Clusters have become a very cost-effective platform forhigh-performance computing. In these systems, the trend is towards theinterconnection network becoming the system bottleneck. Therefore, inthe future, scheduling strategies will have to take into account thecommunication requirements of the applications and the communicationbandwidth that the network can offer. One of the key issues in thesestrategies is the task mapping technique used when the network becomesthe system bottleneck. In this paper, we propose an enhanced version ofa previously proposed mapping technique that takes into account not onlythe existing network resources, but also the traffic generated by theapplications. Also, we evaluate the mapping technique using real MPIapplication traces with timestamps. Evaluation results show that therise of the new mapping technique fully exploits the available networkbandwidth, improving load balancing and increasing the throughput thatcan be delivered by the network
展开▼
