Problems
The following problems occured on the systems we used during the benchmarks:
Sun SunFire 6800
- SunPerf-Lib: had occasional segfaults and non-terminating runs when using SunPerf-lib. Had to preinitialise dlartg_ to overcome this (bug ?).
- Had to use SunPerf 7 (BETA) for a reasonable speedup when using more than one CPU. SunPerf 6 seems to have some issues here.
Compaq Alphaserver GS160
- original benchmark ran without ptmalloc, therefore suffered from severe memory fragmentation
- in the Compaq-Testdrive the new malloc runs only on the Linux-systems; under Tru64 it resulted in SIGBUS / SIGSEGV because incompatibilities with the pthread-lib
- because of the NUMA-architecture, the benchmarks are not really suited to test this machine
IBM Power3
- ESSL-library only supports subset of LAPACK; had to mix CLAPACK and ESSL (replaced long by int in f2c.h !!!)
- C++ compiler produced too many warnings and errors while compiling libbem; benchmark was skipped
HP 9000 Superdome
- overwriting malloc in combination with -lpthread leads to a lock in __ksleep; only system-malloc usable; resolved by overwriting operators new and delete in C++ and NOT malloc (but still problems in free: recieved memory chunks which were not allocated by own malloc)
- couldn“t run benchmarks with more than 8 CPUs; program crashed with SIGBUS; SOLVED: had to increase thread-stack-size.
- C++-compiler only allows -fast (-O2), using -O4 leads to error when linking programs