- January 18, 2013

Benchmarks

* Oracle R Distribution provides*

dramatic performance gains with MKL

Using the recognized R benchmark R-benchmark-25.R test script,

we compared the performance of Oracle

R Distribution with and without the dynamically loaded high performance Math Kernel Library (MKL) from

Intel. The benchmark

results show Oracle R Distribution is significantly faster with the dynamically

loaded high performance library. R users can immediately gain performance enhancements

over open source R, analyzing data on 64-bit architectures and leveraging

parallel processing within specific R functions that invoke computations

performed by these high performance libraries.

The Community-developed

test consists of matrix calculations and functions, program control, matrix multiplication,

Cholesky Factorization, Singular Value Decomposition (SVD), Principal Component

Analysis (PCA), and Linear Discriminant Analysis. Such computations form a core

component of many real-world problems, often taking the majority of compute

time. The ability to speed up these computations means faster results for

faster decision making.

While the benchmark results reported were conducted

using Intel MKL, Oracle R Distribution

also supports AMD Core Math Library (ACML) and Solaris Sun Performance Library.

**Oracle R Distribution 2.15.1 x64 Benchmark Results (time in seconds)**

| ORD with internal BLAS/LAPACK1 thread | ORD + MKL 1 thread | ORD + MKL2 threads | ORD + MKL4 threads | ORD + MKL8 threads | Performance gain ORD + MKL 4 threads | Performance gain ORD + MKL 8 threads |

Matrix Calculations | 11.2 | 1.9 | 1.3 | 1.1 | 0.9 | 9.2x | 11.4x |

Matrix Functions | 7.2 | 1.1 | 0.6 | 0.4 | 0.4 | 17.0x | 17.0x |

Program Control | 1.4 | 1.3 | 1.5 | 1.4 | 0.8 | 0.0x | 0.8x |

Matrix Multiply | 517.6 | 21.2 | 10.9 | 5.8 | 3.1 | 88.2x | 166.0x |

Cholesky Factorization | 25 | 3.9 | 2.1 | 1.3 | 0.8 | 18.2x | 29.4x |

Singular Value Decomposition | 103.5 | 15.1 | 7.8 | 4.9 | 3.4 | 20.1x | 40.9x |

Principal Component Analysis | 490.1 | 42.7 | 24.9 | 15.9 | 11.7 | 29.8x | 40.9x |

Linear Discriminant Analysis | 419.8 | 120.9 | 110.8 | 94.1 | 88.0 | 3.5x | 3.8x |

This benchmark was executed on a 3-node cluster, with 24 cores at 3.07GHz

per CPU and 47 GB RAM, using Linux 5.5.

In the first graph, we see significant performance improvements. For example, SVD with ORD plus MKL executes 20 times faster using 4 threads, and 29 times faster using 8 threads. For Cholesky Factorization, ORD plus MKL is 18 and 30 times faster for 4 and 8 threads, respectively.

In the second graph,we focus on the three longer running tests. Matrix multiplication is 88 and 166 times faster for 4 and 8 threads, respectively. PCA is 30 and 50 times faster, and LDA is over 3 times faster.

This level of performance improvement can significantly reduce application execution time and make interactive, dynamically generated results readily achievable. Note that ORD plus MKL not only impacts performance on the client side, but also when used in combination with R scripts executed using Oracle R Enterprise *Embedded R Execution*. Such R scripts, executing at the database server machine, reap these performance gains as well.

Comments ( 0 )