CPU

A live sample, not a lifetime average. The collector reads /proc twice a second apart and reports the difference, so this is what the user is running now — unlike ps -eo pcpu, where a job that pinned a core for six hours and then went idle still reports a huge number.

100% is one hardware thread fully busy. It is not a share of the machine: a leviathan reports 32 threads, so it tops out at 3200%. The figure under each percentage is the same number divided by 100, and it covers the whole fleet. Expanding a user's row breaks it down by leviathan: how many threads they are using on each machine they are logged into, next to what they have resident there. A host at 0.0 is one they are logged into with nothing running.

Memory

PSS (proportional set size): a page shared between several of a user's processes is divided among them. So a forked R job counts its shared memory once across all its workers rather than once per worker, which is the figure you want when deciding who is actually filling the box.

Load (1 / 5 / 15m)

The Unix load average: how many processes are using a CPU or waiting for one, averaged over the last 1, 5 and 15 minutes. On Linux it also counts processes blocked on uninterruptible I/O.

It is a count, not a percentage — compare it against the host's thread count. On a leviathan (32 threads):

The three numbers show direction. 120 / 40 / 15 means load spiked in the last minute; 15 / 40 / 120 means a big job just finished and the box is draining. Equal numbers mean steady state.

High load with low CPU is an I/O problem, not a CPU one. Because uninterruptible I/O wait counts toward load, a stuck NFS mount to the NAS can drive load into the hundreds while the CPU column sits near zero. That combination is the most useful thing this column tells you.

Snapshot age

Nothing here is live-streamed. Each host writes its own snapshot from cron once a minute; this page fetches all 11 of them in parallel when you load it and stores nothing. So an age of a few seconds up to a minute is normal, and the CPU numbers can be up to a minute behind what is happening right now.

Past 10 minutes a host is flagged stale — its collector has most likely stopped. A host that could not be reached at all is marked down and is left out of every total on this page.