On Wed, Feb 19, 2020 at 11:31:39AM -0800, Andrew Morton wrote: > But what was the nature of these stalls? If they were "stuck in D > state waiting for something" then that's throttling. If they were > "unexpected bursts of in-kernel CPU activity" then I see a better case. It was both. However, the workload was able to perform with no direct reclaim activity and no stalls, while memory.high semantics were never violated. This means that allocation rate was not outstripping reclaim rate, just that both of these things happened in bursts and sputters. If the workload isn't exceeding the memory.high grace buffer, it seems unnecessary to make it wait behind a reclaim invocation that gets delayed on IO but will succeed before the workload comes back for the next allocation. We can just let it get on with its thing, actually work with the memory it just obtained, while background reclaim will free pages as they become reclaimable. This is a win-win situation.
