Hi Paolo, Michael, Stefan and others,
Could anyone merge this patchset if it is okay to you guys?
On 03/30/2018 03:51 PM, guangrong.xiao@xxxxxxxxx wrote:
From: Xiao Guangrong <xiaoguangrong@xxxxxxxxxxx>
Changelog in v3:
Following changes are from Peter's review:
1) use comp_param[i].file and decomp_param[i].compbuf to indicate if
the thread is properly init'd or not
2) save the file which is used by ram loader to the global variable
instead it is cached per decompression thread
Changelog in v2:
Thanks to the review from Dave, Peter, Wei and Jiang Biao, the changes
in this version are:
1) include the performance number in the cover letter
2）add some comments to explain how to use z_stream->opaque in the
patchset
3) allocate a internal buffer for per thread to store the data to
be compressed
4) add a new patch that moves some code to ram_save_host_page() so
that 'goto' can be omitted gracefully
5) split the optimization of compression and decompress into two
separated patches
6) refine and correct code styles
This is the first part of our work to improve compression to make it
be more useful in the production.
The first patch resolves the problem that the migration thread spends
too much CPU resource to compression memory if it jumps to a new block
that causes the network is used very deficient.
The second patch fixes the performance issue that too many VM-exits
happen during live migration if compression is being used, it is caused
by huge memory returned to kernel frequently as the memory is allocated
and freed for every signal call to compress2()
The remaining patches clean the code up dramatically
Performance numbers:
We have tested it on my desktop, i7-4790 + 16G, by locally live migrate
the VM which has 8 vCPUs + 6G memory and the max-bandwidth is limited to
350. During the migration, a workload which has 8 threads repeatedly
written total 6G memory in the VM.
Before this patchset, its bandwidth is ~25 mbps, after applying, the
bandwidth is ~50 mbp.
We also collected the perf data for patch 2 and 3 on our production,
before the patchset:
+ 57.88% kqemu [kernel.kallsyms] [k] queued_spin_lock_slowpath
+ 10.55% kqemu [kernel.kallsyms] [k] __lock_acquire
+ 4.83% kqemu [kernel.kallsyms] [k] flush_tlb_func_common
- 1.16% kqemu [kernel.kallsyms] [k] lock_acquire ▒
- lock_acquire ▒
- 15.68% _raw_spin_lock ▒
+ 29.42% __schedule ▒
+ 29.14% perf_event_context_sched_out ▒
+ 23.60% tdp_page_fault ▒
+ 10.54% do_anonymous_page ▒
+ 2.07% kvm_mmu_notifier_invalidate_range_start ▒
+ 1.83% zap_pte_range ▒
+ 1.44% kvm_mmu_notifier_invalidate_range_end
apply our work:
+ 51.92% kqemu [kernel.kallsyms] [k] queued_spin_lock_slowpath
+ 14.82% kqemu [kernel.kallsyms] [k] __lock_acquire
+ 1.47% kqemu [kernel.kallsyms] [k] mark_lock.clone.0
+ 1.46% kqemu [kernel.kallsyms] [k] native_sched_clock
+ 1.31% kqemu [kernel.kallsyms] [k] lock_acquire
+ 1.24% kqemu libc-2.12.so [.] __memset_sse2
- 14.82% kqemu [kernel.kallsyms] [k] __lock_acquire ▒
- __lock_acquire ▒
- 99.75% lock_acquire ▒
- 18.38% _raw_spin_lock ▒
+ 39.62% tdp_page_fault ▒
+ 31.32% __schedule ▒
+ 27.53% perf_event_context_sched_out ▒
+ 0.58% hrtimer_interrupt
We can see the TLB flush and mmu-lock contention have gone.
Xiao Guangrong (10):
migration: stop compressing page in migration thread
migration: stop compression to allocate and free memory frequently
migration: stop decompression to allocate and free memory frequently
migration: detect compression and decompression errors
migration: introduce control_save_page()
migration: move some code to ram_save_host_page
migration: move calling control_save_page to the common place
migration: move calling save_zero_page to the common place
migration: introduce save_normal_page()
migration: remove ram_save_compressed_page()
migration/qemu-file.c | 43 ++++-
migration/qemu-file.h | 6 +-
migration/ram.c | 482 ++++++++++++++++++++++++++++++--------------------
3 files changed, 324 insertions(+), 207 deletions(-)
