<!DOCTYPE HTML PUBLIC "-//W3C//DTD HTML 4.01 Transitional//EN">
<html>
<head>
<meta content="text/html; charset=UTF-8" http-equiv="Content-Type">
</head>
<body text="#000000" bgcolor="#ffffff">
Hi all,<br>
<br>
Recently I came up with a solution that can help with the thread
pool problem motivated by the following: <br>
<br>
In one of the first implementation of Total Order based commit
protocol (TO), I had the requirement to move the PrepareCommand to
another thread pool. In resume, the TO protocol delivers the
PrepareCommand in a deterministic order in all the nodes, by a
single deliver thread. To ensure consistency, if it delivers two
conflicting transactions, the second transaction must wait until the
first transaction finishes. However, blocking single deliver thread
is not a good solution, because no more transaction can be
validated, even if they don't conflict, while the thread is blocked.<br>
<br>
So, after creating a dependency graph (i.e. the second transaction
knows that it must wait for the first transaction to finish) I move
the PrepareCommand to another thread pool. Initially, I implemented
a new command, called PrepareResponseCommand, that sends back the
reply of the PrepareCommand. This solution has one disadvantage: I
had to implement an ack collector in ISPN, while JGroups already
offers me that with a synchronous communication.<br>
<br>
Recently (2 or 3 months ago) I implemented a simple modification in
JGroups. In a more generic approach, it allows other threads to
reply to a RPC request (such as the PrepareCommand). In the previous
scenario, I replaced the PrepareResponseCommand and the ack
collector implementation with a synchronous RPC invocation. I've
used this solution in other issues in the Cloud-TM's ISPN fork. <br>
<br>
This solution is quite simple to implement and may help you to move
the commands to ISPN internal thread pools. The modifications I've
made are the following:<br>
<br>
1) I added a new interface (see [1]) that is sent to the application
instead of the Message object (see [4]). This interface contains the
Message and it has a method to allow the application send the reply
to that particular request.<br>
2) I added a new object in [4] with the meaning: this return value
is not the reply to the RPC request. This is the returned value that
I return when I want to release the thread, because ISPN should
return some object in the handle() method. Of course, I know that
ISPN will invoke the sendReply() in some other place, otherwise, I
will get a TimeoutException in the sender side.<br>
3) Also I've changed the RequestCorrelator implementation to support
the previous modifications (see [2] and [3])<br>
<br>
In the Cloud-TM's ISPN fork I added a reference in the
BaseCacheRpcCommand to [1] and added the method sendReply() [5]. In
addition, I have the following uses cases working perfectly with
this:<br>
<br>
1) Total Order<br>
<br>
The scenario described in the beginning. The
ParallelTotalOrderManager returns the DO_NOT_REPLY object when it
receives a remote PrepareCommand (see [6] line 77). When the
PrepareCommand is finally processed by the rest of the interceptor
chain, it invokes the PreapreCommand.sendReply() (see [6] line 230).<br>
<br>
2) GMU remote get<br>
<br>
GMU ensures SERIALIZABLE Isolation Level and the remote gets must
ensure that the node that is processing the request has a minimum
version available to ensure data consistency. The problem in ours
initial implementation in large cluster, is the number of remote
gets are very high and all the OOB are being blocked because of this
condition.<br>
<br>
Same thing I've done with the ClusteredRemoteGet as you can in see
[7], line 93 and 105.<br>
<br>
3) GMU CommitCommand<br>
<br>
In GMU, the CommitCommand cannot be processed by any order. If T1 is
serialized before T2, the commit command of T1 must be processed
before the commit command of T2, even if the transactions do not
have conflicts. This generates the same problem above and the same
solution was adopted.<br>
<br>
I know that you have discussed some solutions and I would like to
know what it is your opinion about what I've described. <br>
<br>
If you have questions, please let me know.<br>
<br>
Cheers,<br>
Pedro<br>
<br>
[1] <a
href="https://github.com/pruivo/JGroups/blob/t_cloudtm/src/org/jgroups/blocks/MessageRequest.java"
target="_blank">https://github.com/pruivo/<wbr>JGroups/blob/t_cloudtm/src/<wbr>org/jgroups/blocks/<wbr>MessageRequest.java</a><br>
[2] <a
href="https://github.com/pruivo/JGroups/blob/t_cloudtm/src/org/jgroups/blocks/RequestCorrelator.java#L463"
target="_blank">https://github.com/pruivo/<wbr>JGroups/blob/t_cloudtm/src/<wbr>org/jgroups/blocks/<wbr>RequestCorrelator.java#L463</a><br>
[3] <a
href="https://github.com/pruivo/JGroups/blob/t_cloudtm/src/org/jgroups/blocks/RequestCorrelator.java#L495"
target="_blank">https://github.com/pruivo/<wbr>JGroups/blob/t_cloudtm/src/<wbr>org/jgroups/blocks/<wbr>RequestCorrelator.java#L495</a><br>
[4] <a
href="https://github.com/pruivo/JGroups/blob/t_cloudtm/src/org/jgroups/blocks/RequestHandler.java"
target="_blank">https://github.com/pruivo/<wbr>JGroups/blob/t_cloudtm/src/<wbr>org/jgroups/blocks/<wbr>RequestHandler.java</a><br>
[5] <a
href="https://github.com/pruivo/infinispan/blob/cloudtm_v1/core/src/main/java/org/infinispan/commands/remote/BaseRpcCommand.java#L75"
target="_blank">https://github.com/pruivo/<wbr>infinispan/blob/cloudtm_v1/<wbr>core/src/main/java/org/<wbr>infinispan/commands/remote/<wbr>BaseRpcCommand.java#L75</a><br>
[6]
<a class="moz-txt-link-freetext" href="https://github.com/pruivo/infinispan/blob/cloudtm_v1/core/src/main/java/org/infinispan/transaction/totalorder/ParallelTotalOrderManager.java">https://github.com/pruivo/infinispan/blob/cloudtm_v1/core/src/main/java/org/infinispan/transaction/totalorder/ParallelTotalOrderManager.java</a><br>
[7]
<a class="moz-txt-link-freetext" href="https://github.com/pruivo/infinispan/blob/cloudtm_v1/core/src/main/java/org/infinispan/commands/remote/GMUClusteredGetCommand.java">https://github.com/pruivo/infinispan/blob/cloudtm_v1/core/src/main/java/org/infinispan/commands/remote/GMUClusteredGetCommand.java</a><br>
<br>
<br>
On 2/3/13 11:35 AM, Bela Ban wrote:
<blockquote cite="mid:510E4B98.40608@redhat.com" type="cite">
<pre wrap="">If you send me the details, I'll take a look. I'm pretty busy with
message batching, so I can't promise next week, but soon...
On 2/1/13 11:08 AM, Pedro Ruivo wrote:
</pre>
<blockquote type="cite">
<pre wrap="">Hi,
I had a similar problem when I tried GMU[1] in "large" cluster (40 vms),
because the remote gets and the commit messages (I'm talking about ISPN
commands) must wait for some conditions before being processed.
I solved this problem by adding a feature in JGroups[2] that allows the
request to be moved to another thread, releasing the OOB thread. The
other thread will send the reply of the JGroups Request. Of course, I'm
only moving commands that I know they can block.
I can enter in some detail if you want =)
Cheers,
Pedro
[1] <a class="moz-txt-link-freetext" href="http://www.gsd.inesc-id.pt/~romanop/files/papers/icdcs12.pdf">http://www.gsd.inesc-id.pt/~romanop/files/papers/icdcs12.pdf</a>
[2] I would like to talk with Bela about this, because it makes my life
easier to support total order in ISPN. I'll try to send an email this
weekend =)
On 01-02-2013 08:04, Radim Vansa wrote:
</pre>
<blockquote type="cite">
<pre wrap="">Hi guys,
after dealing with the large cluster for a while I find the way how we use OOB threads in synchronous configuration non-robust.
Imagine a situation where node which is not an owner of the key calls PUT. Then the a RPC is called to the primary owner of that key, which reroutes the request to all other owners and after these reply, it replies back.
There are two problems:
1) If we do simultanously X requests from non-owners to the primary owner where X is OOB TP size, all the OOB threads are waiting for the responses and there is no thread to process the OOB response and release the thread.
2) Node A is primary owner of keyA, non-primary owner of keyB and B is primary of keyB and non-primary of keyA. We got many requests for both keyA and keyB from other nodes, therefore, all OOB threads from both nodes call RPC to the non-primary owner but there's noone who could process the request.
While we wait for the requests to timeout, the nodes with depleted OOB threadpools start suspecting all other nodes because they can't receive heartbeats etc...
You can say "increase your OOB tp size", but that's not always an option, I have currently set it to 1000 threads and it's not enough. In the end, I will be always limited by RAM and something tells me that even nodes with few gigs of RAM should be able to form a huge cluster. We use 160 hotrod worker threads in JDG, that means that 160 * clusterSize = 10240 (64 nodes in my cluster) parallel requests can be executed, and if 10% targets the same node with 1000 OOB threads, it stucks. It's about scaling and robustness.
Not that I'd have any good solution, but I'd really like to start a discussion.
Thinking about it a bit, the problem is that blocking call (calling RPC on primary owner from message handler) can block non-blocking calls (such as RPC response or command that never sends any more messages). Therefore, having a flag on message "this won't send another message" could let the message be executed in different threadpool, which will be never deadlocked. In fact, the pools could share the threads but the non-blocking would have always a few threads spare.
It's a bad solution as maintaining which message could block in the other node is really, really hard (we can be sure only in case of RPC responses), especially when some locks come. I will welcome anything better.
Radim
-----------------------------------------------------------
Radim Vansa
Quality Assurance Engineer
JBoss Datagrid
tel. +420532294559 ext. 62559
Red Hat Czech, s.r.o.
Brno, Purkyňova 99/71, PSČ 612 45
Czech Republic
_______________________________________________
infinispan-dev mailing list
<a class="moz-txt-link-abbreviated" href="mailto:infinispan-dev@lists.jboss.org">infinispan-dev@lists.jboss.org</a>
<a class="moz-txt-link-freetext" href="https://lists.jboss.org/mailman/listinfo/infinispan-dev">https://lists.jboss.org/mailman/listinfo/infinispan-dev</a>
</pre>
</blockquote>
<pre wrap="">_______________________________________________
infinispan-dev mailing list
<a class="moz-txt-link-abbreviated" href="mailto:infinispan-dev@lists.jboss.org">infinispan-dev@lists.jboss.org</a>
<a class="moz-txt-link-freetext" href="https://lists.jboss.org/mailman/listinfo/infinispan-dev">https://lists.jboss.org/mailman/listinfo/infinispan-dev</a>
</pre>
</blockquote>
<pre wrap="">
</pre>
</blockquote>
</body>
</html>