BUG-8494: Cap queue sleep time 09/58209/1
authorRobert Varga <robert.varga@pantheon.tech>
Tue, 30 May 2017 09:41:58 +0000 (11:41 +0200)
committerRobert Varga <robert.varga@pantheon.tech>
Sun, 4 Jun 2017 07:47:42 +0000 (09:47 +0200)
Inconsistency in transmit queue backpressure has caused an observed
delay of over 44 minutes in CSIT. Thile this is an artefact of a bug,
we should not delay application thread for large periods of time, as
that can cause failures in the application (if it has other
time-bound duties like BGP keepalive messages and similar).

Restrict the sleep time to 5 seconds, emitting an INFO-level messages
when we reach this level of sleeping. This should not typically not
occur, as the backpressure should be kicking in much sooner, smooting
out the delay curve.

Change-Id: Ie5f148248caa71791bdda71ddd7e33e5733aa7f8
Signed-off-by: Robert Varga <robert.varga@pantheon.tech>
(cherry picked from commit 15a67bd103c2dc32f28139a7295ac84143c20d0c)

opendaylight/md-sal/cds-access-client/src/main/java/org/opendaylight/controller/cluster/access/client/AbstractClientConnection.java
opendaylight/md-sal/cds-access-client/src/main/java/org/opendaylight/controller/cluster/access/client/ClientActorBehavior.java

index 32becc0..4425c43 100644 (file)
@@ -72,6 +72,10 @@ public abstract class AbstractClientConnection<T extends BackendInfo> {
     // Emit a debug entry if we sleep for more that this amount
     private static final long DEBUG_DELAY_NANOS = TimeUnit.MILLISECONDS.toNanos(100);
 
+    // Upper bound on the time a thread is forced to sleep to keep queue size under control
+    private static final long MAX_DELAY_SECONDS = 5;
+    private static final long MAX_DELAY_NANOS = TimeUnit.SECONDS.toNanos(MAX_DELAY_SECONDS);
+
     private final Lock lock = new ReentrantLock();
     private final ClientActorContext context;
     @GuardedBy("lock")
@@ -133,10 +137,17 @@ public abstract class AbstractClientConnection<T extends BackendInfo> {
      */
     public final void sendRequest(final Request<?, ?> request, final Consumer<Response<?, ?>> callback) {
         final long now = currentTime();
-        final long delay = enqueueEntry(new ConnectionEntry(request, callback, now), now);
+        long delay = enqueueEntry(new ConnectionEntry(request, callback, now), now);
         try {
-            if (delay >= DEBUG_DELAY_NANOS && LOG.isDebugEnabled()) {
-                LOG.debug("Sleeping for {}ms", TimeUnit.NANOSECONDS.toMillis(delay));
+            if (delay >= DEBUG_DELAY_NANOS) {
+                if (delay > MAX_DELAY_NANOS) {
+                    LOG.info("Capping {} throttle delay from {} to {} seconds", this,
+                        TimeUnit.NANOSECONDS.toSeconds(delay), MAX_DELAY_SECONDS);
+                    delay = MAX_DELAY_NANOS;
+                }
+                if (LOG.isDebugEnabled()) {
+                    LOG.debug("Sleeping for {}ms", TimeUnit.NANOSECONDS.toMillis(delay));
+                }
             }
             TimeUnit.NANOSECONDS.sleep(delay);
         } catch (InterruptedException e) {
@@ -172,6 +183,20 @@ public abstract class AbstractClientConnection<T extends BackendInfo> {
     @GuardedBy("lock")
     final void finishReplay(final ReconnectForwarder forwarder) {
         setForwarder(forwarder);
+
+        /*
+         * The process of replaying all messages may have taken a significant chunk of time, depending on type
+         * of messages, queue depth and available processing power. In extreme situations this may have already
+         * exceeded BACKEND_ALIVE_TIMEOUT_NANOS, in which case we are running the risk of not making reasonable forward
+         * progress before we start a reconnect cycle.
+         *
+         * Note that the timer is armed after we have sent the first message, hence we should be seeing a response
+         * from the backend before we see a timeout, simply due to how the mailbox operates.
+         *
+         * At any rate, reset the timestamp once we complete reconnection (which an atomic transition from the
+         * perspective of outside world), as that makes it a bit easier to reason about timing of events.
+         */
+        lastReceivedTicks = currentTime();
         lock.unlock();
     }
 
index 9ba118e..554ffe9 100644 (file)
@@ -9,6 +9,7 @@ package org.opendaylight.controller.cluster.access.client;
 
 import com.google.common.annotations.Beta;
 import com.google.common.base.Preconditions;
+import com.google.common.base.Stopwatch;
 import com.google.common.base.Verify;
 import java.util.Collection;
 import java.util.Map;
@@ -278,6 +279,8 @@ public abstract class ClientActorBehavior<T extends BackendInfo> extends
         LOG.info("{}: resolved shard {} to {}", persistenceId(), shard, backend);
         final long stamp = connectionsLock.writeLock();
         try {
+            final Stopwatch sw = Stopwatch.createStarted();
+
             // Create a new connected connection
             final ConnectedClientConnection<T> newConn = new ConnectedClientConnection<>(conn.context(),
                     conn.cookie(), backend);
@@ -301,7 +304,7 @@ public abstract class ClientActorBehavior<T extends BackendInfo> extends
                 LOG.warn("{}: old connection {} does not match existing {}, new connection {} in limbo",
                     persistenceId(), conn, existing, newConn);
             } else {
-                LOG.info("{}: replaced connection {} with {}", persistenceId(), conn, newConn);
+                LOG.info("{}: replaced connection {} with {} in {}", persistenceId(), conn, newConn, sw);
             }
         } finally {
             connectionsLock.unlockWrite(stamp);

©2013 OpenDaylight, A Linux Foundation Collaborative Project. All Rights Reserved.
OpenDaylight is a registered trademark of The OpenDaylight Project, Inc.
Linux Foundation and OpenDaylight are registered trademarks of the Linux Foundation.
Linux is a registered trademark of Linus Torvalds.