Websites, applications and hosting for your business.
Recommended Services
Supported Scripts
Mr. Stig examines a queue of tasks behind a closed gate beside an active processing mechanism.

The server reports a load average near 2, while its processor appears almost idle. Before deciding that monitoring is wrong or that more CPU is required, identify what the two measurements describe. Linux can report elevated load because tasks are waiting without consuming execution time on the processor at that moment. Those conditions call for investigation rather than an automatic upgrade.

The distinction has practical consequences. Adding virtual processors may leave an application’s dependency wait unchanged. If multiple tasks genuinely compete for CPU, the evidence should instead show that pressure. Collect measurements over the same time window and connect them to the operation the user cannot complete satisfactorily. An isolated dashboard number does not identify that operation or its limiting resource.

Load average is not CPU utilisation

The three load-average numbers are smoothed values for one, five and fifteen minutes. Linux includes runnable tasks and tasks in uninterruptible wait. State R can mean executing or waiting for CPU time. State D indicates uninterruptible wait, often associated with I/O, but does not identify the cause by itself. These counts concern scheduling entities, including threads, rather than a percentage of processor activity.

A value of 2 has a different context with one available processor than with several. Even that comparison is only a starting point. Quotas, application constraints and per-core distribution can change usable capacity. Do not turn the number into an automatic instruction to buy a larger plan. First establish which resources the workload can actually use and what it is waiting for.

Two actual experiments with almost identical load

We used two independent temporary Linux guests in the v86 emulator, each with one virtual processor. The kernel was 4.16.13 i686, with vmstat from procps-ng 4.0.6 and basic iostat from BusyBox 1.35.0. This is an isolated educational experiment using an older kernel and emulated hardware. It neither benchmarks current VPS products nor reports a customer incident.

The first experiment ran four CPU workers. In the second, four writes waited on a deliberately frozen temporary filesystem inside the other guest. Snapshots were taken after forty seconds, followed by interval samples. Each case started independently so that load from the CPU case could not carry into the wait case. The filesystem was thawed afterwards and the writes completed.

Four CPU workers compared with four tasks waiting on the filesystem. The r and CPU-idle values are from subsequent vmstat samples.
MeasureCPUWaiting
1-minute load1.951.99
State4 Ă— R4 Ă— D
vmstat r40
CPU idle0%, 0%, 0%100%, 99%, 100%
vmstat wa0%0%

The subsequent iostat CPU sample also reported 0% idle in the CPU case and 99.01% idle in the wait case. Similar load values described different conditions. The first had work using the processor. The second had four tasks unable to proceed until the specific block was released. Load alone could not distinguish them, even though the practical next investigation would be different.

Why D state does not always imply high iowait

The laboratory cause was known: writes stopped at the deliberately frozen filesystem. All four wchan values showed call_rwsem_down_read_failed, while vmstat b and wa remained zero in these samples. D state and I/O-wait accounting do not describe exactly the same thing. A zero wa does not exclude every wait contributing to load, and the function name alone would not establish the same cause in another incident.

In a real incident, D alone is insufficient to blame storage. Identify what the task is waiting for and correlate it with evidence from that dependency. Zero throughput can also coexist with a stalled I/O operation. Examine completions, latency and queues alongside process states. The controlled freeze establishes the laboratory cause; it does not make a single column a universal diagnosis for production systems.

Keep the measurements in the same time window

On a system with full procps and sysstat tools, an initial read-only collection can include the following commands. A thread view helps when work is distributed across several application threads. Check the fields and permissions available on the particular system. Compact implementations such as BusyBox may support different options from full versions, so do not assume identical flags merely because the command names match.

uptime
vmstat 1 5
ps -eLo pid,tid,stat,wchan:28,comm
iostat -xzy 1 5

Read vmstat’s runnable count, user/system/idle CPU, swap and I/O together. Its initial CPU and I/O report covers time since boot; subsequent reports cover the selected interval. The initial iostat report has the same distinction, and y skips it in the example above. Comparing weeks of cumulative activity with one second of current behaviour can create a misleading explanation of the incident.

Preserve the time, CPU count and constraints, relevant tasks and the application symptom. Low aggregate utilisation on a multicore system can hide a busy thread or saturated core. Load also contains history: an earlier burst may already have ended. Interpret current states and the trend together, rather than assuming that a low instantaneous CPU percentage disproves every possible CPU-related limitation.

Choose the intervention from the limiting condition

Sustained competition for available CPU calls for examining workload and capacity. Waiting on a filesystem, service or synchronisation point calls for investigating that dependency. Our writes did not require a faster processor to resume; they required removal of the known blocking condition. A conclusion about a real server must be based on measurements from that server.

Related Stigmahost services

Stigmahost Virtual Private Servers and Dedicated Servers are infrastructure options to assess against the application’s actual requirements. Bring the incident time, relevant samples and the operation experiencing delay to an upgrade discussion. Those measurements help establish whether the application needs more resources or whether the cause of the wait must be addressed.

Share this Post