Balancing Shared and Distributed Heaps on NUMA Architectures

Malak Aljabri, Hans-Wolfgang Loidl, and Phil Trinder

Abstract:

Due to the varying latencies between memory banks, efficient shared memory access is challenging on modern NUMA architectures. This has a major impact on the shared memory performance of parallel programs, particularly those written inlanguages with automatic memory management.

This paper presents a performance evaluation of distributed and shared heap implementations of parallel Haskell on a state-of-the-art physical shared memory NUMA machine. The evaluation exposes bottlenecks in the shared-memory management, which results in limits to scalability beyond 25 out of the 48 cores. We demonstrate that a hybrid system, GUMSMP, that combines both distributed and shared heap abstractions consistently outperforms the shared memory GHC implementation on seven benchmarks by a factor of 3.3 on average. Specifically, we show that the best results are obtained when sharing memory only within a single NUMA region, and using distributed memory system abstractions across the regions.