root / HISTORY
Historique | Voir | Annoter | Télécharger (2,52 ko)
| 1 | 1 | equemene | ============================================================== |
|---|---|---|---|
| 2 | 1 | equemene | High Performance Computing Linpack Benchmark (HPL) |
| 3 | 1 | equemene | HPL 2.0 - September 10, 2008 |
| 4 | 1 | equemene | ============================================================== |
| 5 | 1 | equemene | |
| 6 | 1 | equemene | History |
| 7 | 1 | equemene | |
| 8 | 1 | equemene | - 09/09/00 Public release of Version 1.0 |
| 9 | 1 | equemene | |
| 10 | 1 | equemene | - 09/27/00 A couple of mistakes in the VSIPL port have been |
| 11 | 1 | equemene | corrected. The tar file as well as the web site were updated |
| 12 | 1 | equemene | on September 27th, 2000. Note that these problems were not |
| 13 | 1 | equemene | affecting the BLAS version of the software in any way. |
| 14 | 1 | equemene | |
| 15 | 1 | equemene | - 01/01/04 Version 1.0a |
| 16 | 1 | equemene | The MPI process grid numbering scheme is now an run-time |
| 17 | 1 | equemene | option. |
| 18 | 1 | equemene | The inlined assembly timer routine that caused the compila- |
| 19 | 1 | equemene | tion to fail when using gcc version 3.3 and above has been |
| 20 | 1 | equemene | removed from the package. |
| 21 | 1 | equemene | Various building problems on the T3E have been fixed; Thanks |
| 22 | 1 | equemene | to Edward Anderson. |
| 23 | 1 | equemene | |
| 24 | 1 | equemene | - 15/12/04 Version 1.0b |
| 25 | 1 | equemene | Weakness of the pseudo-random matrix generator found for pro- |
| 26 | 1 | equemene | blem sizes being power of twos and larger than 2^15; Thanks |
| 27 | 1 | equemene | to Gregory Bauer. This problem has not been fixed. It is thus |
| 28 | 1 | equemene | currently recommended to HPL users willing to test matrices |
| 29 | 1 | equemene | of size larger than 2^15 to not use power twos. |
| 30 | 1 | equemene | |
| 31 | 1 | equemene | When the matrix size is such that one needs > 16 GB per MPI |
| 32 | 1 | equemene | rank, the intermediate calculation (mat.ld+1) * mat.nq in |
| 33 | 1 | equemene | HPL_pdtest.c ends up overflowing because it is done using |
| 34 | 1 | equemene | 32-bit arithmetic. This issue has been fixed by typecasting |
| 35 | 1 | equemene | to size_t; Thanks to John Baron. |
| 36 | 1 | equemene | |
| 37 | 1 | equemene | - 09/10/08 Version 2.0 |
| 38 | 1 | equemene | |
| 39 | 1 | equemene | Piotr Luszczek changed to 64-bit RNG, modified files: |
| 40 | 1 | equemene | -- [M] include/hpl_matgen.h |
| 41 | 1 | equemene | -- [M] testing/matgen/HPL_ladd.c |
| 42 | 1 | equemene | -- [M] testing/matgen/HPL_lmul.c |
| 43 | 1 | equemene | -- [M] testing/matgen/HPL_rand.c |
| 44 | 1 | equemene | -- [M] testing/ptest/HPL_pdinfo.c |
| 45 | 1 | equemene | |
| 46 | 1 | equemene | For a motivation for the change, see: |
| 47 | 1 | equemene | Dongarra and Langou, ``The Problem with the Linpack |
| 48 | 1 | equemene | Benchmark Matrix Generator'', LAWN 206, June 2008. |
| 49 | 1 | equemene | |
| 50 | 1 | equemene | -- [M] testing/ptest/HPL_pdtest.c -- |
| 51 | 1 | equemene | |
| 52 | 1 | equemene | Julien Langou changed the test for correctness from |
| 53 | 1 | equemene | ||Ax-b||_oo / ( eps * ||A||_1 * N ) |
| 54 | 1 | equemene | ||Ax-b||_oo / ( eps * ||A||_1 * ||x||_1 ) |
| 55 | 1 | equemene | ||Ax-b||_oo / ( eps * ||A||_oo * ||x||_oo * N ) |
| 56 | 1 | equemene | to the normwise backward error |
| 57 | 1 | equemene | || r ||_oo / ( eps * ( || x ||_oo * || A ||_oo + || b ||_oo ) * N ) |
| 58 | 1 | equemene | See: |
| 59 | 1 | equemene | Nicholas J. Higham, ``Accuracy and Stability of Numerical Algorithms'', |
| 60 | 1 | equemene | Society for Industrial and Applied Mathematics, Philadelphia, PA, USA, |
| 61 | 1 | equemene | Second Edition, pages = xxx+680, ISBN = 0-89871-521-0, 2002. |
| 62 | 1 | equemene | |
| 63 | 1 | equemene | Note that in our case || b ||_oo is almost for sure |
| 64 | 1 | equemene | 1/2, we compute it anyway. |
| 65 | 1 | equemene | |
| 66 | 1 | equemene | ============================================================== |