Issue

• THP does not obey the zone_reclaim logic.
• The first symptom that you might notice is a spike in CPU load however it's deeper than that.
• The step in the VM which takes a look at the requested allocation and decides whether to steal it from another NUMA node or allocate it locally just isn't there for THP.
• That hinges on the zone_reclaim variable and THP does not look at that variable.
• What will happen is instead of allocating locally to the NUMA node, it allocates off of another NUMA node.
• This causes THP to be fragmented and then reassembled over and over which in turn causes causes a CPU load spike.
• There were other flaws in earlier versions of THPs which also causes the same symptom, one of those was addressed in 783284.
• Theoretically adding code to have THP obey the zone reclaim variable should be easy with the requisite kernel knowledge/skill.
• It isn't hard to add the 3-8 lines of code, the problem is all the other subtle implications.