On Thu, 2014-01-02 at 19:42 -0500, Debabrata Banerjee wrote:

Currently because of how mm behaves (3.10.y) the code even before the
patch is a problem. I believe what may fix it is if instead of just
removing the conditional on __GFP_WAIT, the initial order > 0
allocation should be made GFP_ATOMIC, then fallback to the original
gfp mask for the order-0 allocations.

On systems that have highly fragmented main memory with pressure,
skb_page_frag_refill() causes problems. mm enters significant
compaction cycles on all cpu's which in itself is bad (add
considerable spinlock contention in isolate_migratepages_range() for
several seconds in kernel at 100% cpu), however even without this
happening basically we have large memory reclaimation when only
allocations from order-3 were necessary. For example, I might see half
the existing page cache on a system (2GB out of 8GB) reclaimed in a
burst, which effectively means the application has to wait even longer
after this compact/reclaim cycle for those pages to be read back from
disk. This is a significant reduction in useful memory from before
skb_page_frag_refill() existed, as one of our systems could run in
steady state will little free memory and 100% fragmentation. Now I see
10-30x more memory free (read: not utilized). Order > 0 allocations
were happening rarely before, now it happens consistently from this
function.

My suggestion above would avoid mm going through
__alloc_pages_direct_compact() and triggering the bad events above. It
will take me several days to try this experiment.

My suggestion is to use a recent kernel, and/or eventually backport the
mm fixes if any.

order-3 allocations should not reclaim 2GB out of 8GB.

There is a reason PAGE_ALLOC_COSTLY_ORDER exists and is 3