Conclusion



next up previous
Next: References Up: Possibilities for Active Messaging Previous: Optimization

Conclusion

Overall, active messaging could prove to be a worthwhile addition to the PVM suite if done carefully. However, an AM implementation will not provide an overwhelming ``out-of-the-box'' improvement over PVM's direct routed psend and precv APIs. AM will provide the users the tools to redesign their applications to communicate more efficiently. The asynchronous nature of AM combined with overlapped communication and computation can lead to orders of magnitude speedup, especially on algorithms that interleave massive data movement and computation (FFT for example). The most likely option for a first implementation of AM would be to use UDP or some relatively simple device-specific API. This would allow detailed study and suggest further optimizations to be made to the model and its algorithms. Successive experiments could then be performed using different network access methods until an acceptable balance of portability and performance is obtained. If it can be shown that the AM model can support PVM's level of functionality and that active messages routinely generate more robust and efficient distributed applications, then the issue of the conversion of the PVM suite to use this paradigm should be seriously considered.



Jack Dongarra
Tue Feb 7 21:45:39 EST 1995