Since the kernel update on the Cluster we see severe performance problems when running a parallel sander on more than 4 CPUs. The effect cannot not yet be explained. Further investigations are on the way… see comments for the solution

Reasonable performance after update of switch firmware: After updating/upgrading the firmware of the HP ProCurve switches 2848 the performance of parallel AMBER runs is again reasonable. Tests with 16 CPUs (8 nodes) and a few with 32 CPUs (16 nodes) show reasobale runtimes again. Strange that the performance with the old switch firmware so strongly depended on the version of the Linux kernel on the compute nodes …