Experience using Adaptive Concurrency Control for Transactional Memory with Lee's Routing Algorithm (poster presentation)

Mohammad Ansari, Christos Kotselidis, Kimberly Jarvis, Mikel Lujan, Chris Kirkham and Ian Watson

The 13th ACM SIGPLAN Symposium on Principles and Practice of Parallel Programming (PPoPP 2008)
Salt Lake City, Utah, February 20-23, 2008

---

Abstract

Transactional applications may exhibit fluctuating amounts of contention during execution. Excessive numbers of threads executing transactions can produce phases with high contention. This high contention materializes as increased number of conflicts and aborted transactions which hurt performance and resource utilization. Similarly, a limited number of threads executing transactions will under-perform in phases with low contention. Current Transactional Memory (TM) implementations have overlooked this behavior.

This paper presents the first application of adaptive concurrency control to TM in order to dynamically adjust the number of threads executing transactions concurrently. Four adaptive schemes are implemented in DSTM2, a software TM implementation, and evaluated against a TM benchmark with complex and realistic behavior. Adaptive concurrency control complements existing contention management policies that capture which transaction should be aborted when two transactions conflict. Compared with existing contention management policies without adaptive concurrency control running with 8 threads, the best case using adaptive concurrency control shows a performance improvement of 247% and resource usage improvement of 82%, and an average performance improvement of 38% and resource usage improvement of 53%.

---

  

START Conference Manager (V2.54.5)