Modeling node bandwidth limits and their effects on vector combining algorithms
Each node in a message-passing multicomputer typically has several communication links. However, the maximum aggregate communication speed of a node is often less than the sum of its individual link speeds. Such computers are called node bandwidth limited (NBL). The NBL constraint is important when choosing algorithms because it can change the relative performance of different algorithms that accomplish the same task. This paper introduces a model of communication performance for NBL computers and uses the model to analyze the overall performance of three algorithms for vector combining (global sum) on the Intel Touchstone DELTA computer. Each of the three algorithms is found to be at least 33% faster than the other two for some combinations of machine size and vector length. The NBL constraint is shown to significantly affect the conditions under which each algorithm is fastest.
- Research Organization:
- Pacific Northwest National Lab. (PNNL), Richland, WA (United States)
- Sponsoring Organization:
- USDOE, Washington, DC (United States)
- DOE Contract Number:
- AC06-76RL01830
- OSTI ID:
- 10133355
- Report Number(s):
- PNL-SA-20425; ON: DE92010829
- Resource Relation:
- Other Information: PBD: 13 Jan 1992
- Country of Publication:
- United States
- Language:
- English
Similar Records
Global combine on mesh architectures with wormhole routing
An asynchronous algorithm for scattering information between the active nodes of a multicomputer system