Determining Cap Values

If a project cap is set too low, there might not be enough memory for the workload to proceed effectively under normal conditions. The paging that occurs because the workload requires more memory has a negative effect on system performance.

Projects that have caps set too high can consume available physical memory before their caps are exceeded. In this case, physical memory is effectively managed by the kernel and not by rcapd.

In determining caps on projects, consider these factors.

Impact on I/O system

The daemon can attempt to reduce a project workload's physical memory usage whenever the sampled usage exceeds the project's cap. During cap enforcement, the swap devices and other devices that contain files that the workload has mapped are used. The performance of the swap devices is a critical factor in determining the performance of a workload that routinely exceeds its cap. The execution of the workload is similar to running it on a machine with the same amount of physical memory as the workload's cap.

Impact on CPU usage

The daemon's CPU usage varies with the number of processes in the project workloads it is capping and the sizes of the workloads' address spaces.

A small portion of the daemon's CPU time is spent sampling the usage of each workload. Adding processes to workloads increases the time spent sampling usage.

Another portion of the daemon's CPU time is spent enforcing caps when they are exceeded. The time spent is proportional to the amount of virtual memory involved. CPU time spent increases or decreases in response to corresponding changes in the total size of a workload's address space. This information is reported in the vm column of rcapstat output. See Monitoring Resource Utilization With rcapstat and the rcapstat(1) man page for more information.

Reporting on shared memory

The daemon cannot determine which pages of memory are shared with other processes or which are mapped multiple times within the same process. Since rcapd assumes that each page is unique, this results in a discrepancy between the reported (estimated) RSS and the actual RSS.

Certain workloads, such as databases, use shared memory extensively. For these workloads, you can sample a project's regular usage to determine a suitable initial cap value. Use output from the prstat command with the -J option. See the prstat(1M) man page.