Using Skip Graphs for Increased NUMA Locality

摘要

High-performance simulations and parallel frameworks often rely on highly scalable, concurrent data structures for system scalability. With an increased availability of NUMA architectures, we present a technique to promote NUMA-aware data parallelism inside a concurrent data structure, bringing significant quantitative and qualitative improvements on NUMA locality, as well as reduced contention for synchronized memory accesses. Our architecture is based on a data-partitioned, concurrent skip graph indexed by thread-local sequential maps. We implemented maps and relaxed priority queues using such technique. Maps show up to 6x higher CAS locality, up to a 68.6% reduction on the number of remote CAS operations, and an increase from 88.3% to 99% on the CAS success rate compared to a control implementation (subject to the same optimizations, and implementation practices). Remote memory accesses are not only reduced in number, but the larger the NUMA distance between threads, the larger the reduction is. Relaxed priority queues implemented using our technique show similar scalability improvements, with provable reduction in contention and decrease in relaxation in one of our implementations.