Technical partnership with Intel: Westmere-EX benchmarks
Following our technical partnership with Intel software teams, Free Field Technologies implemented Intel MKL® multithreaded BLAS library (10.2.7) in order to efficiently parallelize ACTRAN direct solver on SMP machines without any extra memory consumption.
Applied to an industrial AIRBUS model with near 7 Mdofs, a single node with 4 processors Intel® Xeon® 7600 @ 2.3 Ghz achieved a 19 speedup on factorization phase, using more than 200 GB of memory, from 114 hours to 6 hours. This optimization was used as basic blocks on each node (2 processors Xeon® 5570) of a cluster, using MPI as hybrid parallelism level over nodes: same test case then ran in less than 50 minutes with 32 nodes (256 cores).
These results demonstrate obvious performance gain obtained with hybrid parallelism approach (MPI + MKL® Multithreaded BLAS). It also paves the way for new industrial model sizing that were not or hardly reachable previously. It opens door to new level of quality and engineering skills.
New advances in SMP architecture are always interesting: the linear speedup observed here at no RAM additional cost could allow us to address big models on a single node very efficiently while aiming at much larger computations when running on clusters.
A.Mosson - Numerical Acoustic Engineer – AIRBUS