Messages in this thread | | | Date | Wed, 12 Dec 2018 10:50:51 +0100 | From | Michal Hocko <> | Subject | Re: [LKP] [mm] ac5b2c1891: vm-scalability.throughput -61.3% regression |
| |
On Tue 11-12-18 16:37:22, David Rientjes wrote: [...] > Since it depends on the workload, specifically workloads that fit within a > single node, I think the reasonable approach would be to have a sane > default regardless of the use of MADV_HUGEPAGE or thp defrag settings and > then optimzie for the minority of cases where the workload does not fit in > a single node. I'm assuming there is no debate about these larger > workloads being in the minority, although we have single machines where > this encompasses the totality of their workloads.
Your assumption is wrong I believe. This is the fundamental disagreement we are discussing here. You are essentially arguing for node_reclaim (formerly zone_reclaim) behavior for THP pages. All that without any actual data on wider variety of workloads. As the matter of _fact_ we know that node_reclaim behavior is not a suitable default. We did that mistake in the past and we had to revert that default _exactly_ because a wider variety of workloads suffered from over reclaim and performance issues as a result of constant reclaim
You have also haven't explained why you do care so much about remote THP while you do not care about remote base bages (the page allocator falls back to those as soon as the kswapd doesn't keep pace with the allocation rate; THP or high order pages in general is analoguous with kcompactd doing a pro-active compaction). Like the base pages we do not want larger pages to fallback to a remote node too easily. There is no question about that I believe.
I can be convinced that larger pages really require a different behavior than base pages but you should better show _real_ numbers on a wider variety workloads to back your claims. I have only heard hand waving and a very vague and quite doubtful numbers for a non-disclosed benchmark without a clear indication on how it relates to real world workloads. So color me unconvinced. -- Michal Hocko SUSE Labs
| |