Synchronization in parallel programs is a major performance bottleneck in multiprocessor systems. shared data is protected by locks and a lot of time is spent on the competition arising at the lock hand-off. In order ...
详细信息
Synchronization in parallel programs is a major performance bottleneck in multiprocessor systems. shared data is protected by locks and a lot of time is spent on the competition arising at the lock hand-off. In order to be serialized, requests to the same cache line can either be bounced （NACKed） or buffered in the coherence controller. In this paper, we focus mainly on systems whose coherence controllers buffer requests. In a lock hand-off, a burst of requests to the same line arrive at the coherence controller. During lock hand-off only the requests from the winning processor contribute to progress of the computation, since the winning processor is the only one that will advance the work. This key observation leads us to propose a hardware mechanism we call request bypassing, which allows requests from the winning processor to bypass the requests buffered in the coherence controller keeping the lock line. We present an inexpensive implementation of request bypassing that reduces the time spent on all the execution phases of a critical section （acquiring the lock, accessing shared data, and releasing the lock） and which, as a consequence, speeds up the whole parallel computation. This mechanism requires neither compiler or programmer support nor ISA or coherence protocol changes. By simulating a 32-processor system, we show that using request bypassing does not degrade but rather improves performance in three applications with low synchronization rates, while in those having a large amount of synchronization activity （the remaining four）, we see reductions in execution time and in lock stall time ranging from 14% to 39% and from 52% to 7170, respectively. We compare request bypassing with a previously proposed technique called read combining and with a system that bounces requests, observing a significantly lower execution time with the bypassing scheme. Finally, we analyze the sensitivity of our results to some key hardware and software parameters.
While the desire to use commodity parts in the communication architecture of a DSM multiprocessor offers advantages in cost and design time, the impact on application performance is unclear. We study this performance ...
详细信息
While the desire to use commodity parts in the communication architecture of a DSM multiprocessor offers advantages in cost and design time, the impact on application performance is unclear. We study this performance impact through detailed simulation, analytical modeling, and experiments on a flexible DSM prototype, using a range of parallel applications. We adapt the logP model to characterize the communication architectures of DSM machines. The l (network latency) and o (controller occupancy) parameters are the keys to performance in these machines, with the g (node-to-network bandwidth) parameter becoming important only for the fastest controllers. We show that, of all the logP parameters, controller occupancy has the greatest impact on application performance. Of the two contributions of occupancy to performance degradation-the latency it adds and the contention it induces-it is the contention component that governs performance regardless of network latency, showing a quadratic dependence on o. As expected, techniques to reduce the impact of latency make controller occupancy a greater bottleneck. Surprisingly, the performance impact of occupancy is substantial, even for highly-tuned applications and even in the absence of latency hiding techniques. Scaling the problem size is often used as a technique to overcome limitations in communication latency and bandwidth. Through experiments on a DSM prototype, we show that there are important classes of applications for which the performance lost by using higher occupancy controllers cannot be regained easily, if at all, by scaling the problem size.
暂无评论