Using the Data

The outputs from the monitoring utilities can provide insights into obtaining the best performance from your cluster. If you are new to cluster computing, you will want to note the relationship between the different machine resources, including CPU utilization, swap usage, and network utilization. Following are some useful guidelines:

• Low CPU usage with high network traffic might indicate that your system is I/O bound and could benefit from faster network components.

• Low network load and high CPU usage indicate that your system performance could improve with faster CPUs.

• Medium to high swap usage is always bad. This indicates that memory is oversubscribed, and application pieces must be moved to the much slower disk sub-system. This can be a substantial bottleneck, and is a sure sign that additional RAM is needed.

Any of these issues could be helped with application optimization, but sometimes it is more economical to add resources than to change working software.

For best performance of a computational workload, make sure your compute nodes have ample memory for the application and problem set. Also, use diskless compute nodes or configure local disks for scratch file space rather than swap space.