1 / 25

The role of virtual machine technologies for scientific computing

The role of virtual machine technologies for scientific computing. Roldan Pozo Mathematical Software Group National Institute of Standards and Technology. Clusters and Computational Grids for Scientific Computing 9-28-2004.

zelda
Télécharger la présentation

The role of virtual machine technologies for scientific computing

An Image/Link below is provided (as is) to download presentation Download Policy: Content on the Website is provided to you AS IS for your information and personal use and may not be sold / licensed / shared on other websites without getting consent from its author. Content is provided to you AS IS for your information and personal use only. Download presentation by click this link. While downloading, if for some reason you are not able to download a presentation, the publisher may have deleted the file from their server. During download, if you can't get a presentation, the file might be deleted by the publisher.

E N D

Presentation Transcript

  1. The role of virtual machine technologies for scientific computing Roldan Pozo Mathematical Software Group National Institute of Standards and Technology Clusters and Computational Grids for Scientific Computing 9-28-2004

  2. Java and .NET for numerics • what works • what doesn’t • what can be improved • what can we realistically expect

  3. Why we originally looked at Java • portability of the Java Virtual Machine (JVM) • minimize memory leaks and pointer errors • security model • network-aware environment • Parallel and Distributed computing • Threads • Remote Method Invocation (RMI) • Integrated graphics • Widely adopted • embedded systems, browsers, devices • widely adopted for teaching, development

  4. Microsoft .NET • provides portable executables and runtime environment • based on Windows Win32 API • C# is the canonical language for .NET • compilers (C#, C++, Fortran, Basic) target same platform • interoperability from different languages • ECMA Standard Specification for C# and CLI • Shared Source CLI (Rotor) available for • Windows XP, FreeBSD, Mac OS X

  5. Java concerns • Performance • early JVMs too slow • poor optimizing compilers • virtual machine • security model • array bounds checking

  6. Java concerns • lack of scientific software • computational libraries • numerical interfaces • considerable effort to port from f77/C/C++ • no universal translator

  7. Java Numerics Working Group • industry-wide consortium to establish tools, APIs, and libraries • Sun, IBM, Intel, Compaq/Digital/HP MathWorks, VNI, NAG • component of Java Grande Forum • updated Java’s floating-point model • served as focal point for numeric activities • proposals and implementations for • complex, arrays, mult-adds, fdlibm • libraries, compilers, language extensions

  8. Numeric issues • array bounds checking • floating point model subset of IEEE 754 • no IEEE extended formats (80 bit FPU stack) • many compiler optimizations (e.g. associativity) disallowed • cannot use hardware acceleration • e.g no mapping sin(x) to x86 opcodes • no fused multiply-adds

  9. What we needed • floating point model • true multidimensional arrays • complex data types • lightweight objects • operator overloading • generic typing (templates)

  10. Java Linear Algebra Libraries • JAMA (NIST & MathWorks) • provides APIs and implementations for • basic vector, matrix operations • LU, QR, SVD and Cholesky factorizations • Eigenvalue decompositions • JLAPACK (UTK) • JamPack (NIST & UMD)

  11. SciMark Benchmark • Numerical benchmark for Java, C/C++, C# • composite results for five kernels: • FFT (complex, 1D) • Successive Over-relaxation • Monte Carlo integration • Sparse matrix multiply • dense LU factorization • results in Mflops • two sizes: small, large Pozo, Miller (NIST) Java & C, Werner Vogels (Cornell, Rotor HPC Project) C#

  12. SciMark database of results • Over 2,200 separate results for • Hardware platforms • laptops, desktops, and PDAs: Intel, Apple, Sun, IBM, Apple, Sharp, AMD • Java Virtual Machines (JVM) • Sun, IBM, Apple, Blackdown, Microsoft, Compaq, HP, Symantec, Netscape, BEA, NSIcom, Golden Code, FreeBSD • Operating Systems • Windows 95/98/2000/ME/NT/XP, OS/2, AIX, NetBSD, Linux, IRIX, FreeBSD, Ultrix, Solaris, SunOS, Apple OS X, Compaq, NETWARE, OSF1

  13. Some SciMark 2.0 results

  14. JVMs have improved over time(Scimark scores) SciMark : 333 MHz Sun Ultra 10

  15. SciMark: Java, C, C#Large out-of-cache problem sizes gcc -O6 -funroll-all-loops Borland bcc32 -6 -0 (Intel P4 2.5GHz, Windows XP .NET MS C/C++ v. 13.10)

  16. SciMark: Java, C, C#Small in-cache problem sizes gcc -O6 -funroll-all-loops Borland bcc32 -6 -0 Intel P4 2.5GHz, Windows XP .NET MS C/C++ v. 13.10 .NET C# v. 7.0)

  17. C# multiarray performanceScimark small in-cache problem sizes (Intel P4 2.8GHz, Windows XP, CLR v. 1.0.3705.288)

  18. Fortran performance in .NETTime (secs) for Polyhedron 2004 F90 Benchmarks FTN95: Salford FTN95 /omptimize /p6 FTN95.NET: Salford FTN95 /optimize /clr Intel: ifort /03 /Qipo /QxN /link /stack: 64000000 3.06 GHz Intel Pentium 4, 512K cache, 768 MB Windows XP Pro

  19. Java and Fortran PerformanceIBM High Performance Fortran compiler (Moriera, et. al.) *IBM RS/6000 67MHz POWER2 (266 Mflops peak) AIX Fortran, HPJC

  20. Java Numerical Software • General • Apflot, Colt, JADE, Java3D (matrix), ArciMath BigDecimal, IBM Alphaworks, Jsci, Spline++, JMSL (Visual Numerics), jCrunch, mpjava, RngPack, OpsResearch, • Linear Algebra • JAMA, Jampack, Java LAPACK, Matrix Toolkit, Owlpack, JSP • Language extensions (compilers) • multi-dimensional arrays (IBM), complex numbers (zeta) , HPJava, Cj, Titanium

  21. Scientific .NET Software • Center Space Software: NMath library • mathematical, financial, engineering, scientific • linear algebra, random numbers, integration, interpolation, statistics, linear regression • native interface to BLAS/LAPACK • callable from any .NET language • .NET Analytics: dnAnalytics library • complex, double matrices, LU, Cholesky, QR, SVD • Matrix Library in C# or with native interfaces • native interface to ATLAS, CLAPACK, Intel’s MKL, Apple’s Vector Library (veclib) • Compatible with Mono on Windows, Linux, MacOS

  22. Java Conclusions • Java’s strength: • binary portability; single software distribution • competitive performance • strict security model; numerical reproducibility • Java’s obstacles: • no standard support for multidimensional arrays, complex numbers, and operator overloading • limited numeric software and library support • no blind conversion of C/C++/Fortran codes • Can be solved (technologically) • but need standards and support

  23. .NET/C# Conclusions • .NET’s strengths: • common platforms for Microsoft languages • shared source CLI (Rotor) implementation, ECMA standard • interlanguage compatibility (Fortran, C/C++) • competitive performance • C# provides structs, operator overloading, multiarrays • more flexible security model; limited reproducibility • .NET’s obstacles: • not as widely available on all platforms • various projects (e.g. Mono, Rotor) are addressing this • compatibility with Fortran and ANSI C/C++ is not 100% • limited numeric software and library support • Can be solved but also need standards and support

  24. Resources • Java Numerics Working Group • math.nist.gov/JavaNumerics • Java Grande Benchmarks • www.javagrande.org • SciMark • (Java, C ) math.nist.gov/SciMark • (C#, .NET) rotor.cs.cornell.edu/SciMark/

More Related