qemu-devel
[Top][All Lists]
Advanced

[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index]

Re: [Qemu-devel] [PATCH V2 1/4] net/colo-compare.c: Add checkpoint min p


From: Zhang Chen
Subject: Re: [Qemu-devel] [PATCH V2 1/4] net/colo-compare.c: Add checkpoint min period to optimize performance
Date: Tue, 18 Jul 2017 10:20:13 +0800
User-agent: Mozilla/5.0 (X11; Linux x86_64; rv:52.0) Gecko/20100101 Thunderbird/52.1.1



On 07/17/2017 08:24 PM, Dr. David Alan Gilbert wrote:
* Zhang Chen (address@hidden) wrote:

On 07/14/2017 08:10 PM, Dr. David Alan Gilbert wrote:
* Zhang Chen (address@hidden) wrote:
If colo-compare find out the first different packet that means
the following packet almost is different. we needn't do a lot
of checkpoint in this time, so we set the no-need-checkpoint
peroid, default just set 3 second.

Signed-off-by: Zhang Chen <address@hidden>
---
   net/colo-compare.c | 13 ++++++++++++-
   1 file changed, 12 insertions(+), 1 deletion(-)

diff --git a/net/colo-compare.c b/net/colo-compare.c
index 6d500e1..0f8e198 100644
--- a/net/colo-compare.c
+++ b/net/colo-compare.c
@@ -40,6 +40,9 @@
   /* TODO: Should be configurable */
   #define REGULAR_PACKET_CHECK_MS 3000
+/* TODO: Should be configurable */
Yes it should!

+#define CHECKPOINT_MIN_TIME 3000
+
   /*
     + CompareState ++
     |               |
@@ -455,6 +458,7 @@ static void colo_compare_connection(void *opaque, void 
*user_data)
       Packet *pkt = NULL;
       GList *result = NULL;
       int ret;
+    static int64_t checkpoint_time_ms;
       while (!g_queue_is_empty(&conn->primary_list) &&
              !g_queue_is_empty(&conn->secondary_list)) {
@@ -494,7 +498,14 @@ static void colo_compare_connection(void *opaque, void 
*user_data)
                */
               trace_colo_compare_main("packet different");
               g_queue_push_tail(&conn->primary_list, pkt);
-            /* TODO: colo_notify_checkpoint();*/
+
+            if (pkt->creation_ms - checkpoint_time_ms > CHECKPOINT_MIN_TIME) {
+                /*
+                 * TODO: Notify colo frame to do checkpoint.
+                 * colo_compare_inconsistent_notify();
+                 */
+                checkpoint_time_ms = pkt->creation_ms;
+            }
You need to be careful how this interacts with the actual start of the
checkpoint.   Lets say you have two miscompared packets close to each
other:


      miscompare!
           checkpoint
      miscompare!
           ignore it because it was close to the 1st one

     That means we never trigger the 2nd checkpoint and it'll carry on
until the maximum checkpoint length.

     But also, I think you need to consider what happens to future packets
being compared; you can't release any packets now until the checkpoint
as soon as you know there's a miscompare.
We need some time to do the checkpoint, and in this period we can ignore
the miscompare to get better performance. Like that:

currently:

     miscompare!
          notify checkpoint
     miscompare!
          notify checkpoint
     miscompare!
          notify checkpoint
     miscompare!
          notify checkpoint
     vm_stop and do checkpoint

     vm_start and finish checkpoint

     vm_stop and do checkpoint

     vm_start and finish checkpoint

     vm_stop and do checkpoint

     vm_start and finish checkpoint

     vm_stop and do checkpoint

     vm_start and finish checkpoint


running normally.


after:

     miscompare!
          notify checkpoint
     miscompare!
          ignore
     miscompare!
          ignore
     miscompare!
          ignore
     vm_stop and do checkpoint

     vm_start and finish checkpoint

running normally.
Yes, but you must make sure that you don't
ignore any miscompares after the start of the next checkpoint - I don't
see how you avoid that.

Good catch, I will fix it in next version.


Also we must be careful about packets released after the 1st miscompare.

Yes, after the 1st miscompare, all ignored packet will be enqueued.
Then, we will flush all packet in the queue during do checkpoint.

Thanks
Zhang Chen


Dave


Thanks
Zhang Chen


Dave

               break;
           }
       }
--
2.7.4




--
Dr. David Alan Gilbert / address@hidden / Manchester, UK


.

--
Thanks
Zhang Chen



--
Dr. David Alan Gilbert / address@hidden / Manchester, UK


.


--
Thanks
Zhang Chen






reply via email to

[Prev in Thread] Current Thread [Next in Thread]