Mailing-List: contact dev-help@activemq.apache.org; run by ezmlm
Precedence: bulk
Reply-To: dev@activemq.apache.org
Date: Thu, 19 May 2011 18:10:49 +0000 (UTC)
From: "Stirling Chow (JIRA)" <jira@apache.org>
To: dev@activemq.apache.org
Message-ID: 
 <869902131.27502.1305828649422.JavaMail.tomcat@hel.zones.apache.org>
In-Reply-To: 
 <646013700.27280.1305825887318.JavaMail.tomcat@hel.zones.apache.org>
Subject: [jira] [Updated] (AMQ-3331) When a producer from a network bridge
 is blocked by producer flow control, all producers from the network bridge
 get blocked.
MIME-Version: 1.0
Content-Type: text/plain; charset=utf-8
Content-Transfer-Encoding: 7bit


     [ https://issues.apache.org/jira/browse/AMQ-3331?page=com.atlassian.jira.plugin.system.issuetabpanels:all-tabpanel ]

Stirling Chow updated AMQ-3331:
-------------------------------

    Description: 
Symptom
=======
Broker A produces messages to two queues, Q1 and Q2.  Broker B consumes messages from two queues, Q1 and Q2.  Broker A is connected by a demand forwarding bridge, over TCP, to Broker B so that messages produced to Q1/Q2 will be forwarded to the consumers on Broker B.

At some point, Broker B's instance of Q2 becomes full (e.g., because the Q2 consumer is slow), and this triggers producer flow control to halt new messages being sent to Broker B's Q2 over the bridge.  Broker A's instances of Q1/Q2 are not full, so the producers on Broker A are not blocked.

If the messages produced by Broker A are *persistent*, we see this behaviour over the course of the production of 1000 messages to both Q1/Q2, where Broker B's Q2 becomes full on the 500th message:

{noformat}
Broker A       Bridge        Broker B
========                     ========
0->1000->0     ------>       0->1000->...
0->1000->500                 0->500->...
{noformat}

The above results, which assume network and consumer prefetch sizes of 1, are what we expected, namely: 
# Broker A produces 1000 messages to Q1 without blocking and all of these messages are forwarded to Broker B's Q1 without blocking, eventually being consumed by Broker B's Q1 consumer.
# Broker A produces 1000 messages to Q2 without blocking and 500 of these messages are forwarded to Broker B's Q2 before producer flow control blocks the flow until Broker B's Q2 consumer can start reducing the queue size.


This is good because the bridge treats Q1 and Q2 independently (i.e., producer flow control on Q2 does not block the messages forwarded to Q1).

If the messages produced by Broker A are *non-persistent*, we see this behaviour:

{noformat}
Broker A       Bridge        Broker B
========                     ========
0->1000->500   ------>       0->500->...
0->1000->500                 0->500->...
{noformat}

The above results, which assume network and consumer prefetch sizes of 1, are not what we expected, namely: producer flow control on Broker B's instance of Q2 blocks the forwarding of messages to Broker B's instance of Q1.

This is not good because producer flow control on Q2 essentially triggers producer flow control on Q1, even though Q1 is *not* full.

It also seems strange (and almost non-intuitive until you understand the cause), that peristent messages should behave better than non-persistent messages.  The same difference in behaviour can also be observed with persistent messages if Broker A these outside a JMS transaction (e.g., AUTO_ACKNOWLEDGE) versus inside a JMS transaction: outside behaves appropriately, with Q1 independent of Q2, but inside behaves the same as the non-persistent case with Q1 blocked by Q2.

These observations are contrary to the AMQ 5.0 documentation regarding producer flow control: {quote}As of ActiveMQ 5.0, we can now individually flow control each producer on a shared connection without having to suspend the entire connection.{quote}

Cause
=====

The difference in behaviour between persistent and non-persistent (and transactionaly/non-transactional) is due to the two ways that org.apache.activemq.broker.region.Queue implements producer flow control:

{code:title=Queue#send(...)}
// We can avoid blocking due to low usage if the producer is
// sending
// a sync message or if it is using a producer window
if (producerInfo.getWindowSize() > 0 || message.isResponseRequired()) {
{code}

and

{code:title=Queue#send(...)}
} else 
  if (memoryUsage.isFull()) {
    waitForSpace(context, memoryUsage, "Usage Manager Memory Limit is full. Producer ("
                                + message.getProducerId() + ") stopped to prevent flooding "
                                + getActiveMQDestination().getQualifiedName() + "."
                                + " See http://activemq.apache.org/producer-flow-control.html for more info");
  }
{code}

There is only a single transport thread that services the TCP socket on Broker B.  This TCP socket is the "remote" end of the A->B bridge and is responsible for *sequentially* enqueueing to Broker B's queues all messages from Broker A.  When a non-persistent or transactional message is sent to Broker A's queues, it has +responseRequired=true+, which is preserved when the bridge forwards the message to Broker B's queues.  If producer flow control is triggered on Broker B's queue, the first method of producer flow control will be used: the transport thread will not block, but the repsonse will be held back until the queue has room.  As a result, the transport thread is free to continue enqueueing messages from the bridge, particularly those destined for queues that are not full (NOTE: since the network prefetch is 1 no new messages to the full queue will be forwarded until the response is returned).

When a persistent or non-transactional message is sent to Broker A's queues, it has {tt}responseRequired=false{tt}, which is preserved when the bridge forwards the message to Broker B's queues.  If producer flow control is triggered on Broker B's queue, the second method of producer flow control will be used: the transport thread will be blocked.  As a result, no other messages from the bridge will be forwarded, even those destined for queues that are not full.


  was:
Symptom
=======
Broker A produces messages to two queues, Q1 and Q2.  Broker B consumes messages from two queues, Q1 and Q2.  Broker A is connected by a demand forwarding bridge, over TCP, to Broker B so that messages produced to Q1/Q2 will be forwarded to the consumers on Broker B.

At some point, Broker B's instance of Q2 becomes full (e.g., because the Q2 consumer is slow), and this triggers producer flow control to halt new messages being sent to Broker B's Q2 over the bridge.  Broker A's instances of Q1/Q2 are not full, so the producers on Broker A are not blocked.

If the messages produced by Broker A are *persistent*, we see this behaviour over the course of the production of 1000 messages to both Q1/Q2, where Broker B's Q2 becomes full on the 500th message:

{noformat}
Broker A       Bridge        Broker B
========                     ========
0->1000->0     ------>       0->1000->...
0->1000->500                 0->500->...
{noformat}

The above results, which assume network and consumer prefetch sizes of 1, are what we expected, namely: 
# Broker A produces 1000 messages to Q1 without blocking and all of these messages are forwarded to Broker B's Q1 without blocking, eventually being consumed by Broker B's Q1 consumer.
# Broker A produces 1000 messages to Q2 without blocking and 500 of these messages are forwarded to Broker B's Q2 before producer flow control blocks the flow until Broker B's Q2 consumer can start reducing the queue size.


This is good because the bridge treats Q1 and Q2 independently (i.e., producer flow control on Q2 does not block the messages forwarded to Q1).

If the messages produced by Broker A are *non-persistent*, we see this behaviour:

{noformat}
Broker A       Bridge        Broker B
========                     ========
0->1000->500   ------>       0->500->...
0->1000->500                 0->500->...
{noformat}

The above results, which assume network and consumer prefetch sizes of 1, are not what we expected, namely: producer flow control on Broker B's instance of Q2 blocks the forwarding of messages to Broker B's instance of Q1.

This is not good because producer flow control on Q2 essentially triggers producer flow control on Q1, even though Q1 is *not* full.

It also seems strange (and almost non-intuitive until you understand the cause), that peristent messages should behave better than non-persistent messages.  The same difference in behaviour can also be observed with persistent messages if Broker A these outside a JMS transaction (e.g., AUTO_ACKNOWLEDGE) versus inside a JMS transaction: outside behaves appropriately, with Q1 independent of Q2, but inside behaves the same as the non-persistent case with Q1 blocked by Q2.

These observations are contrary to the AMQ 5.0 documentation regarding producer flow control: {quote}As of ActiveMQ 5.0, we can now individually flow control each producer on a shared connection without having to suspend the entire connection.{quote}

Cause
=====

The difference in behaviour between persistent and non-persistent (and transactionaly/non-transactional) is due to the two ways that org.apache.activemq.broker.region.Queue implements producer flow control:

{code:title=Queue#send(...)}
// We can avoid blocking due to low usage if the producer is
// sending
// a sync message or if it is using a producer window
if (producerInfo.getWindowSize() > 0 || message.isResponseRequired()) {
{code}

and

{code:title=Queue#send(...)}
} else 
  if (memoryUsage.isFull()) {
    waitForSpace(context, memoryUsage, "Usage Manager Memory Limit is full. Producer ("
                                + message.getProducerId() + ") stopped to prevent flooding "
                                + getActiveMQDestination().getQualifiedName() + "."
                                + " See http://activemq.apache.org/producer-flow-control.html for more info");
  }
{code}

There is only a single transport thread that services the TCP socket on Broker B.  This TCP socket is the "remote" end of the A->B bridge and is responsible for *sequentially* enqueueing to Broker B's queues all messages from Broker A.  When a non-persistent or transactional message is sent to Broker A's queues, it has {tt}responseRequired=true{tt}, which is preserved when the bridge forwards the message to Broker B's queues.  If producer flow control is triggered on Broker B's queue, the first method of producer flow control will be used: the transport thread will not block, but the repsonse will be held back until the queue has room.  As a result, the transport thread is free to continue enqueueing messages from the bridge, particularly those destined for queues that are not full (NOTE: since the network prefetch is 1 no new messages to the full queue will be forwarded until the response is returned).

When a persistent or non-transactional message is sent to Broker A's queues, it has {tt}responseRequired=false{tt}, which is preserved when the bridge forwards the message to Broker B's queues.  If producer flow control is triggered on Broker B's queue, the second method of producer flow control will be used: the transport thread will be blocked.  As a result, no other messages from the bridge will be forwarded, even those destined for queues that are not full.


> When a producer from a network bridge is blocked by producer flow control, all producers from the network bridge get blocked.
> -----------------------------------------------------------------------------------------------------------------------------
>
>                 Key: AMQ-3331
>                 URL: https://issues.apache.org/jira/browse/AMQ-3331
>             Project: ActiveMQ
>          Issue Type: Bug
>          Components: Broker, Test Cases, Transport
>    Affects Versions: 5.5.0
>            Reporter: Stirling Chow
>
> Symptom
> =======
> Broker A produces messages to two queues, Q1 and Q2.  Broker B consumes messages from two queues, Q1 and Q2.  Broker A is connected by a demand forwarding bridge, over TCP, to Broker B so that messages produced to Q1/Q2 will be forwarded to the consumers on Broker B.
> At some point, Broker B's instance of Q2 becomes full (e.g., because the Q2 consumer is slow), and this triggers producer flow control to halt new messages being sent to Broker B's Q2 over the bridge.  Broker A's instances of Q1/Q2 are not full, so the producers on Broker A are not blocked.
> If the messages produced by Broker A are *persistent*, we see this behaviour over the course of the production of 1000 messages to both Q1/Q2, where Broker B's Q2 becomes full on the 500th message:
> {noformat}
> Broker A       Bridge        Broker B
> ========                     ========
> 0->1000->0     ------>       0->1000->...
> 0->1000->500                 0->500->...
> {noformat}
> The above results, which assume network and consumer prefetch sizes of 1, are what we expected, namely: 
> # Broker A produces 1000 messages to Q1 without blocking and all of these messages are forwarded to Broker B's Q1 without blocking, eventually being consumed by Broker B's Q1 consumer.
> # Broker A produces 1000 messages to Q2 without blocking and 500 of these messages are forwarded to Broker B's Q2 before producer flow control blocks the flow until Broker B's Q2 consumer can start reducing the queue size.
> This is good because the bridge treats Q1 and Q2 independently (i.e., producer flow control on Q2 does not block the messages forwarded to Q1).
> If the messages produced by Broker A are *non-persistent*, we see this behaviour:
> {noformat}
> Broker A       Bridge        Broker B
> ========                     ========
> 0->1000->500   ------>       0->500->...
> 0->1000->500                 0->500->...
> {noformat}
> The above results, which assume network and consumer prefetch sizes of 1, are not what we expected, namely: producer flow control on Broker B's instance of Q2 blocks the forwarding of messages to Broker B's instance of Q1.
> This is not good because producer flow control on Q2 essentially triggers producer flow control on Q1, even though Q1 is *not* full.
> It also seems strange (and almost non-intuitive until you understand the cause), that peristent messages should behave better than non-persistent messages.  The same difference in behaviour can also be observed with persistent messages if Broker A these outside a JMS transaction (e.g., AUTO_ACKNOWLEDGE) versus inside a JMS transaction: outside behaves appropriately, with Q1 independent of Q2, but inside behaves the same as the non-persistent case with Q1 blocked by Q2.
> These observations are contrary to the AMQ 5.0 documentation regarding producer flow control: {quote}As of ActiveMQ 5.0, we can now individually flow control each producer on a shared connection without having to suspend the entire connection.{quote}
> Cause
> =====
> The difference in behaviour between persistent and non-persistent (and transactionaly/non-transactional) is due to the two ways that org.apache.activemq.broker.region.Queue implements producer flow control:
> {code:title=Queue#send(...)}
> // We can avoid blocking due to low usage if the producer is
> // sending
> // a sync message or if it is using a producer window
> if (producerInfo.getWindowSize() > 0 || message.isResponseRequired()) {
> {code}
> and
> {code:title=Queue#send(...)}
> } else 
>   if (memoryUsage.isFull()) {
>     waitForSpace(context, memoryUsage, "Usage Manager Memory Limit is full. Producer ("
>                                 + message.getProducerId() + ") stopped to prevent flooding "
>                                 + getActiveMQDestination().getQualifiedName() + "."
>                                 + " See http://activemq.apache.org/producer-flow-control.html for more info");
>   }
> {code}
> There is only a single transport thread that services the TCP socket on Broker B.  This TCP socket is the "remote" end of the A->B bridge and is responsible for *sequentially* enqueueing to Broker B's queues all messages from Broker A.  When a non-persistent or transactional message is sent to Broker A's queues, it has +responseRequired=true+, which is preserved when the bridge forwards the message to Broker B's queues.  If producer flow control is triggered on Broker B's queue, the first method of producer flow control will be used: the transport thread will not block, but the repsonse will be held back until the queue has room.  As a result, the transport thread is free to continue enqueueing messages from the bridge, particularly those destined for queues that are not full (NOTE: since the network prefetch is 1 no new messages to the full queue will be forwarded until the response is returned).
> When a persistent or non-transactional message is sent to Broker A's queues, it has {tt}responseRequired=false{tt}, which is preserved when the bridge forwards the message to Broker B's queues.  If producer flow control is triggered on Broker B's queue, the second method of producer flow control will be used: the transport thread will be blocked.  As a result, no other messages from the bridge will be forwarded, even those destined for queues that are not full.

--
This message is automatically generated by JIRA.
For more information on JIRA, see: http://www.atlassian.com/software/jira