Re: [External] Re: [PATCH v14 0/8] Free some vmemmap pages of HugeTLB page

[Muchun Song]

On Sat, Feb 6, 2021 at 12:01 AM Joao Martins <joao.m.martins@xxxxxxxxxx> wrote:
>
> On 2/4/21 3:50 AM, Muchun Song wrote:
> > Hi all,
> >
>
> [...]
>
> > When a HugeTLB is freed to the buddy system, we should allocate 6 pages for
> > vmemmap pages and restore the previous mapping relationship.
> >
> > Apart from 2MB HugeTLB page, we also have 1GB HugeTLB page. It is similar
> > to the 2MB HugeTLB page. We also can use this approach to free the vmemmap
> > pages.
> >
> > In this case, for the 1GB HugeTLB page, we can save 4094 pages. This is a
> > very substantial gain. On our server, run some SPDK/QEMU applications which
> > will use 1024GB hugetlbpage. With this feature enabled, we can save ~16GB
> > (1G hugepage)/~12GB (2MB hugepage) memory.
> >
> > Because there are vmemmap page tables reconstruction on the freeing/allocating
> > path, it increases some overhead. Here are some overhead analysis.
>
> [...]
>
> > Although the overhead has increased, the overhead is not significant. Like Mike
> > said, "However, remember that the majority of use cases create hugetlb pages at
> > or shortly after boot time and add them to the pool. So, additional overhead is
> > at pool creation time. There is no change to 'normal run time' operations of
> > getting a page from or returning a page to the pool (think page fault/unmap)".
> >
>
> Despite the overhead and in addition to the memory gains from this series ...
> there's an additional benefit there isn't talked here with your vmemmap page
> reuse trick. That is page (un)pinners will see an improvement and I presume because
> there are fewer memmap pages and thus the tail/head pages are staying in cache more
> often.
>
> Out of the box I saw (when comparing linux-next against linux-next + this series)
> with gup_test and pinning a 16G hugetlb file (with 1G pages):
>
>         get_user_pages(): ~32k -> ~9k
>         unpin_user_pages(): ~75k -> ~70k
>
> Usually any tight loop fetching compound_head(), or reading tail pages data (e.g.
> compound_head) benefit a lot. There's some unpinning inefficiencies I am fixing[0], but
> with that in added it shows even more:
>
>         unpin_user_pages(): ~27k -> ~3.8k
>
> FWIW, I was also seeing that with devdax and the ZONE_DEVICE vmemmap page reuse equivalent
> series[1] but it was mixed with other numbers.

It's really a surprise. Thank you very much for the test data.
Very nice. Thanks again.


>
> Anyways, JFYI :)
>
>         Joao
>
> [0] https://lore.kernel.org/linux-mm/20210204202500.26474-1-joao.m.martins@xxxxxxxxxx/
> [1] https://lore.kernel.org/linux-mm/20201208172901.17384-1-joao.m.martins@xxxxxxxxxx/