Understanding Linux Load Average: Principles and Calculations

This article provides a comprehensive explanation of Linux load average, addressing common misconceptions and detailing the underlying mechanisms. It begins by explaining how load average is typically viewed using the top command, which retrieves data from the /proc/loadavg pseudo-file. The article then traces the data flow from user space to the kernel, explaining how the kernel defines and handles this pseudo-file.

The core of the article explains how load average is calculated in the kernel. It describes a two-step process: first, a high-resolution timer periodically collects the number of running and uninterruptible processes from each CPU's run queue, aggregating this into a system-wide instantaneous load value. Second, a timer interrupt uses an exponential weighted moving average (EWMA) algorithm to compute the 1, 5, and 15-minute load averages from this instantaneous value.

The article clarifies that modern Linux load average includes not only runnable processes but also uninterruptible processes (those waiting on I/O), which is why load average is not always directly correlated with CPU usage. It explains the historical context for this design choice, noting that it was introduced to better reflect overall system resource demand, not just CPU demand.

Finally, the article summarizes the key points: how load average is calculated, why it may not correlate directly with CPU usage, and how the kernel exposes this data to user applications through the /proc/loadavg interface.