A) Stream a T2/t1(stream A) T3/t1(local Stream) T2/t1(local Stream) T3/t1(stream A) (a) Stream a T2/t1(stream A) T3/t1(local Stream) T2/t1(local Stream) T3/t1(stream A

25 Modeling and simulation of a network interface handling two priority levels. We propose a scheme called Double Queue Scheme to reduce the eeects of local jobs' communication on the performance of parallel jobs. It divides the only one input and output queues in traditional network interfaces into two ones. One is for the communication of parallel jobs, and the other is for the communication of local jobs. Two queues are given diierent priorities to adjust the allocation of communication bandwidth among them. By giving the parallel jobs' queue a higher priority, through simulation, we nd that the scheme will improve the performance of parallel jobs considerablely with a slight eeect on local jobs communication especially when the sizes of local jobs' communication is small and moderate. How to schedule parallel jobs and local jobs is another issue that will aaect the performance of both jobs. We have proposed a power preservation approach 6] to schedule the jobs in a NOW. But the communication overhead is not considered there. Currently, we are integrating them and evaluating the performance when both processors and networks are taken into account. Acknowledgement: We are grateful to the anonymous referees for their rigorous and constructive comments. 24 queue scheme in the message lter under the same condition as 2), and T 3 was measured. Again, the curves of T 2 =T 1 in the environment without using the double queue scheme and T 3 =T 1 using the double queue scheme are plotted in Figure 10 to observe the slowdowns of both the testing streams (streams A and B) and the local stream when the message size of the local stream increased from 2 Kbytes to 16 Kbytes. (b) stream B T2/T1(stream B) T3/T1(local stream) T2/T1(local stream) T3/T1(stream B) Figure 10: Experiments with the one-to-all and all-to-one combinations. Figure 10(a) and (b) show that the message transmission times of both testing streams A and B were delayed up to 45 times and 30 times, respectively, when message types are not distinguished. The local stream itself were delayed by streams A and B up to 7 and 12 times, respectively. The double queue scheme in the message lter reduced the delay for stream A to 6 times (about 8 times reduction), and for stream B to 9 times (3 times reduction). The message lter is more eeective in reducing the message transmission delay on a small …