When a large number of processors try to access a common variable, referred to as hot-spot accesses, the resulting memory contention not only can seriously degrade performance, but may also cause tree saturation in the interconnection network which blocks both hot-spot and regular requests alike. It has been shown that even if only a small percentage of all requests are to a hot-spot location, these requests can cause very serious performance problems, and it is proposed that special networks do the necessary combining of requests in order to keep the interconnection network and memory contention from becoming a bottleneck. The authors propose instead a software combining-tree concept and show that it is effective in decreasing memory contention and preventing tree saturation because it distributes hot-spot accesses over a software tree whose nodes can be dispersed among many memory modules. Thus it is an inexpensive alternative to expensive combining networks.
|Original language||English (US)|
|Title of host publication||Proceedings of the International Conference on Parallel Processing|
|Editors||Kai Hwang, Steven M. Jacobs, Earl E. Swartzlander|
|Number of pages||8|
|State||Published - Dec 1 1986|