root / INSTALL @ 9
Historique | Voir | Annoter | Télécharger (3,02 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 | 1) Retrieve the tar file, then |
7 | 1 | equemene | |
8 | 1 | equemene | gunzip hpl.tgz; tar -xvf hpl.tar |
9 | 1 | equemene | |
10 | 1 | equemene | this will create an hpl directory, that we call below the |
11 | 1 | equemene | top-level directory. |
12 | 1 | equemene | |
13 | 1 | equemene | 2) Create a file Make.<arch> in the top-level directory. For |
14 | 1 | equemene | this purpose, you may want to re-use one contained in the |
15 | 1 | equemene | setup directory. This file essentially contains the compilers |
16 | 1 | equemene | and librairies with their paths to be used. |
17 | 1 | equemene | |
18 | 1 | equemene | 3) Type "make arch=<arch>". This should create an executable |
19 | 1 | equemene | in the bin/<arch> directory called xhpl. |
20 | 1 | equemene | |
21 | 1 | equemene | For example, on our Linux PII cluster, I create a file called |
22 | 1 | equemene | Make.Linux_PII in the top-level directory. Then, I type |
23 | 1 | equemene | "make arch=Linux_PII" |
24 | 1 | equemene | This creates the executable file bin/Linux_PII/xhpl. |
25 | 1 | equemene | |
26 | 1 | equemene | 4) Quick check: run a few tests: |
27 | 1 | equemene | |
28 | 1 | equemene | cd bin/<arch> |
29 | 1 | equemene | mpirun -np 4 xhpl |
30 | 1 | equemene | |
31 | 1 | equemene | 5) Tuning: Most of the performance parameters can be tuned, |
32 | 1 | equemene | by modifying the input file bin/HPL.dat. See the file TUNING |
33 | 1 | equemene | in the top-level directory. |
34 | 1 | equemene | |
35 | 1 | equemene | ============================================================== |
36 | 1 | equemene | |
37 | 1 | equemene | Compile time options: At the end of the "model" Make.<arch>, |
38 | 1 | equemene | --------------------- the user is given the opportunity to |
39 | 1 | equemene | compile the software with some specific compile options. The |
40 | 1 | equemene | list of this options and their meaning are: |
41 | 1 | equemene | |
42 | 1 | equemene | -DHPL_COPY_L |
43 | 1 | equemene | force the copy of the panel L before bcast; |
44 | 1 | equemene | |
45 | 1 | equemene | -DHPL_CALL_CBLAS |
46 | 1 | equemene | call the cblas interface; |
47 | 1 | equemene | |
48 | 1 | equemene | -DHPL_CALL_VSIPL |
49 | 1 | equemene | call the vsip library; |
50 | 1 | equemene | |
51 | 1 | equemene | -DHPL_DETAILED_TIMING |
52 | 1 | equemene | enables detail timers; |
53 | 1 | equemene | |
54 | 1 | equemene | The user must choose between either the BLAS Fortran 77 |
55 | 1 | equemene | interface, or the BLAS C interface, or the VSIPL library |
56 | 1 | equemene | depending on which computational kernels are available on his |
57 | 1 | equemene | system. Only one of these options should be selected. If you |
58 | 1 | equemene | choose the BLAS Fortran 77 interface, it is necessary to fill |
59 | 1 | equemene | out the machine-specific C to Fortran 77 interface section of |
60 | 1 | equemene | the Make.<arch> file. To do this, please refer to the |
61 | 1 | equemene | Make.<arch> examples contained in the setup directory. |
62 | 1 | equemene | |
63 | 1 | equemene | By default HPL will: |
64 | 1 | equemene | *) not copy L before broadcast, |
65 | 1 | equemene | *) call the BLAS Fortran 77 interface, |
66 | 1 | equemene | *) not display detailed timing information. |
67 | 1 | equemene | |
68 | 1 | equemene | As an example, suppose one wants HPL to copy the panel of |
69 | 1 | equemene | columns into a contiguous buffer before broadcasting. In |
70 | 1 | equemene | theory, it would be more efficient to let HPL create the |
71 | 1 | equemene | appropriate MPI user-defined data type since this may avoid |
72 | 1 | equemene | the data copy. So, it is a strange idea, but one insists. To |
73 | 1 | equemene | achieve this one would add -DHPL_COPY_L to the definition of |
74 | 1 | equemene | HPL_OPTS at the end of the file Make.<arch>. Issue then a |
75 | 1 | equemene | "make clean arch=<arch>; make build arch=<arch>" and the xhpl |
76 | 1 | equemene | executable will be re-build with that feature in. |
77 | 1 | equemene | ============================================================== |
78 | 1 | equemene | |
79 | 1 | equemene | Check out the website www.netlib.org/benchmark/hpl for the |
80 | 1 | equemene | latest information. |
81 | 1 | equemene | ============================================================== |