Re: [Qemu-devel] [PATCH 1/8] migration: stop compressing page in migration thread

[Dr. David Alan Gilbert]

* Xiao Guangrong (address@hidden) wrote:
> 
> 
> On 03/26/2018 05:02 PM, Peter Xu wrote:
> > On Thu, Mar 22, 2018 at 07:38:07PM +0800, Xiao Guangrong wrote:
> > > 
> > > 
> > > On 03/21/2018 04:19 PM, Peter Xu wrote:
> > > > On Fri, Mar 16, 2018 at 04:05:14PM +0800, Xiao Guangrong wrote:
> > > > > 
> > > > > Hi David,
> > > > > 
> > > > > Thanks for your review.
> > > > > 
> > > > > On 03/15/2018 06:25 PM, Dr. David Alan Gilbert wrote:
> > > > > 
> > > > > > >     migration/ram.c | 32 ++++++++++++++++----------------
> > > > > > 
> > > > > > Hi,
> > > > > >      Do you have some performance numbers to show this helps?  Were 
> > > > > > those
> > > > > > taken on a normal system or were they taken with one of the 
> > > > > > compression
> > > > > > accelerators (which I think the compression migration was designed 
> > > > > > for)?
> > > > > 
> > > > > Yes, i 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 mbps.
> > > > 
> > > > Hi, Guangrong,
> > > > 
> > > > Not really review comments, but I got some questions. :)
> > > 
> > > Your comments are always valuable to me! :)
> > > 
> > > > 
> > > > IIUC this patch will only change the behavior when last_sent_block
> > > > changed.  I see that the performance is doubled after the change,
> > > > which is really promising.  However I don't fully understand why it
> > > > brings such a big difference considering that IMHO current code is
> > > > sending dirty pages per-RAMBlock.  I mean, IMHO last_sent_block should
> > > > not change frequently?  Or am I wrong?
> > > 
> > > It's depends on the configuration, each memory-region which is ram or
> > > file backend has a RAMBlock.
> > > 
> > > Actually, more benefits comes from the fact that the performance & 
> > > throughput
> > > of the multithreads has been improved as the threads is fed by the
> > > migration thread and the result is consumed by the migration
> > > thread.
> > 
> > I'm not sure whether I got your points - I think you mean that the
> > compression threads and the migration thread can form a better
> > pipeline if the migration thread does not do any compression at all.
> > 
> > I think I agree with that.
> > 
> > However it does not really explain to me on why a very rare event
> > (sending the first page of a RAMBlock, considering bitmap sync is
> > rare) can greatly affect the performance (it shows a doubled boost).
> > 
> 
> I understand it is trick indeed, but it is not very hard to explain.
> Multi-threads (using 8 CPUs in our test) keep idle for a long time
> for the origin code, however, after our patch, as the normal is
> posted out async-ly that it's extremely fast as you said (the network
> is almost idle for current implementation) so it has a long time that
> the CPUs can be used effectively to generate more compressed data than
> before.

One thing to try, to explain Peter's worry, would be, for testing, to
add a counter to see how often this case triggers, and perhaps add
some debug to see when;  Peter's right that flipping between the
RAMBlocks seems odd, unless you're either doing lots of iterations or
have lots of separate RAMBlocks for some reason.

Dave

> > Btw, about the numbers: IMHO the numbers might not be really "true
> > numbers".  Or say, even the bandwidth is doubled, IMHO it does not
> > mean the performance is doubled. Becasue the data has changed.
> > 
> > Previously there were only compressed pages, and now for each cycle of
> > RAMBlock looping we'll send a normal page (then we'll get more thing
> > to send).  So IMHO we don't really know whether we sent more pages
> > with this patch, we can only know we sent more bytes (e.g., an extreme
> > case is that the extra 25Mbps/s are all caused by those normal pages,
> > and we can be sending exactly the same number of pages like before, or
> > even worse?).
> > 
> 
> Current implementation uses CPU very ineffectively (it's our next work
> to be posted out) that the network is almost idle so posting more data
> out is a better choice，further more, migration thread plays a role for
> parallel, it'd better to make it fast.
> 
> > > 
> > > > 
> > > > Another follow-up question would be: have you measured how long time
> > > > needed to compress a 4k page, and how many time to send it?  I think
> > > > "sending the page" is not really meaningful considering that we just
> > > > put a page into the buffer (which should be extremely fast since we
> > > > don't really flush it every time), however I would be curious on how
> > > > slow would compressing a page be.
> > > 
> > > I haven't benchmark the performance of zlib, i think it is CPU intensive
> > > workload, particularly, there no compression-accelerator (e.g, QAT) on
> > > our production. BTW, we were using lzo instead of zlib which worked
> > > better for some workload.
> > 
> > Never mind. Good to know about that.
> > 
> > > 
> > > Putting a page into buffer should depend on the network, i,e, if the
> > > network is congested it should take long time. :)
> > 
> > Again, considering that I don't know much on compression (especially I
> > hardly used that) mine are only questions, which should not block your
> > patches to be either queued/merged/reposted when proper. :)
> 
> Yes, i see. The discussion can potentially raise a better solution.
> 
> Thanks for your comment, Peter!
--
Dr. David Alan Gilbert / address@hidden / Manchester, UK