On 23.01.20 15:05, David Hildenbrand wrote:
On 23.01.20 11:20, Alexander Graf wrote:
Hi Alex,
On 22.01.20 18:43, Alexander Duyck wrote:
This series provides an asynchronous means of reporting free guest pages
to a hypervisor so that the memory associated with those pages can be
dropped and reused by other processes and/or guests on the host. Using
this it is possible to avoid unnecessary I/O to disk and greatly improve
performance in the case of memory overcommit on the host.
When enabled we will be performing a scan of free memory every 2 seconds
while pages of sufficiently high order are being freed. In each pass at
least one sixteenth of each free list will be reported. By doing this we
avoid racing against other threads that may be causing a high amount of
memory churn.
The lowest page order currently scanned when reporting pages is
pageblock_order so that this feature will not interfere with the use of
Transparent Huge Pages in the case of virtualization.
Currently this is only in use by virtio-balloon however there is the hope
that at some point in the future other hypervisors might be able to make
use of it. In the virtio-balloon/QEMU implementation the hypervisor is
currently using MADV_DONTNEED to indicate to the host kernel that the page
is currently free. It will be zeroed and faulted back into the guest the
next time the page is accessed.
To track if a page is reported or not the Uptodate flag was repurposed and
used as a Reported flag for Buddy pages. We walk though the free list
isolating pages and adding them to the scatterlist until we either
encounter the end of the list, processed as many pages as were listed in
nr_free prior to us starting, or have filled the scatterlist with pages to
be reported. If we fill the scatterlist before we reach the end of the
list we rotate the list so that the first unreported page we encounter is
moved to the head of the list as that is where we will resume after we
have freed the reported pages back into the tail of the list.
Below are the results from various benchmarks. I primarily focused on two
tests. The first is the will-it-scale/page_fault2 test, and the other is
a modified version of will-it-scale/page_fault1 that was enabled to use
THP. I did this as it allows for better visibility into different parts
of the memory subsystem. The guest is running with 32G for RAM on one
node of a E5-2630 v3. The host has had some features such as CPU turbo
disabled in the BIOS.
Test page_fault1 (THP) page_fault2
Name tasks Process Iter STDEV Process Iter STDEV
Baseline 1 1012402.50 0.14% 361855.25 0.81%
16 8827457.25 0.09% 3282347.00 0.34%
Patches Applied 1 1007897.00 0.23% 361887.00 0.26%
16 8784741.75 0.39% 3240669.25 0.48%
Patches Enabled 1 1010227.50 0.39% 359749.25 0.56%
16 8756219.00 0.24% 3226608.75 0.97%
Patches Enabled 1 1050982.00 4.26% 357966.25 0.14%
page shuffle 16 8672601.25 0.49% 3223177.75 0.40%
Patches enabled 1 1003238.00 0.22% 360211.00 0.22%
shuffle w/ RFC 16 8767010.50 0.32% 3199874.00 0.71%
The results above are for a baseline with a linux-next-20191219 kernel,
that kernel with this patch set applied but page reporting disabled in
virtio-balloon, the patches applied and page reporting fully enabled, the
patches enabled with page shuffling enabled, and the patches applied with
page shuffling enabled and an RFC patch that makes used of MADV_FREE in
QEMU. These results include the deviation seen between the average value
reported here versus the high and/or low value. I observed that during the
test memory usage for the first three tests never dropped whereas with the
patches fully enabled the VM would drop to using only a few GB of the
host's memory when switching from memhog to page fault tests.
Any of the overhead visible with this patch set enabled seems due to page
faults caused by accessing the reported pages and the host zeroing the page
before giving it back to the guest. This overhead is much more visible when
using THP than with standard 4K pages. In addition page shuffling seemed to
increase the amount of faults generated due to an increase in memory churn.
The overhead is reduced when using MADV_FREE as we can avoid the extra
zeroing of the pages when they are reintroduced to the host, as can be seen
when the RFC is applied with shuffling enabled.
The overall guest size is kept fairly small to only a few GB while the test
is running. If the host memory were oversubscribed this patch set should
result in a performance improvement as swapping memory in the host can be
avoided.
I really like the approach overall. Voluntarily propagating free memory
from a guest to the host has been a sore point ever since KVM was
around. This solution looks like a very elegant way to do so.
The big piece I'm missing is the page cache. Linux will by default try
to keep the free list as small as it can in favor of page cache, so most
of the benefit of this patch set will be void in real world scenarios.
One approach is to move (parts of) the page cache from the guest to the
hypervisor - e.g., using emulated NVDIMM or virtio-pmem.
Whether you can do that depends heavily on your virtualization
environment. On a host with single tenant VMs, that's definitely
feasible. In a Kubernetes environment, it might also be feasible.
But when you have VMs that assume that the host is interfering with them
as little as possible, it becomes harder:
How do you ensure fairness across different VMs' page cache that is
munged into a single big host one?
Do you even have host page cache or are you using SR-IOV / mdev for
storage for performance reasons?
The puzzle is still incomplete, even with NVDIMM exposure to the guest
as an option unfortunately :).
Alex
Amazon Development Center Germany GmbH
Krausenstr. 38
10117 Berlin
Geschaeftsfuehrung: Christian Schlaeger, Jonathan Weiss
Eingetragen am Amtsgericht Charlottenburg unter HRB 149173 B
Sitz: Berlin
Ust-ID: DE 289 237 879
