On the scalability of image and signal processing parallel applications on emerging cc-NUMA many-cores

Nowadays, single-chip cache-coherent multi-cores up to 100 cores are a reality and many-cores of hundreds of cores are planned in the near future. This technological shift undertaking by the high-end computer-industry is converging with the design motivation of other domains like embedded and HPC industries. In this paper, we propose to investigate the scalability of the same four unmodified, shared-memory, image and signal processing oriented parallel applications on two targets: (i) embedded - TSAR, a single-chip 256-cores based, Cycle-Accurate-Bit-Accurate simulated, cc-NUMA many-core; and (ii) high-end - an AMD Opteron Interlagos, 64-core based, cc-NUMA many-core. Beside our scalability results on both cc-NUMA targets, our contributions include two operating system mechanisms: (i) a distributed, client/server based, scheduler design allowing the kernel to offer scalable inter-threads synchronization mechanisms; and (ii) a kernel-level memory affinity technique named Auto-Next-Touch allowing the kernel to transparently and automatically migrate physical pages in order to enforce the locality of thread's memory accesses. Although these two mechanisms are implemented and evaluated in ALMOS (Advanced Locality Management Operating System) running on the TSAR target, they remain applicable to other shared-memory operating systems.