A CPU-bound system's problems will be anticipated in the load average. This is one of the most often misunderstood performance metrics. A system's load is the number of processes that are ready to run, including ones currently running. This number is measured at regular, frequent intervals. The load average, therefore, is the average of this load number over a particular interval of time. There actually are three load averages measured over different periods of time: the most recent minute, the last five minutes, and the last 15 minutes.
If there are four processes that are performing computations while the system is doing nothing else, then the load average will be approximately four. The three different measurements are present to indicate recent trends. If the one-minute load average is ten, the five-minute figure is four, and the 15-minute average is one, then you know not only that you have a heavily loaded machine, but that the load has increased rapidly over the past few minutes. If you got rid of several runaway processes, you would expect the one-minute average to drop quickly, but the 15-minute average would continue to remain high.
But what is a "high" load average? There is no single answer to this question. The average does not consider the number of processors in a system. A load average of eight on a single processor machine could be indicative of a problem, but on a four-processor system it probably is normal.
What type of process is "ready to run"? Any process that needs some time on the CPU meets that description. And a process that is currently running on a CPU is also considered "ready to run" (some older versions of SunOS [5.0, 5.1, and 5.2] got that wrong). But the system also counts processes that are in a state called "short-term disk wait." Such processes are not quite ready to be scheduled for a CPU, but will be soon.
The inclusion of processes in short-term wait can skew the load average higher than one might expect. A process that spends a great deal of time reading and writing to a local disk will be included in the load average. Waiting for input from a tty device is never considered a short-term wait, so a process waiting for user input is not counted in the load average. A process that spends most of its time reading and writing to a serial line, such as uucico, kermit, or Z-modem, will not count toward a higher load. In most cases, network I/O is not considered short-term. But there is one important exception: NFS. A process waiting for an NFS read is regarded as a short-term disk wait. As a result, if an NFS server goes down, the load average on the clients will soar even though the CPU itself is not being used.
The load average is calculated primarily for the purposes of the UNIX process-scheduling algorithm. The current system load is included when calculating a process's CPU utilization, which in turn will affect the priority that a process is given for scheduling. Because it is convenient to see that number, utilities let us examine the load. But remember that this number is an internal kernel event calculated for a specific purpose, and may not always make sense to you and me.
The load averages should not be used as the sole means of assessing a system's performance. The average is an excellent indicator, but other information also needs to be considered. If you see a high load average, you should find out why it's high. The answer may be that the high number is actually perfectly normal and that nothing needs to be done to improve performance.