Possible Bug in CPU Usage Calculation — Only User Time Considered, Ignoring System/Other Times

[trevor211]

Hi,

I noticed that in perftest, the CPU usage is calculated using values from the first line of /proc/stat, but only the user and idle fields are used. Specifically, the calculation looks like this:

ustat_diff = user[1] - user[0];  // only user time
idle_diff = idle[1] - idle[0];   // idle time
cpu_usage = (ustat_diff / (ustat_diff + idle_diff)) * 100;

However, this approach seems incomplete because it ignores other CPU usage fields such as system, nice, irq, etc. A more accurate calculation should include all CPU time components. Typically, it should be:

total_diff = Δuser + Δnice + Δsystem + Δidle + Δiowait + Δirq + Δsoftirq + Δsteal + ...
idle_diff = Δidle + Δiowait
cpu_usage = (total_diff - idle_diff) / total_diff * 100

In short:

cpu_usage = (Δtotal - Δidle) / Δtotal

By only considering user time, the current calculation underestimates total CPU usage, especially under workloads that heavily use system or irq time.

This becomes particularly problematic in two-sided RDMA mode (e.g., RDMA Send/Recv), where the application might use event-driven programming (e.g., using poll/select/epoll). In such cases, user time remains low while system time or interrupt time may be high. As a result, the current method can misleadingly show near-zero CPU usage, even when the system is busy handling RDMA events.

Is this an oversight or intentional simplification? If it's not intentional, I believe this might be a bug worth correcting.

Thanks!