A Linux box shows a load average of 40 but CPU utilization is only ten percent. What is going on?

On Linux the load average counts processes in uninterruptible sleep as well as runnable ones, so a high load with idle CPUs almost always means processes are blocked on I/O. Check iowait and device utilization with iostat, look for processes in D state, and consider a slow disk, a stalled network filesystem, or a saturated storage volume. Occasionally it is lock contention or a hung mount rather than throughput.

Load average on Linux is not a CPU queue, it counts runnable plus uninterruptible processes, so processes stuck waiting on disk or a network mount inflate it while the CPUs sit idle. So my first move is to confirm that: ps looking for processes in D state, and vmstat or top to check the iowait column. Then iostat to see whether a device is at one hundred percent utilization with a nasty await time, which points at a saturated or failing disk, or at a cloud volume that has run out of burst credits, something I have hit more than once on a general purpose volume that was fine for months and then suddenly was not. If the D state processes are all touching the same mount, I look at the network filesystem, because a hung NFS export produces exactly this picture, huge load and nothing happening. If storage looks healthy I widen out to lock contention or a driver problem. The key point is that restarting the application here would tell me nothing, because the application is the victim, not the cause.