[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index]
Re: [Qemu-devel] [PATCH V2 0/3] Introduce COLO-compare
Dr. David Alan Gilbert
Re: [Qemu-devel] [PATCH V2 0/3] Introduce COLO-compare
Thu, 31 Mar 2016 10:43:06 +0100
* Li Zhijian (address@hidden) wrote:
> On 03/30/2016 08:05 PM, Dr. David Alan Gilbert wrote:
> >* Zhang Chen (address@hidden) wrote:
> >>COLO-compare is a part of COLO project. It is used
> >>to compare the network package to help COLO decide
> >>whether to do checkpoint.
> >Hi Zhang Chen,
> > I've put comments on the individual patches, but some more general things:
> > 1) Please add a coment giving the example of the command line for the
> > primary
> > and secondary use of this module - it helps make it easier to
> > understand the patches.
> > 2) There's no tracing in here - please add some; I found when I tried to
> > get
> > COLO working I needed to use lots of tracing and debugging to
> > understand the
> > packet flow.
> > 3) Add comments; e.g. for each function say which thread is using it and
> > where
> > the packets are coming from; e.g.
> > called from the main thread on the primary for packets arriving
> > over the socket
> > from the secondary.
> > There's just so many packets going in so many directions it would make
> > it
> > easier to follow.
> > 4) A more fundamental problem is what happens if the secondary never
> > sends anything
> > on the socket, the result is you end up running until the end of the
> > long COLO
> > checkpoint without triggering a discompare - in my world I added a
> > timeout (400ms)
> > for an unmatched packet from the primary, where if no matching packet
> > was received
> > a checkpoint would be triggered.
> > 5) I see the packet comparison is still the simple memcmpy that you had
> > in December;
> > are you planning on doing anything more complicated; you must be seing
> > most packets
> > miscompare?
> >You can see my current world at;
> >which has my basic TCP comparison (it's only tracking incoming connections)
> >and I know it's
> >not complete either. It mostly works OK, although I've got an occasional seg
> >(which makes me wonder if I need to add the conn_list_lock I see you added).
> > I'm also
> >not doing any TCP reassembly which is probably needed.
> Thank you very much for your comments.
> I just see you tree, you put in a lot of work(tcp comparison enhance,
> number re-write, timeout...)
> Actually, this compare module is just in a RFC stage(only including compare
> frame), there are
> many works to be done:
> 1) Integrate to COLO frame(and Let COLO primary and secondary at running
Yes; although I think you've had most of the code for that, also feel free to
any of the bits I've changed.
> 2) ip segment defrag
Yes, this seems the hardest bit to me. It is needed for some workloads; for
a simple case if just an ssh connection, the differences in timing between the
and secondary you see that the primary might send two short packets and the
might send one long packet with the same data.
> 3) comparison base on the sequence number(tcp and udp) if packet has
> Because tcp re-transmission is quit common. IRC, your code will compare
> the whole tcp
> packet(sequence number will be compare)
Yes; once the secondary reworks the sequence number I found the seuence number
matched most of the time on the primary. One problem (depending on the
is that the ack number might not match and I'm not sure of the best fix, for
a) Send packet 1
b) Send packet 2
c) Receive response
The order of b & c is random - if the response on the same socket arrives (c)
then the ack number in packet 2 is different; on one workload this caused a lot
but only if the timing is just wrong.
(I also had to turn off TCP timestamping to get useful comparisons)
> 4) packet belongs to the same connection is sort by sequence number
> 5) Out-Of-Oder packet handle
I think 4 & 5 both happen as part of the defrag; out of order packets were a
me on some of the workloads; I had to turn off multiqueue in one case.
> 6) cleanup the un-active conn_list which maybe closed. the simple way is to
> introduce a
> timer to record whether a connection have packet come within a timeout,
> connection gone
> beyond this timeout should be cleanup.
At the moment that isn't a big problem because when you receive the next
can flush the list.
The only case where you have to deal with this is for continuous failover when
original secondary is promoted to primary, then it's connection list has to live
on longer for any connections it created prior to the failover.
Perhaps this gets more complex with defrag?
> 7) Dave point out above (4)
> 8) something I miss...
> For Various reasons, not all the works can be done immediately, So we hope to
> discuss and
> decide which function have the high priority.
> Any comments and suggestions are welcome.
Yes, there's a lot of work; as I say, feel free to use any of my patches
from the world above, I wasn't planning on doing much more work on that set.
> IMO, a compare frame and a COLO frame hack patch could be simple enough.
I think you'd have to show that you got some useful comparison matches;
if it almost always failed the comparison then I can't see the point.
> >> - add jhash.h
> >> - initial patch
> >>Zhang Chen (3):
> >> colo-compare: introduce colo compare initlization
> >> colo-compare: track connection and enqueue packet
> >> colo-compare: introduce packet comparison thread
> >> include/qemu/jhash.h | 59 ++++
> >> net/Makefile.objs | 1 +
> >> net/colo-compare.c | 782
> >> +++++++++++++++++++++++++++++++++++++++++++++++++++
> >> vl.c | 3 +-
> >> 4 files changed, 844 insertions(+), 1 deletion(-)
> >> create mode 100644 include/qemu/jhash.h
> >> create mode 100644 net/colo-compare.c
> >Dr. David Alan Gilbert / address@hidden / Manchester, UK
> Best regards.
> Li Zhijian (8555)
Dr. David Alan Gilbert / address@hidden / Manchester, UK
[Qemu-devel] [PATCH V2 3/3] colo-compare: introduce packet comparison thread, Zhang Chen, 2016/03/30
Re: [Qemu-devel] [PATCH V2 0/3] Introduce COLO-compare, Dr. David Alan Gilbert, 2016/03/30
Re: [Qemu-devel] [PATCH V2 0/3] Introduce COLO-compare, Zhang Chen, 2016/03/31
- Re: [Qemu-devel] [PATCH V2 2/3] colo-compare: track connection and enqueue packet, (continued)