Memory utilization

Memory is not managed as a single component, such as a CPU or disk, but as a collection of small components called pages.

The size of a typical page in memory can range from 1 to 8 kilobytes, depending on your operating system. A computer with 64 megabytes of memory and a page size of 2 kilobytes contains approximately 32,000 pages.

When the operating system needs to allocate memory for use by a process, it scavenges any unused pages within memory that it can find. If no free pages exist, the memory-management system has to choose pages that other processes are still using and that seem least likely to be needed in the short run. CPU cycles are required to select those pages. The process of locating such pages is called a page scan. CPU utilization increases when a page scan is required.

Memory-management systems typically use a least recently used algorithm to select pages that can be copied out to disk and then freed for use by other processes. When the CPU has identified pages that it can appropriate, it pages out the old page images by copying the old data from those pages to a dedicated disk. The disk or disk partition that stores the page images is called the swap disk, swap space, or swap area. This paging activity requires CPU cycles as well as I/O operations.

Eventually, page images that have been copied to the swap disk must be brought back in for use by the processes that require them. If there are still too few free pages, more must be paged out to make room. As memory comes under increasing demand and paging activity increases, this activity can reach a point at which the CPU is almost fully occupied with paging activity. A system in this condition is said to be thrashing. When a computer is thrashing, all useful work comes to a halt.

To prevent thrashing, some operating systems use a coarser memory-management algorithm after paging activity crosses a certain threshold. This algorithm is called swapping. When the memory-management system resorts to swapping, it appropriates all pages that constitute an entire process image at once, rather than a page at a time.

Swapping frees up more memory with each operation. However, as swapping continues, every process that is swapped out must be read in again, dramatically increasing disk I/O to the swap device and the time required to switch between processes. Performance is then limited to the speed at which data can be transferred from the swap disk back into memory. Swapping is a symptom of a system that is severely overloaded, and throughput is impaired.

Many systems provide information about paging activity that includes the number of page scans performed, the number of pages sent out of memory (paged out), and the number of pages brought in from memory (paged in):

Paging out is the critical factor because the operating system pages out only when it cannot find pages that are free already.
A high rate of page scans provides an early indicator that memory utilization is becoming a bottleneck.
Pages for terminated processes are freed in place and simply reused, so paging-in activity does not provide an accurate reflection of the load on memory. A high rate of paging in can result from a high rate of process turnover with no significant performance impact.

Although the principle for estimating the service time for memory is the same as that described in Resource utilization and performance, you use a different formula to estimate the performance impact of memory utilization than you do for other system components.

You can use the following formula to calculate the expected paging delay for a given CPU utilization level and paging rate:

PD= (C/(1-U)) * R * T

PD: is the paging delay.
C: is the CPU service time for a transaction.
U: is the CPU utilization (expressed as a decimal).
R: is the paging-out rate.
T: is the service time for the swap device.

As paging increases, CPU utilization also increases, and these increases are compounded. If a paging rate of 10 per second accounts for 5 percent of CPU utilization, increasing the paging rate to 20 per second might increase CPU utilization by an additional 5 percent. Further increases in paging lead to even sharper increases in CPU utilization, until the expected service time for CPU requests becomes unacceptable.