1. Description of multiple page faults for received zero pages
a. -mem-prealloc feature and hugepage backend are not enabled on the
destination
b. After receiving the zero pages, the destination first determines if
the current page content is 0 via buffer_is_zero, this may cause a
read page fault
perf record -e page-faults information below
13.75% 13.75% multifdrecv_0 qemu-system-x86_64 [.] buffer_zero_avx512
11.85% 11.85% multifdrecv_1 qemu-system-x86_64 [.] buffer_zero_avx512
multifd_recv_thread
nocomp_recv
multifd_recv_zero_page_process
buffer_is_zero
select_accel_fn
buffer_zero_avx512
c. Other page faults mainly come from writing operations to normal and
zero pages.
2. Solution
a. During the multifd migration process, the received pages are tracked
through RAMBlock's receivedmap.
b. If received zero page is not set in recvbitmap, the destination will not
check whether the page content is 0, thus avoiding the occurrence of
read fault.
c. If the zero page has been set in receivedmap, set the page with 0
directly.
There are two reasons for this
1. It's unlikely a zero page if it's sent once or more.
2. For the 1st time destination received a zero page, it must be a zero
page, so no need to scan for the 1st round.
3. Test Result 16 vCPUs and 64G memory VM, multifd number is 2,
and 100G network bandwidth
3.1 Test case: 16 vCPUs are idle and only 2G memory are used
+-----------+--------+--------+----------+
|MultiFD | total |downtime| Page |
|Nocomp | time | | Faults |
| | (ms) | (ms) | |
+-----------+--------+--------+----------+
|with | | | |
|recvbitmap | 7335| 180| 2716|
+-----------+--------+--------+----------+
|without | | | |
|recvbitmap | 7771| 153| 121357|
+-----------+--------+--------+----------+
+-----------+--------+--------+--------+-------+--------+-------------+
|MultiFD | total |downtime| SVM |SVM | IOTLB | IO PageFault|
|QPL | time | | IO TLB |IO Page| MaxTime| MaxTime |
| | (ms) | (ms) | Flush |Faults | (us) | (us) |
+-----------+--------+--------+--------+-------+--------+-------------+
|with | | | | | | |
|recvbitmap | 10224| 175| 410| 27429| 1| 447|
+-----------+--------+--------+--------+-------+--------+-------------+
|without | | | | | | |
|recvbitmap | 11253| 153| 80756| 38655| 25| 18349|
+-----------+--------+--------+--------+-------+--------+-------------+
3.2 Test case: 16 vCPUs are idle and 56G memory(not zero) are used
+-----------+--------+--------+----------+
|MultiFD | total |downtime| Page |
|Nocomp | time | | Faults |
| | (ms) | (ms) | |
+-----------+--------+--------+----------+
|with | | | |
|recvbitmap | 16825| 165| 52967|
+-----------+--------+--------+----------+
|without | | | |
|recvbitmap | 12987| 159| 2672677|
+-----------+--------+--------+----------+
+-----------+--------+--------+--------+-------+--------+-------------+
|MultiFD | total |downtime| SVM |SVM | IOTLB | IO PageFault|
|QPL | time | | IO TLB |IO Page| MaxTime| MaxTime |
| | (ms) | (ms) | Flush |Faults | (us) | (us) |
+-----------+--------+--------+--------+-------+--------+-------------+
|with | | | | | | |
|recvbitmap | 132315| 77| 890| 937105| 60| 9581|
+-----------+--------+--------+--------+-------+--------+-------------+
|without | | | | | | |
|recvbitmap | >138333| N/A| 1647701| 981899| 43| 21018|
+-----------+--------+--------+--------+-------+--------+-------------+
>From the test result, both of page faults and IOTLB Flush operations can
be significantly reduced. The reason is that zero page processing does not
trigger read faults, and a large number of zero pages do not even trigger
write faults (Test 3.1), because it is considered that after the destination
is started, the content of unaccessed pages is 0.
I have a concern here, the RAM memory is allocated by mmap with anonymous
flag, and if the first received zero page is not set to 0 explicitly, does
this ensure that the received zero pages memory data is 0?
In this case, the performance impact of live migration is not big
because the destination is not the bottleneck.
When using QPL (SVM-capable device), even if IOTLB is improved, the
overall performance will still be seriously degraded because a large
number of IO page faults are still generated.
Previous discussion link:
1.
https://lore.kernel.org/all/CAAYibXib+TWnJpV22E=adncdBmwXJRqgRjJXK7X71J=bdfa...@mail.gmail.com/
2.
https://lore.kernel.org/all/ph7pr11mb594123f7eefebfce219af100a3...@ph7pr11mb5941.namprd11.prod.outlook.com/
Yuan Liu (1):
migration/multifd: solve zero page causing multiple page faults
migration/multifd-zero-page.c | 4 +++-
migration/multifd-zlib.c | 1 +
migration/multifd-zstd.c | 1 +
migration/multifd.c | 1 +
migration/ram.c | 4 ++++
migration/ram.h | 1 +
6 files changed, 11 insertions(+), 1 deletion(-)
--
2.39.3
