Loading...
Please wait, while we are loading the content...
The Effects of Latency and Occupancy in Distributed Shared Memory Multiprocessors
| Content Provider | Semantic Scholar |
|---|---|
| Author | Holt, Chris Heinrich, Mark Singh, Jaswinder Pal Rothberg, Edward Hennessy, John Brian |
| Copyright Year | 2011 |
| Abstract | Many designers of distributed shared memory (DSM) multiprocessors are proposing the use of commodity parts, not only in the processor and memory subsystem but also in the communication architecture. While the desire to use commodity parts in the communication architecture offers potential advantages in cost and design time, the impact on the performance of applications is unclear. In this paper we study this performance impact through detailed simulation and analytical modeling, using a range of important applications and computational kernels. We characterize the communication architectures of DSM machines by four parameters, similar to those in the logP model. The l (latency) and o (controller occupancy, or controller bandwidth) parameters are the keys to performance in these machines, with the g (gap or node-to-network bandwidth) parameter not being a bottleneck in recent and upcoming machines. Conventional wisdom is that latency is the dominant performance bottleneck in DSM machines. We show, however, that controller occupancy also has a substantial impact on application performance, for the occupancies that are being proposed for a range of cache-coherent DSM machines. Moreover, performance is affected more by the contention that controller occupancy induces than the latency it adds. As expected, techniques to reduce the impact of network latency make controller occupancy a greater bottleneck. What is surprising, however, is that 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. We show that in many structured computations occupancyinduced contention is not alleviated by increasing problem size, and that there are important classes of applications for which the performance lost by using higher latency networks or higher occupancy controllers cannot be regained easily, if at all, by scaling the problem size. |
| File Format | PDF HTM / HTML |
| Alternate Webpage(s) | http://csl.cs.ucf.edu/~heinrich/papers/Effects.pdf |
| Alternate Webpage(s) | https://www.researchgate.net/profile/Mark_Heinrich/publication/2261805_The_Effects_of_Latency_and_Occupancy_in_Distributed_Shared_Memory_Multiprocessors/links/02bfe510fc8f4f3e94000000.pdf |
| Language | English |
| Access Restriction | Open |
| Content Type | Text |
| Resource Type | Article |
