12-14-2009, 03:32 AM
Hi all!
The main problem swith shifting to 32 bit is, that you can't use more than 2 GB per process - which can cripple your simulation when working with large models.
on 64 bit systems with a lot of RAM you might also want to use huge pages if supported by your application. This speeds up the memory subsystem of your machine considerably.
Huge pages can have several megabytes, while a normal page on x86 has 4 Kbyte. you can imagine the overhead of managing 32GB of ram with 4k pages.
I am currently trying to compile Aster 10.0 under opensuse 11.1 with the Intel compilers and AMD acml libraries, but have quite some problems getting it work.
The machine is a dual opteron 252 with 8 Gb of RAM.
I was able to compile with gnu-compilers, but then almost halve of the testcases don't run through, which is not good enough. According to the code-aster forum that points to compilation errors in the fortran parts.
The Problem with the intel compilers lies mainly in the point that i was not able to convince the aster install script to use the AMD libraries with the intel compilers - which should be possible. As those libraries speed some floating point comptations up by a factor of up to ten and most of the FP code to at least the factor two it would be foolich not to use that when compiling.
The main problem swith shifting to 32 bit is, that you can't use more than 2 GB per process - which can cripple your simulation when working with large models.
on 64 bit systems with a lot of RAM you might also want to use huge pages if supported by your application. This speeds up the memory subsystem of your machine considerably.
Huge pages can have several megabytes, while a normal page on x86 has 4 Kbyte. you can imagine the overhead of managing 32GB of ram with 4k pages.
I am currently trying to compile Aster 10.0 under opensuse 11.1 with the Intel compilers and AMD acml libraries, but have quite some problems getting it work.
The machine is a dual opteron 252 with 8 Gb of RAM.
I was able to compile with gnu-compilers, but then almost halve of the testcases don't run through, which is not good enough. According to the code-aster forum that points to compilation errors in the fortran parts.
The Problem with the intel compilers lies mainly in the point that i was not able to convince the aster install script to use the AMD libraries with the intel compilers - which should be possible. As those libraries speed some floating point comptations up by a factor of up to ten and most of the FP code to at least the factor two it would be foolich not to use that when compiling.
