lkml.org 
[lkml]   [2019]   [Aug]   [23]   [last100]   RSS Feed
Views: [wrap][no wrap]   [headers]  [forward] 
 
Messages in this thread
/
Date
From
SubjectRe: [PATCH v6 0/6] mm / virtio: Provide support for unused page reporting

> On Thu, 2019-08-22 at 06:43 -0400, Pankaj Gupta wrote:
> > > This series provides an asynchronous means of reporting to a hypervisor
> > > that a guest page is no longer in use and can have the data associated
> > > with it dropped. To do this I have implemented functionality that allows
> > > for what I am referring to as unused page reporting
> > >
> > > The functionality for this is fairly simple. When enabled it will
> > > allocate
> > > statistics to track the number of reported pages in a given free area.
> > > When the number of free pages exceeds this value plus a high water value,
> > > currently 32, it will begin performing page reporting which consists of
> > > pulling pages off of free list and placing them into a scatter list. The
> > > scatterlist is then given to the page reporting device and it will
> > > perform
> > > the required action to make the pages "reported", in the case of
> > > virtio-balloon this results in the pages being madvised as MADV_DONTNEED
> > > and as such they are forced out of the guest. After this they are placed
> > > back on the free list, and an additional bit is added if they are not
> > > merged indicating that they are a reported buddy page instead of a
> > > standard buddy page. The cycle then repeats with additional non-reported
> > > pages being pulled until the free areas all consist of reported pages.
> > >
> > > I am leaving a number of things hard-coded such as limiting the lowest
> > > order processed to PAGEBLOCK_ORDER, and have left it up to the guest to
> > > determine what the limit is on how many pages it wants to allocate to
> > > process the hints. The upper limit for this is based on the size of the
> > > queue used to store the scattergather list.
> > >
> > > My primary testing has just been to verify the memory is being freed
> > > after
> > > allocation by running memhog 40g on a 40g guest and watching the total
> > > free memory via /proc/meminfo on the host. With this I have verified most
> > > of the memory is freed after each iteration.
> >
> > I tried to go through the entire patch series. I can see you reported a
> > -3.27 drop from the baseline. If its because of re-faulting the page after
> > host has freed them? Can we avoid freeing all the pages from the guest
> > free_area
> > and keep some pages(maybe some mixed order), so that next allocation is
> > done from
> > the guest itself than faulting to host. This will work with real workload
> > where
> > allocation and deallocation happen at regular intervals.
> >
> > This can be further optimized based on other factors like host memory
> > pressure etc.
> >
> > Thanks,
> > Pankaj
>
> When I originally started implementing and testing this code I was seeing
> less than a 1% regression. I didn't feel like that was really an accurate
> result since it wasn't putting much stress on the changed code so I have
> modified my tests and kernel so that I have memory shuffting and THP
> enabled. In addition I have gone out of my way to lock things down to a
> single NUMA node on my host system as the code I had would sometimes
> perform better than baseline when running the test due to the fact that
> memory was being freed back to the hose and then reallocated which
> actually allowed for better NUMA locality.
>
> The general idea was I wanted to know what the worst case penalty would be
> for running this code, and it turns out most of that is just the cost of
> faulting back in the pages. By enabling memory shuffling I am forcing the
> memory to churn as pages are added to both the head and tail of the
> free_list. The test itself was modified so that it didn't allocate order 0
> pages and instead was allocating transparent huge pages so the effects
> were as visible as possible. Without that the page faulting overhead would
> mostly fall into the noise of having to allocate the memory as order 0
> pages, that is what I had essentially seen earlier when I was running the
> stock page_fault1 test.

Right. I think the reason is this test is allocating THP's in guest, host side
you are still using order 0 pages, I assume?

>
> This code does no hinting on anything smaller than either MAX_ORDER - 1 or
> HUGETLB_PAGE_ORDER pages, and it only starts when there are at least 32 of
> them available to hint on. This results in us not starting to perform the
> hinting until there is 64MB to 128MB of memory sitting in the higher order
> regions of the zone.

o.k

>
> The hinting itself stops as soon as we run out of unhinted pages to pull
> from. When this occurs we let any pages that are freed after that
> accumulate until we get back to 32 pages being free in a given order.
> During this time we should build up the cache of warm pages that you
> mentioned, assuming that shuffling is not enabled.

I was thinking about something like retaining pages to a lower watermark here.
Looks like we still might have few lower order pages in free list if they are
not merged to orders which are hinted.

>
> As far as further optimizations I don't think there is anything here that
> prevents us from doing that. For now I am focused on just getting the
> basics in place so we have a foundation to start from.

Agree. Thanks for explaining.

Best rgards,
Pankaj

>
> Thanks.
>
> - Alex
>
>

\
 
 \ /
  Last update: 2019-08-23 07:19    [W:0.821 / U:0.924 seconds]
©2003-2020 Jasper Spaans|hosted at Digital Ocean and TransIP|Read the blog|Advertise on this site