Re: [PATCH 2/5] mm/page_alloc: Add a bulk page allocator

On Fri, Mar 12, 2021 at 04:03:50PM +0000, Mel Gorman wrote:

On Fri, Mar 12, 2021 at 02:58:14PM +0000, Matthew Wilcox wrote:

quoted

On Fri, Mar 12, 2021 at 12:46:09PM +0100, Jesper Dangaard Brouer wrote:

quoted

In my page_pool patch I'm bulk allocating 64 pages. I wanted to ask if
this is too much? (PP_ALLOC_CACHE_REFILL=64).

The mlx5 driver have a while loop for allocation 64 pages, which it
used in this case, that is why 64 is chosen.  If we choose a lower
bulk number, then the bulk-alloc will just be called more times.

The thing about batching is that smaller batches are often better.
Let's suppose you need to allocate 100 pages for something, and the page
allocator takes up 90% of your latency budget.  Batching just ten pages
at a time is going to reduce the overhead to 9%.  Going to 64 pages
reduces the overhead from 9% to 2% -- maybe that's important, but
possibly not.

I do not think that something like that can be properly accessed in
advance. It heavily depends on whether the caller is willing to amortise
the cost of the batch allocation or if the timing of the bulk request is
critical every single time.

quoted

quoted

The result of the API is to deliver pages as a double-linked list via
LRU (page->lru member).  If you are planning to use llist, then how to
handle this API change later?

Have you notice that the two users store the struct-page pointers in an
array?  We could have the caller provide the array to store struct-page
pointers, like we do with kmem_cache_alloc_bulk API.

My preference would be for a pagevec.  That does limit you to 15 pages
per call [1], but I do think that might be enough.  And the overhead of
manipulating a linked list isn't free.

I'm opposed to a pagevec because it unnecessarily limits the caller.  The
sunrpc user for example knows how many pages it needs at the time the bulk
allocator is called but it's not the same value every time.  When tracing,
I found it sometimes requested 1 page (most common request actually) and
other times requested 200+ pages. Forcing it to call the batch allocator
in chunks of 15 means the caller incurs the cost of multiple allocation
requests which is almost as bad as calling __alloc_pages in a loop.

Well, no.  It reduces the cost by a factor of 15 -- or by 93%.  200 is
an interesting example because putting 200 pages on a list costs 200 *
64 bytes of dirty cachelines, or 12KiB.  That's larger than some CPU L1
caches (mine's 48KB, 12-way set associative), but I think it's safe to say
some of those 200 cache lines are going to force others out into L2 cache.
Compared to a smaller batch of 15 pages in a pagevec, it'll dirty two cache
lines (admittedly the 15 struct pages are also going to get dirtied by being
allocated and then by being set up for whatever use they're getting, but
they should stay in L1 cache while that's happening).

I'm not claiming the pagevec is definitely a win, but it's very
unclear which tradeoff is actually going to lead to better performance.
Hopefully Jesper or Chuck can do some tests and figure out what actually
works better with their hardware & usage patterns.

I think the first version should have an easy API to start with. Optimise
the implementation if it is a bottleneck. Only make the API harder to
use if the callers are really willing to always allocate and size the
array in advance and it's shown that it really makes a big difference
performance-wise.

I'm not entirely sure that a pagevec is harder to use than a list_head.