDIBMP? GDI? JPEG Joint Photographic Experts Group GIFGraphics Interchange Format PNG Exchangeable Image File TIFFTag Image File

27、FormatGDI?GDI? Windows GDI? GDI+? Graphics myGraphicsObject.DrawLineparameters Graphics GDI+ GDIDC GraphicsGDI Pen?Path Image? FontGDI? SelectObject45GDI+? GraphicsPen? Brush 0 CalcUnitData,? 5-3 +CalcUnitData+SetValuein t : double , in c : double, in p : double+SetValuein svaluearray : strin

28、g+SetValuein cua : CalcUnitData+Drawin g : Graphics, in x : int, in y : int+TotalTime : double 0+ComAndSynTime : double 0+ParaAbleTime : double 0+ColorTotalTime : Color+ColorComAndSynTime : Color+ColorParaAbleTime : ColorCalcUnitData? 5-3CalcUnitData?OnPaintPaintEventArgs e5-446? 5-45-55-547 6.1.16.

29、1.2FEA Finite Element Analysis?1 234485 6 IYVJzVzVZJYI?+ 6-1 VZVZYVY ?+? 11 6-2HEVHx ExJ Ey HyIEy Hz ?i? i? ? E11? H11 Z xy JV Z yx VJ 49? 6-1 6-2 6-1? minYY YY ? zz ? E11 H11 E11 H11Hz0 H xz0H x 1E x6-21 ZZ 111 ?+ VZV2 ZZ min221 ?+ VZV3 Z Y331 ?+ VYV6-3 minZZ X YZYYmaxYYminZZ Z, ZY504 Z minY 441 ?+

30、 VYV6-3? H EHSZ1Z 1?HS01 0L2 0 6-4 L2104433 ? L V min22212102222?+? + ?dSVVdZYdVzVYVVI 6-5? 0VVZVZYVY ?+? 110 + VYV1 6-6 6-6? 0 1 2 3 41 2 346-1 6-3 min0+ ? IIVI 6-7? + dZYdVzVYVI 22222121 ? 6-2 6-251 yzCzCyCCzyV 3210, +iC? V? iV Tiiiii MVMVVVV 1,3,2,1 + iV V TTMTT VVV 121 , + V 11 +×+ NM ji,?

31、ji ji ihj zyV , ijiijiiijjiijijiijhyzjVhyzhzjVhyzyjVhyzhzyjVzzyyV111,11 +?+?+?+6-8? 6-5 ji, eeTe VKVjiI 21, 1234214334124321kkkkkkkkkkkkkkkkKeek V1 V2 V3 V MV 1+M1 2N1h2hMh6-252jhjhhjjikiijijijii339,1 + jhjhhjjikiijijijii6318,2 +? jhjhhjjikiijijijii3618,3 ?+ jhjhhjjikiijijijii6636,4 ekVKVI T ? 21 6-

32、9?K ? 11 +×+ NM 2 2 j,1 eTeeeTe fVVTVjI 112 21 ? Nj ,2,1+ +?0000000000141121112112100112112111214122222222j1 jjjjjjjjT jjje Te jVjVjVjVV 1,1, 2211 + +? 0,0,316,61322221jjjjf jje ? jI 2 53111112 21 fVVTVI TT ? 11111 21+ ? MTMMMTM fVVTVIMT eMT 1+MV eV 1+Mf eMf 1+ +14112111211210011211211121410000

33、00000011111111jjjjjjjjTjjjjeM Tmmmme jVjVjVjVV 1,1, 111 + + +?+ 316,613,0,0 11111jjjjf jjem 6-3 6-4043 II fVKVVIIIIIVI TT ?+ ? 21210 6-10? VK ?K T1 T Mf? 1f1+Mf6-10 fKV K54+?MMMMMMMTABBAABBAABBAABBTA1133222211111+1321MMVVVVV? +1100Mff 6-11?+NiKNiKNiKNiKNiKNiKiKiKiKiKiKiKAi,1,1,2,2,1,1,1,1,1221122112

34、21?+NiKNiKNiKNiKNiKNiKiKiKiKiKiKiKiKBi,1,1,2,2,2,1,1,1,1,3443344433443ii ,1TMT +? +NtNtNtNtNtNtttttttttttTiiiiiiiiiiiiiiiij21221122112211221111332222111111j2i? Mj 1i eiet i 321 55etet ii 12 211f1+Mf+?+? +NNNf iij13221121 11221ijijVzVJ? 1? V? JHExyHEVJZ ? yxHEZ 21 z? ? 10 ?6.3 K V f?-? K? K V f56 1 1

35、212 2 2323 3 341, , 1, 1 1M M M M MM M MA BB A BB A BKB A BB A+ +?6.12 K E F? I1 12 1 1212 2 23 2 231223 3 34 23, 1, 1 1 , 1 100M M MM M M M M MA B F GIB A B F GE IB A B E IF GB A E I F+ + + + ,E F F V Z E F V f E Z f1 112 2231 1, 1000M MM MZI fE I ZE IZ fE I + + 1Z E f? Z F V Z1 12 1 12 23 2 2, 11

36、110M M MM MMF G V ZF G V ZF GV ZF+ + 1V F Z?V1,E F 57;12,1,111,1,1,1,111iiiiiiiiiiiiiiiGEAFBEBGdoMtoiforAFF?+2Z ;12;1,111?+iiiii ZEfZdoMtoiforfZ3 ?V ;1;11,11111+?iiiiiiMMMVGZVdotoMiforZVFF 597.1PC1? PC 7-17-1 CPURAM1 P4 2.66G 512M 10M/100M 2 P4 2.2G 512M 10M/100M2Windows XP+MPICH2 MTDCP+ MTDCP3?207-

37、2607-2? s?121?28.17751 6.9736 4.18708 4.174980.0473328 0.668067 15*10? 8.13384 6.8814 4.10014 3.47464 1.812 0.906105.6 87.3652 51.1731 50.81410.0834978 7.67814 15*20 ? 105.556 87.3323 51.05 43.1035 1.892 0.94616.4852 13.2077 8.2364 8.305210.0712191 1.7632 30*10 ? 16.4428 13.1733 8.1204 6.51111 1.795

38、 0.898208.558 178.163 101.351 100.4030.1178 16.1041 30*20 ? 208.513 178.065 101.192 84.2688 1.917 0.958461.143 332.729 222.953 204.1430.267564 32.9006 60*20 ? 461.098 332.694 222.644 171.207 1.859 0.929?7-161?7-17-112 ?7-71 1 1 2 1.917 0.958 3 2.864 0.955 4 3.445 0.8617-27-2?7-37-3 64 4 7-8 7-8 4s4

39、10.127 8 20.1275 16 39.9545 20 49.6823 40 100.698?7-47-4? 4? 7-27-3? 1 7-44? 23 4? 365? 4 67 12MPIMPI 3 4CPU 5 6CUP 68 1 2 3 69 104711, 2004 2 /0>. 2004 3? MPI2001 4 Michael J.Quinn MPI? OpenMP C 2004 5 Barry Wilkinson Michael Allen 2 20056 Jack Dongarra Ian Foster Geoffrey Fox William Gropp? 2005 7 MICHAEL J.QUINN C MPI? OpenMP 2005 8 C.Xavier /.ngar? 20049 2003 10 Cameron Hughes Tracey HughesC+? 2004 11 Timothy G.Matts