Comparison of lock thrashing avoidance methods and its performance implications for lock design
Proceedings of the third international workshop on Large-scale system and application performance
| Hi-index | 0.00 |
Multi-core architectures have been adopted in various computing environments. Predictions based on Moore's Law state that thousands of cores can be integrated on a single chip within 10 years. To achieve better performance and scalability on multi-cores, applications should be multi-threaded, and therefore threads assigned on different cores can execute concurrently. However, lock contention in kernels can affect the scalability so significantly that the speedup decreases with the increasing number of cores (thrashing). Existing efforts to address this problem mainly focus on deferring lock thrashing, and therefore these techniques cannot prevent thrashing fundamentally. In this paper, we propose to use lock-aware scheduling to avoid thrashing. Our method detects thrashing on a per-thread basis and migrates contended threads to a smaller set of cores. The optimal number of cores is determined by maximizing the proposed normalized throughput model of migrated threads. The proposed method is implemented in Linux 2.6.29.4 and evaluated on a 32-core system. Experimental results on a series of lock-intensive micro- and macro-benchmarks show the effectiveness: for 3 of 5 workloads exhibiting thrashing behaviour, lock-aware scheduling can detect the speedup decrease accurately and sustain the maximal speedup, for the remaining 2 workloads, the performance can be improved greatly although the maximal speedup is not sustained, for 1 workload which does not suffer thrashing, the method introduces negligible runtime overhead.