Sensitivity to Network Over-Utilization

This experiment attempts to isolate the effects of network use on each of the workloads. The test is performed as follows: a program (known affectionately as the "Network Hog") runs on the cluster (on machine 3, for example), reading a 100 MB file repeatedly from a server machine (machine 0, in the figure). The workload script is then run on on of the remaining machines (1 or 2), with its files located on the NFS Server. Remember, these machines are all on an ethernet; thus, the contention for the network is what is being measured.

Due to time constraints, we were only able to look at the CPU and I/O Workloads.

CPU Workload

Not surprisingly, the CPU Workload did not suffer much from network over-utilization. Since there is little or no I/O in all of the scripts, the most any one program was slowed by was about 6%.

IO Workload

In stark contrast to the CPU Workload, the I/O Workload suffers a fair amount, slowing down from 20% to 80%. By forcing processes to wait longer for I/O, we lessen CPU utilization. The point to be made here is a simple one: one process that claims a large fraction of a shared resource can have a detrimental affect on many processes in the system. The scheduling implication of this is that processes that can be classified as "network hogs" should be harnessed so that the whole system can make adequate progress.

We also hope to show (in the near future) the effects of the "network hog" on a switched-based network; in that experiment, the only contention will be for the switch resources.