Distributed-sum termination detection supporting multithreaded execution
Parallel Computing
| Hi-index | 0.00 |
Efficient determination of processing termination atbarrier synchronization points can occupy an importantrole in the overall throughput of parallel and distributedcomputing systems. Even though relatively efficienttermination detection techniques have been proposed forcertain environments, no effective performance analysismethodology has been introduced to determine applicationattributes that favor the use of a particular terminationdetection technique. This fact has hindered the adoptionand development of termination detection schemes. Thispaper addresses this problem by developing acommunication pattern based methodology to improve theprecision of the theoretical performance of terminationdetection techniques in lieu of laborious experiments orpotentially subjective benchmarking studies. By measuringmessage complexity from the idle period respect, itprovides a simple and effective way to evaluate existingtermination detection techniques or design newtermination detection algorithms.