|
From: | Hyman Huang |
Subject: | Re: [PATCH] softmmu/dirtylimit: Fix usleep early return on signal |
Date: | Mon, 16 Oct 2023 23:37:01 +0800 |
User-agent: | Mozilla/5.0 (Windows NT 10.0; Win64; x64; rv:102.0) Gecko/20100101 Thunderbird/102.15.1 |
在 2023/9/5 10:17, alloc young 写道:
On 2023/9/4 21:27, Yong Huang wrote:Use the command line below with initrd built in tests/migration, guest runs stress at 5GiB/s. Migration with dirty limit on and default parameters, migration can't finish within 1 hour. The default vcpu dirty limit set is 1 MB/s, however, the mig_src.log show copy rate at 128MiB/s. With this patch, migration finsih in 39s.On Fri, Sep 1, 2023 at 10:19 AM <alloc.young@outlook.com <mailto:alloc.young@outlook.com>> wrote:From: alloc <alloc.young@outlook.com <mailto:alloc.young@outlook.com>>Timeout functions like usleep can return early on signal, which reducesmore dirty pages than expected. In dirtylimit case, dirtyrate meter thread needs to kick all vcpus out to sync. The callchain: vcpu_calculate_dirtyrate global_dirty_log_sync memory_global_dirty_log_sync kvm_log_sync_global kvm_dirty_ring_flushkvm_cpu_synchronize_kick_all <---- send vcpu signalFor long time sleep, use qemu_cond_timedwait_iothread to handle cpu stopevent. The Dirty Limit algorithm seeks to keep the vCPU dirty page rate within the set limit; since it focuses more emphasis on processing time and precision, I feel that improvement should strive for the same result. Could you please provide the final test results showing the impact of that improvement?/usr/libexec/qemu-kvm -display none -vga none -name mig_src,debug-threads=on -monitor stdio -accel kvm,dirty-ring-size=4096 -cpu host -kernel /boot/vmlinuz-5.14.0-70.22.1.el9_0.x86_64 -initrd /root/initrd-stress.img -append noapic edd=off printk.time=1 noreplace-smp cgroup_disable=memory pci=noearly console=ttyS0 debug ramsize=1 ncpus=1 -chardev file,id=charserial0,path=/var/log/mig_src.log -serial chardev:charserial0 -m 1536 -smp 1
I calculated the migration time and success probability using the QEMU test/migration/guestperf.py script. With the patch, performance seems to not increase. Would you kindly post the test case's specifics? IMHO, conducting tests more than once and obtaining probability statistics data could be more compelling. Thanks, Yong
Signed-off-by: alloc <alloc.young@outlook.com <mailto:alloc.young@outlook.com>> --- softmmu/dirtylimit.c | 19 +++++++++++++++++-- 1 file changed, 17 insertions(+), 2 deletions(-) diff --git a/softmmu/dirtylimit.c b/softmmu/dirtylimit.c index fa959d7743..ee938c636d 100644 --- a/softmmu/dirtylimit.c +++ b/softmmu/dirtylimit.c @@ -411,13 +411,28 @@ void dirtylimit_set_all(uint64_t quota, void dirtylimit_vcpu_execute(CPUState *cpu) { + int64_t sleep_us, endtime_us; + + dirtylimit_state_lock(); if (dirtylimit_in_service() && dirtylimit_vcpu_get_state(cpu->cpu_index)->enabled && cpu->throttle_us_per_full) { trace_dirtylimit_vcpu_execute(cpu->cpu_index, cpu->throttle_us_per_full); - usleep(cpu->throttle_us_per_full); - } + sleep_us = cpu->throttle_us_per_full; + dirtylimit_state_unlock();+ endtime_us = qemu_clock_get_us(QEMU_CLOCK_REALTIME) + sleep_us;+ while (sleep_us > 0 && !cpu->stop) { + if (sleep_us > SCALE_US) { + qemu_mutex_lock_iothread(); + qemu_cond_timedwait_iothread(cpu->halt_cond, sleep_us / SCALE_US); + qemu_mutex_unlock_iothread(); + } else + g_usleep(sleep_us); + sleep_us = endtime_us - qemu_clock_get_us(QEMU_CLOCK_REALTIME); + } + } else + dirtylimit_state_unlock(); } static void dirtylimit_init(void) -- 2.39.3 -- Best regards
[Prev in Thread] | Current Thread | [Next in Thread] |