load average/run queue metrics

[Michal Toth]

Hi all,

i was wondering whether the load average metrics, as shown by uptime or top does take the number of cpu's into account. So if average lenght of run queue is let's say 7, does that translate to the same level of (indicated) utilization (other things equal) on system with four and on system with just one cpu?

thanks

[Patrick Wallek · Accepted solution]

In a sense, yes. If you look at the 'uptime' man page you will see that uptime reports "...the average number of jobs in the run queue over the last 1, 5, and 15 minutes for the active processors."

So a 1 minute runqueue of 7 on a 4 CPU system means that there was an average of 1.75 jobs waiting on each processor. On a single CPU system, there were 7 jobs waiting on the processor.

[Michal Toth]

so there is only one system-wide run queue (well perhaps more, if processor sets are used)?

[Bill Hassell]

> so there is only one system-wide run queue (well perhaps more, if processor sets are used)?

For uptime and top, the run queue is global and is built by linking all runnable processes and threads together. Technically, there are two run queues: global and per-processor. Processor sets and other workload managers essentially adjust internal priorities but the run queue metric remains as defined.

Note that the run queue is all running and ready-to-run processes. A process is ready to run when it is waiting for nothing (like I/O or a system call). The run queue is a misleading metric because it is only measuring CPU cycles, not total workload. In general, an ideal run queue is approximately equal to the number of processors for simple compute-bound processes. But there are many circumstances where the run queue can be much larger than the processor count and yet the system runs fine. It is only one of many metrics needed to analyze system performance.

[Michal Toth]

thanks for bringing some light into the issue