SC17 Denver, CO

P44: Increasing Throughput of Multiprogram HPC Workloads: Evaluating a SMT Co-Scheduling Approach

Authors: Elias Lundmark (University West Sweden), Chris Persson (University West Sweden), Andreas de Blanche (University West Sweden, Tetra Pak), Thomas Lundqvist (University West Sweden)

Abstract: Simultaneous Multithreading (SMT) is a technique that allows for more efficient processor utilization by scheduling multiple threads on a single physical core, thus increasing the instruction level parallelism as it mixes instructions from several threads. Previous research has shown an average throughput increase of around 20% with an SMT level of two, e.g. two threads per core. However, a bad combination of threads can actually result in decreased performance. To be conservative, many HPC-systems have SMT disabled, thus, limiting the number of scheduling slots in the system to one per core. However, for SMT to not hurt performance, we need to determine which threads should share a core. In this poster, we use 30 random SPEC CPU job mixed on a twelve-core Broadwell based node, to study the impact of enabling SMT using two different co-scheduling strategies. The results show that SMT can increase performance especially when using no-same-program co-scheduling.
Award: Best Poster Finalist (BP): no

Poster: pdf
Two-page extended abstract: pdf

Poster Index