I upgrade my beowulf clusters to mandrake 5.3 from 5.2, then upgrade mpich-1.1.1 to mpich-1.1.2. I found that the performance of NPB LU class A benchmark drop from 400Mgflops to 300Mgflops !.First I though it is mpich-1.1.2 which cause the problem,but when i switch back to mpich-1.1.1 the result is same (300Mgflops) . So any ideas why the performance drop ? Normaly,we should get BETTER when upgrading .. And, did anyone try out the performance with kernel 2.2.2 ? How the benchmark result ?
