Interpret load and CPU

Load average is not a CPU percentage. It counts runnable tasks and tasks in uninterruptible sleep, averaged over 1, 5, and 15 minutes. That last part is why Linux load numbers confuse people. A machine doing zero CPU work can show a load of 20 if twenty processes are stuck waiting on a dead disk.

Read load against the core count

uptime
# 15:04:33 up 12 days,  3:12,  1 user,  load average: 6.42, 5.90, 2.15
nproc
# 4

Compare load with CPU count. Load 6.4 on 4 cores means demand exceeds capacity. On average, 2+ tasks are waiting at any moment. The three numbers also give you direction. Here the 15-minute average is 2.15, so this started recently and is getting worse.

Find out what kind of demand it is

top splits CPU time into categories on its %Cpu(s) line:

%Cpu(s):  8.1 us,  2.3 sy,  0.0 ni, 41.2 id, 48.0 wa,  0.0 hi,  0.4 si

The fields that matter: us is user code, sy is kernel work, id is idle, and wa is I/O wait. That is CPU sitting idle while tasks wait for storage. This sample is the classic pattern: load is high, but 41% of CPU is idle and 48% is wa. The bottleneck is the disk, not the processor. Chasing a “CPU problem” here wastes the afternoon.

Then check per-core numbers, because averages hide saturation:

mpstat -P ALL 5 1

CPU    %usr   %sys  %iowait   %idle
all   26.10   3.02     0.51   70.37
  0   99.20   0.60     0.00    0.20
  1    2.10   3.80     0.70   93.40

One saturated core can limit a single-threaded application while total CPU still looks moderate. 26% overall, but core 0 is pinned at 99% and that’s exactly where your Node process lives. pidstat 5 (or pressing 1 inside top) tells you which process owns the busy core.

Capture five timed samples during a controlled load rather than staring at one refresh. A single snapshot catches noise. Five samples show whether demand is user CPU, system CPU, I/O wait, or one constrained process.

The trap is the one this lesson opened with: high load with idle CPUs. Load counts D-state tasks too, so slow storage or a hung NFS mount inflates it. When load looks scary, check wa and process states before you blame the CPU.