From: ahemavathy (ahemavathy_at_hotmail.com)
Date: Sun Jan 25 2004 - 14:10:35 PST
This paper describes five thread management techniques on memory shared Multiprocessors and studies the performance implications of each of them. The five techniques are 1) single lock 2) Multiple Locks 3) Local free lists 4) idle queue and 5) local ready queue. The data structures needed by each thread are a program counter, stack and a control block. The thread operations : creation, startup and finish manipulate each of the shared data structures : ready queue, stack free list and the control block free list. The measurement results after comparing the five techniques clearly show that local ready queue is most efficient. Only local ready queues can support higher throughput.
Next the paper explores three spin-lock management alternatives : 1) spin on Xchgb - whose drawback is it makes the bus busier 2) spin on Read - where the spinning is done in cache thereby reducing bus traffic and 3) Ethernet-style backoff where when a failure occurs, it waits for sometime before trying again. Even though each scheme has its advantages and disadvantages, in most cases backoff strategy seems to be a better solution.
Next by modeling bus and lock contention, they study the effects of a combination of factors on the performance.
I liked this paper. The results are presented in clear graphs and it was easy to follow. It was interesting to read about each of the thread management techniques and how each of them look better than the previous one at first but eventually have their own drawbacks.
This archive was generated by hypermail 2.1.6 : Sun Jan 25 2004 - 14:11:04 PST