RE: [PATCH v1 0/1] COLO: migrate dirty ram pages before colo checkpoint

[Zhanghailiang <zhang.zhanghailiang@huawei.com>]

> -----Original Message-----
> From: Derek Su [mailto:jwsu1986@gmail.com]
> Sent: Thursday, August 13, 2020 6:28 PM
> To: Lukas Straub <lukasstraub2@web.de>
> Cc: Derek Su <dereksu@qnap.com>; qemu-devel@nongnu.org; Zhanghailiang
> <zhang.zhanghailiang@huawei.com>; chyang@qnap.com; quintela@redhat.com;
> dgilbert@redhat.com; ctcheng@qnap.com
> Subject: Re: [PATCH v1 0/1] COLO: migrate dirty ram pages before colo
> checkpoint
> 
> On Fri, Jul 31, 2020 at 3:52 PM Lukas Straub <lukasstraub2@web.de> wrote:
> >
> > On Sun, 21 Jun 2020 10:10:03 +0800
> > Derek Su <dereksu@qnap.com> wrote:
> >
> > > This series is to reduce the guest's downtime during colo checkpoint
> > > by migrating dirty ram pages as many as possible before colo checkpoint.
> > >
> > > If the iteration count reaches COLO_RAM_MIGRATE_ITERATION_MAX or ram
> > > pending size is lower than 'x-colo-migrate-ram-threshold', stop the
> > > ram migration and do colo checkpoint.
> > >
> > > Test environment:
> > > The both primary VM and secondary VM has 1GiB ram and 10GbE NIC for
> > > FT traffic.
> > > One fio buffer write job runs on the guest.
> > > The result shows the total primary VM downtime is decreased by ~40%.
> > >
> > > Please help to review it and suggestions are welcomed.
> > > Thanks.
> >
> > Hello Derek,
> > Sorry for the late reply.
> > I think this is not a good idea, because it unnecessarily introduces a delay
> between checkpoint request and the checkpoint itself and thus impairs network
> bound workloads due to increased network latency. Workloads that are
> independent from network don't cause many checkpoints anyway, so it doesn't
> help there either.
> >
> 

Hi Derek,

Actually, There is a quit interesting question we should think: 
What will happen if VM continues to run after detected a mismatched state between PVM and SVM,
According to the rules of COLO, we should stop VMs immediately to sync the state between PVM and SVM,
But here, you choose them to continue to run for a while, then there may be more client's network packages
Coming, and may cause more memory pages dirty, another side effect is the new network packages will not
Be sent out with high probability, because their replies should be different since the state between PVM and SVM is different.

So, IMHO, it makes non-sense to let VMs to continue to run after detected them in different state.
Besides, I don't think it is easy to construct this case in tests.


Thanks,
Hailiang

s> Hello, Lukas & Zhanghailiang
> 
> Thanks for your opinions.
> I went through my patch, and I feel a little confused and would like to dig into it
> more.
> 
> In this patch, colo_migrate_ram_before_checkpoint() is before
> COLO_MESSAGE_CHECKPOINT_REQUEST, so the SVM and PVM should not enter
> the pause state.
> 
> In the meanwhile, the packets to PVM/SVM can still be compared and notify
> inconsistency if mismatched, right?
> Is it possible to introduce extra network latency?
> 
> In my test (randwrite to disk by fio with direct=0), the ping from another client to
> the PVM  using generic colo and colo used this patch are below.
> The network latency does not increase as my expectation.
> 
> generic colo
> ```
> 64 bytes from 192.168.80.18: icmp_seq=87 ttl=64 time=28.109 ms
> 64 bytes from 192.168.80.18: icmp_seq=88 ttl=64 time=16.747 ms
> 64 bytes from 192.168.80.18: icmp_seq=89 ttl=64 time=2388.779 ms
> <----checkpoint start
> 64 bytes from 192.168.80.18: icmp_seq=90 ttl=64 time=1385.792 ms
> 64 bytes from 192.168.80.18: icmp_seq=91 ttl=64 time=384.896 ms
> <----checkpoint end
> 64 bytes from 192.168.80.18: icmp_seq=92 ttl=64 time=3.895 ms
> 64 bytes from 192.168.80.18: icmp_seq=93 ttl=64 time=1.020 ms
> 64 bytes from 192.168.80.18: icmp_seq=94 ttl=64 time=0.865 ms
> 64 bytes from 192.168.80.18: icmp_seq=95 ttl=64 time=0.854 ms
> 64 bytes from 192.168.80.18: icmp_seq=96 ttl=64 time=28.359 ms
> 64 bytes from 192.168.80.18: icmp_seq=97 ttl=64 time=12.309 ms
> 64 bytes from 192.168.80.18: icmp_seq=98 ttl=64 time=0.870 ms
> 64 bytes from 192.168.80.18: icmp_seq=99 ttl=64 time=2371.733 ms
> 64 bytes from 192.168.80.18: icmp_seq=100 ttl=64 time=1371.440 ms
> 64 bytes from 192.168.80.18: icmp_seq=101 ttl=64 time=366.414 ms
> 64 bytes from 192.168.80.18: icmp_seq=102 ttl=64 time=0.818 ms
> 64 bytes from 192.168.80.18: icmp_seq=103 ttl=64 time=0.997 ms ```
> 
> colo used this patch
> ```
> 64 bytes from 192.168.80.18: icmp_seq=72 ttl=64 time=1.417 ms
> 64 bytes from 192.168.80.18: icmp_seq=73 ttl=64 time=0.931 ms
> 64 bytes from 192.168.80.18: icmp_seq=74 ttl=64 time=0.876 ms
> 64 bytes from 192.168.80.18: icmp_seq=75 ttl=64 time=1184.034 ms
> <----checkpoint start
> 64 bytes from 192.168.80.18: icmp_seq=76 ttl=64 time=181.297 ms
> <----checkpoint end
> 64 bytes from 192.168.80.18: icmp_seq=77 ttl=64 time=0.865 ms
> 64 bytes from 192.168.80.18: icmp_seq=78 ttl=64 time=0.858 ms
> 64 bytes from 192.168.80.18: icmp_seq=79 ttl=64 time=1.247 ms
> 64 bytes from 192.168.80.18: icmp_seq=80 ttl=64 time=0.946 ms
> 64 bytes from 192.168.80.18: icmp_seq=81 ttl=64 time=0.855 ms
> 64 bytes from 192.168.80.18: icmp_seq=82 ttl=64 time=0.868 ms
> 64 bytes from 192.168.80.18: icmp_seq=83 ttl=64 time=0.749 ms
> 64 bytes from 192.168.80.18: icmp_seq=84 ttl=64 time=2.154 ms
> 64 bytes from 192.168.80.18: icmp_seq=85 ttl=64 time=1499.186 ms
> 64 bytes from 192.168.80.18: icmp_seq=86 ttl=64 time=496.173 ms
> 64 bytes from 192.168.80.18: icmp_seq=87 ttl=64 time=0.854 ms
> 64 bytes from 192.168.80.18: icmp_seq=88 ttl=64 time=0.774 ms ```
> 
> Thank you.
> 
> Regards,
> Derek
> 
> > Hailang did have a patch to migrate ram between checkpoints, which should
> help all workloads, but it wasn't merged back then. I think you can pick it up again,
> rebase and address David's and Eric's comments:
> > https://lore.kernel.org/qemu-devel/20200217012049.22988-3-zhang.zhangh
> > ailiang@huawei.com/T/#u
> >
> > Hailang, are you ok with that?
> >
> > Regards,
> > Lukas Straub