Parallel Management of Large Dynamic Shared Memory Space: a Hierarchical FEM Application
Résumé
We show in this paper the memory management issues raised by a parallel irregular and dynamic hierarchical application, which constantly allocates and deallocates data over an extremely large virtual address space. First, we show that if memory caches data locality is necessary, a lack of virtual pages locality may greatly affect the obtained performance. Second, fragmentation and contention problems associated with the required parallel dynamic memory allocation are presented. We propose practical solutions and discuss experimentation results obtained on a cache-coherent non uniform memory access (ccNUMA) distributed shared memory SGI Origin2000 machine.