Re: [PATCH v3] mm: add thp_utilization metrics to /proc/thp_utilization

[Yu Zhao]

On Thu, Aug 11, 2022 at 4:12 PM Yang Shi <shy828301@xxxxxxxxx> wrote:
>
> On Thu, Aug 11, 2022 at 2:55 PM Yu Zhao <yuzhao@xxxxxxxxxx> wrote:
> >
> > On Thu, Aug 11, 2022 at 1:20 PM Alex Zhu (Kernel) <alexlzhu@xxxxxx> wrote:
> > >
> > > Hi Yu,
> > >
> > > I’ve updated your patch set from last year to work with folio and am testing it now. The functionality in split_huge_page() is the same as what I have. Was there any follow up work done later?
> >
> > Yes, but it won't change the landscape any time soon (see below). So
> > please feel free to continue along your current direction.
> >
> > > If not, I would like to incorporate this into what I have, and then resubmit. Will reference the original patchset. We need this functionality for the shrinker, but even the changes to split_huge_page() by itself it should show some performance improvement when used by the existing deferred_split_huge_page().
> >
> > SGTM. Thanks!
> >
> > A side note:
> >
> > I'm working on a new mode: THP=auto, meaning the kernel will detect
> > internal fragmentation of 2MB compound pages to decide whether to map
> > them by PMDs or split them under memory pressure. The general workflow
> > of this new mode is as follows.
>
> I tend to agree that avoiding allocating THP in the first place is the
> preferred way to avoid internal fragmentation. But I got some
> questions about your design/implementation:
>
> >
> > In the page fault path:
> > 1. Compound pages are allocated as usual.
> > 2. Each is mapped by 512 consecutive PTEs rather than a PMD.
> > 3. There will be more TLB misses but the same number of page faults.
> > 4. TLB coalescing can mitigate the performance degradation.
>
> Why not just allocate base pages in the first place? Khugepaged has
> max_pte_none tunable to detect internal fragmentation. If you worry
> about zero page, you could add max_pte_zero tunable.
>
> Or did you investigate whether the new MADV_COLLAPSE may be helpful or
> not? It leaves the decision to the userspace.

There are two problems we have to workaround.
1. External fragmentation that prevents later compound page allocations.
2. The cost of taking mmap_lock for write.

IIRC, the first reference I listed describes the first problem. (It
uses a similar reservation technique.) From a very high level, smaller
page allocations add more entropy than larger ones and accelerate the
system toward equilibrium, in which state the system can't allocate
more THPs without exerting additional force (compaction).

Reserving compound pages delays the equilibrium whereas MADV_COLLAPSE
tries to reverse the equilibrium. The latter has a higher cost. In
addition, it needs to take mmap_lock for write.

> > In khugepaged:
> > 1. Check the dirty bit in the PTEs mapping a compound page, to
> > determine its utilization.
> > 2. Remap compound pages that meet a certain utilization threshold by
> > PMDs in place, i.e., no migrations.
> >
> > In the reclaim path, e.g., MGLRU page table scanning:
> > 1. Decide whether compound pages mapped by PTEs should be split based
> > on their utilizations and memory pressure, e.g., reclaim priority.
> > 2. Clean subpages should be freed directly after split, rather than swapped out.
> >
> > N.B.
> > 1. This workflow relies on the dirty bit rather examining the content of a page.
> > 2. Sampling can be done by periodically switching between a PMD and
> > 512 consecutive PTEs.
> > 3. It only needs to hold mmap_lock for read because this special mode
> > (512 consecutive PTEs) is not considered the split mode.
> > 4. Don't hold your breath :)
> >
> > Other references:
> > 1. https://www.usenix.org/system/files/atc20-zhu-weixi_0.pdf
> > 2. https://www.usenix.org/system/files/osdi21-hunter.pdf