Linux load average explained: what is too high?
In short
The load average is the average number of processes that are running, waiting for a CPU or blocked in uninterruptible I/O, over the last 1, 5 and 15 minutes. It only becomes meaningful when divided by the number of CPU cores: a sustained value above 1.0 per core means work is queueing. On Linux a high load can also come from slow disks or network storage while the CPUs are idle.
The load average tells you how many processes, on average, wanted to run over the last 1, 5 and 15 minutes. Whether a value is too high depends on the number of CPU cores: a load of 4.0 is a fully used machine with four cores and a badly overloaded one with a single core. As a rule of thumb, a sustained load above 1.0 per core means that work is waiting. On Linux there is one important extra: processes blocked on disk or network I/O count as well, so a high load does not always mean a busy CPU.
Where do I see the load average?
uptime 21:11:22 up 3 days, 3:27, 1 user, load average: 0.99, 0.25, 0.14The same three numbers appear in the header of top and come from the kernel:
cat /proc/loadavg0.99 0.25 0.14 2/461 10988The fourth field shows currently runnable scheduling entities and the total number (2/461), the fifth is the most recently assigned process ID.
What do the three numbers mean?
They are exponentially damped moving averages over roughly 1, 5 and 15 minutes. They are not plain arithmetic means; recent activity weighs more, and old activity fades out gradually. Read them as a trend:
| Pattern | Meaning |
|---|---|
| 1 min above 5 min above 15 min | load is rising, something started recently |
| 1 min below 5 min below 15 min | load is falling, the peak is over |
| all three about equal and high | sustained overload, not a spike |
In the example above the 1-minute value is 0.99 while the 15-minute value is 0.14: something started a moment ago. For alerting, the 15-minute value is the calmer signal; the 1-minute value jumps with every cron job.
What is a high load average for my server?
Find the number of logical CPUs:
nproc4Then divide. A load of 3.2 on four cores is 0.8 per core.
| Load per core | Interpretation |
|---|---|
| below 0.7 | headroom for peaks |
| around 1.0 | fully used, no reserve |
| sustained above 1.0 | processes wait; response times grow |
| several times the core count | severe overload or blocked I/O |
Treat these as orientation, not as limits. A batch server may run at 1.0 per core all day and be perfectly healthy; an interactive web server feels slow well before that. Hyper-threads count as CPUs in nproc, but two threads on one core do not deliver twice the performance, so on such machines trouble starts a little earlier than the number suggests. On virtual servers, also keep an eye on steal time (see below): the cores you see are not always yours alone.
Why can the load be high while the CPU is idle?
On most Unix systems the load counts processes that are running or ready to run. Linux additionally counts tasks in uninterruptible sleep (state D), typically waiting for disk or network storage. That is deliberate: a machine whose processes all wait for a dying disk is overloaded, even if the CPU has nothing to do.
So the first question for any high load is: CPU or I/O?
vmstat 1 5procs -----------memory---------- ---swap-- -----io---- -system-- -------cpu-------
r b swpd free buff cache si so bi bo in cs us sy id wa st gu
6 0 0 812344 10148 451391 0 0 12 40 1317 2890 91 7 2 0 0 0
0 7 0 801120 10148 452034 0 0 9480 115 4201 6674 3 4 11 82 0 0The first line of vmstat output shows averages since boot; read the following ones. The two sample lines above illustrate the two cases (they would not normally appear together):
rhigh,usplussynear 100,wanear 0: CPU-bound. More processes want CPU than there are cores.bhigh,wahigh,idmoderate: I/O-bound. Processes are blocked waiting for storage.siandsonot zero: the system is swapping, which turns a memory shortage into an I/O problem.sthigh: the hypervisor gives CPU time to other guests; your processes wait although nothing on your server is at fault.
How do I find the processes behind the load?
List everything that is currently running (R) or in uninterruptible sleep (D):
ps -eo state,pid,user,comm --sort=state | awk '$1 ~ /^[RD]/'For CPU-bound load, sort by CPU usage:
ps -eo pid,user,comm,%cpu --sort=-%cpu | head -8
top -o %CPUFor I/O-bound load, the sysstat package (apt install sysstat, same package name on the RHEL family) provides the detail:
iostat -x 1 3 # per device: %util and await
pidstat -d 1 3 # per process: kB read and written per secondA device with %util near 100 and a growing await is saturated. If a network mount hangs, df or ls on that path will hang as well, and every process touching it adds one to the load.
Current kernels also expose pressure stall information, which answers the question "is anything actually waiting?" more directly than load does:
cat /proc/pressure/cpu
cat /proc/pressure/ioThe avg10, avg60 and avg300 values are the percentage of time in which at least some tasks were stalled on that resource. The files exist only if the kernel has the feature enabled, which is the case on most current distribution kernels.
What load average does not tell you
Load is a measure of demand on the kernel's run queue, nothing more. It does not see a web server whose workers are all occupied by slow clients, or a PHP-FPM pool that has reached its process limit: those processes sleep while they wait for a socket, so load stays low while visitors see timeouts. If the site is slow and the load is harmless, look at the service itself, for example as described in PHP-FPM: server reached pm.max_children.
The opposite also happens: a brief load peak during a nightly backup is harmless, and an alert for it only teaches people to ignore alerts.
Common mistakes
- Judging the number without the core count. "Load 8" says nothing until you know whether the machine has 2 or 32 cores.
- Equating load with CPU usage. A load of 20 with idle CPUs points to storage, not to a need for more cores.
- Alerting on the 1-minute value. It reacts to every short job. Use the 5- or 15-minute value, per core, and require it to persist.
- Comparing servers by raw load. Compare load per core, and compare like with like: a database server and a build machine have different normal levels.
- Ignoring steal time on virtual machines. If
stis high, tuning your own processes will not help; talk to the hosting provider or move the machine. - Looking at load alone during an incident. Combine it with memory, swap activity, disk fill level and the service metrics. A full disk, for example, often shows up as hanging processes first.
Checklist
uptime: read the three values as a trend, rising or falling?nproc: divide the load by the number of cores.vmstat 1 5: is it CPU (r,us,sy) or I/O (b,wa)?- Swapping? Check
siandso. - Virtual machine? Check
st. pssorted by CPU, oriostat -xandpidstat -d, to find the process or device.- If load is low but the site is slow, check web server workers and PHP-FPM.
- Set alert thresholds per core and on the calmer averages.
How Lynx Heartbeat helps
The Lynx Heartbeat agent measures the load every minute and relates it to the number of cores of each server, so the same thresholds make sense on a small virtual machine and on a large database host. Next to it you see CPU, memory, swap, disks and the processes that currently use the most, by name only. Because load misses saturated services, the agent also reads the status of Apache, PHP-FPM and MySQL. Alerts are debounced, so a short peak during a backup does not wake anyone; more on that in alert fatigue, quiet hours and escalation. The full list of measurements is on the features page.
Frequently asked questions
What is a good load average on Linux?
There is no absolute number. Divide the load by the number of CPU cores from nproc: sustained values below about 0.7 per core are comfortable, around 1.0 per core the machine is fully used, and clearly above 1.0 per core work is waiting.
What do the three load average numbers mean?
They are exponentially damped averages over roughly the last 1, 5 and 15 minutes. If the 1-minute value is higher than the 15-minute value, load is rising; if it is lower, load is falling.
Is load average the same as CPU usage?
No. CPU usage is the share of time the processors were busy. Load average counts processes that are running or waiting, and on Linux it includes processes blocked in uninterruptible I/O. A server can have a high load with almost idle CPUs.
Why is the load high although the CPU is idle?
Processes are waiting for something other than CPU, usually a slow or overloaded disk, a hanging network mount or swapping. Check the wa column in vmstat or %iowait in mpstat and look for processes in state D.
Can a website be down while the load average is low?
Yes. If all web server workers or all PHP-FPM children are occupied by waiting connections, new requests queue up while the processes mostly sleep. Load does not show this; the service's own status page does.