You are viewing a plain text version of this content. The canonical link for it is here.
Posted to users@qpid.apache.org by Ramayan Tiwari <ra...@gmail.com> on 2017/01/04 22:54:51 UTC
Re: Qpid broker 6.0.4 performance issues
Hi Lorenz,
Happy new year to everyone, hope you guys had fun!
I am doing performance test runs to figure out a reasonable threshold for
direct memory, considering our use case of small message payloads. I have a
few more questions:
1. I assume there is no way to disable direct memory (so that we always use
heap to store messages), however, I was wondering if this would be
something that can be offered as part of broker config (unless it a
significant change and unfeasible to do it).
2. We monitor Qpid broker memory and stop enqueue at a certain threshold to
protect broker heap. Now that message payload is stored in DM, we would
like to track that as well. However, DM size reported by Mbean doesn't
correlate directly with the sum of message payload (possibly because the
byte buffers are not allocated/de-allocated that frequently). So to get
around that, we are considering using the sum queue sizes to get actual DM
usage.
Thanks
Ramayan
On Wed, Dec 21, 2016 at 5:17 AM, Lorenz Quack <qu...@gmail.com>
wrote:
> Hi,
>
> Regarding 0.32 behaviour, it checked to see whether to flow a message
> to disk when putting a message on the Queue the same way Qpid 6 does.
> In that sense 6 is not more or less aggressive. However, the
> algorithm behind the decision whether or not to flow to disk has
> changed. This change was done as part of a larger effort to isolate
> VirtualHosts from each other. I would have to go back and check how
> the algorithm worked previously but I would assume that it just
> considered the total (estimated) amount of memory used and did no per
> VH or per Queue allocation. This means 6 effectively lowers the
> threshold on individual Queues especially in the case where some
> VirtualHosts and/or Queues are used less than others. On the upside,
> the broker is fairer in its resource management and a single
> VirtualHost can no longer use up all available memory. How exactly
> that trade-off between fairness and efficient use of available memory
> is made is debatable but I don't think we want to go back to the pre-6
> model of just lumping everything together.
>
> Given your numbers (1 VH, 6000 Qs) each Queue would initially be
> allocated 1/6000th of 60% of 8 GB ≈ 1 MB. In then end state the full
> Queues should end up with approximately 780 MB but as you noticed the
> threshold is only recalculated periodically during housekeeping (by
> default every 30 s) or when a VH or Queue is added or deleted. If you
> have DEBUG logging you should see periodic messages like "Allocating
> target size to queues [...]" if not then I am afraid you won't be
> able to tell the current thresholds because they are only reported
> once when flowToDisk becomes active/inactive.
>
> So I think your analysis is probably correct that the revision of the
> threshold is always "behind" the publishing, raising it on every
> revision but never far enough to prevent flowToDisk. This is not
> ideal. We will have to address this. However, I am afraid that in
> the current release there is no way to influence the algorithm other
> than setting the available memory and broker.flowToDiskThreshold.
>
> Regarding the MemoryStore, the algorithm triggering flowToDisk is the
> same for all stores, just the implementation of the actual writing
> messages to disk differs. For the MemoryStore it is a noop, i.e., the
> message is not flown to disk and remains in memory. Performancewise
> we do not do a lot of testing with the MemoryStore because it is not a
> typical use-case and mainly used for unit and system testing. I would
> assume that the better distribution you are seeing is coincidental
> since that part of the code should be relatively independent of the
> store type. Unfortunately, I cannot see any of your graphs. I
> believe the mailing list strips all attachments.
>
> Regarding a recommendation of how to configure your DM vs Heap I would
> like to refer you to our documentation [1], especially section
> "9.11.6. Memory Tuning the Broker". There we provide formulas to
> estimate the memory consumption of the broker for both DM and Heap.
> Note that these are estimates and you should test your chosen settings
> under a typical peak workload. Given that your messages are small you
> will probably want to favour Heap over DM but I am reluctant to make
> an explicit recommendation.
>
> Kind regards,
> Lorenz
>
> P.S.: I am going on a 2 day vacation later today but feel free to
> continue this conversation with others on this list.
>
> [1] https://qpid.apache.org/releases/qpid-java-6.1.0/java-broker
> /book/Java-Broker-Runtime-Memory.html
>
> On 20/12/16 20:37, Ramayan Tiwari wrote:
>
>> Hi Lorenz,
>>
>> Thanks a lot for your response and explaining the flow to disk algorithm
>> in detail. I have described the test setup in detail in the first email of
>> this thread, to summarize the points again:
>> a) There is only one virtual host.
>> b) There are 6000 queues in this virtual host, but messages are only
>> enqueued to 10 queues.
>> c) Every queue gets equal number of messages (100k) at the start of the
>> test (we do not start dequeue till all the 1 million messages are enqueued).
>> d) Heap and DM memory are equal (8GB each) and DM flow to disk threshold
>> is 60%.
>>
>> I looked at QUE-1014/15 log lines and following is what I notice:
>> a) These log lines are not present in 0.32 broker's log, which means that
>> its not doing any flow to disk. Is flow to disk behavior different in the
>> two brokers, it looks like 6.0.x is a lot more aggressive in this regard.
>>
>> b) Since all the 1 million messages are enqueued at the start of test
>> (takes about 7 mins to enqueue), flow to disk threshold revisions performed
>> by the housekeeping task are not able to catch up. Or the rate with which
>> thresholds are revised can not catch up with the rate of enqueue. In my
>> test, revisions once happened twice (4 seconds and 5 mins after test start)
>> and then on, the threshold was not revised for the queues.
>>
>> To make sure that we are not getting penalized by writing to disk, I also
>> did a test using Memory store type and compared the result with BDB store
>> type. Apparently, BDB store is slightly more efficient (2.7%) in terms of
>> number of messages delivered. Memory store also takes more broker CPU (3%
>> more on average), but its better in terms of distributing messages in a
>> round robin manner from all the queues. See the attached graphs for details.
>>
>> I do notice that flow to disk behavior is almost exactly same
>> (QUE-1014/15 log lines are present) when running with Memory store. I am
>> wondering what does flow to disk does when we use Memory store?
>>
>> Since our average messages size is less than 1KB, I am really looking
>> forward to some recommendation around the % allocation for DM vs Heap.
>>
>> Thanks
>> Ramayan
>>
>>
>> On Tue, Dec 20, 2016 at 4:02 AM, Lorenz Quack <quack.lorenz@gmail.com
>> <ma...@gmail.com>> wrote:
>>
>> Hello Ramayan,
>>
>> glad to hear that the patch is (mostly) working for you.
>> To address your points:
>>
>> 1. If indeed in one case flow to disk is kicking in while in
>> the other one it is not, then I am not surprised that
>> there is a 5% difference. The question is whether the
>> flow to disk is expected or not which leads to
>>
>> 2. The direct memory utilization not exceeding a certain
>> value is a strong indication that flow to disk is active.
>> Could you verify that by checking the logs (QUE-1014/15)?
>> If the flow to disk limit is exceeded then it is expected
>> that 2 million messages consume the same amount of direct
>> memory as 1 million messages. Could you share a little
>> more about the test setup? How many VirtualHost are
>> running on the broker? How many Queues are on each
>> VirtualHost? What is the Queue depth of those Queues?
>> All of those factors influence the actual flow to disk
>> threshold. This is to ensure some fairness between
>> VirtualHosts as far as memory consumption is concerned.
>> Below I explain how threshold allocation is currently
>> performed. We are considering changing the algorithm in
>> the future or making it tunable. Your ideas, requirements,
>> and input on this would certainly be of interest to us.
>>
>> Looking forward to hearing from you.
>>
>> Kind regards,
>> Lorenz
>>
>>
>> Algorithm for flow to disk threshold:
>>
>> 1. Take the total amount of the broker.flowToDiskThreshold and
>> divide it amongst all active VirtualHosts as follows
>>
>> a. Half of broker.flowToDiskThreshold is evenly devided
>> amongst the VHs to ensure a minimum amount is available to
>> each VH.
>>
>> b. The remaining half is allocated proportional to the current
>> usage pattern. For example, if VH1 is currently using 3
>> MB, VH2 is using 1 MB and VH3 is using 0 MB, then of the
>> remaining half 3/4 will be allocated to VH1, 1/4 to VH2,
>> and nothing to VH3. If all VHs are empty distribute this
>> half evenly like in 1.a.
>>
>> 2. The VirtualHosts allocate their available memory to their
>> Queues in a proportional fashion as explained above (1.b).
>>
>>
>> Example:
>>
>> * The broker.flowToDiskThreshold is set to 10 GB.
>>
>> * Two Virtual Hosts with 10 Queues each.
>>
>> * VH1 all 10 Queues are empty.
>>
>> * VH2 all Queues contain 10 MB except of one Queue that
>> contains 100 MB.
>>
>> * According to 1.a each VirtualHost is allocated half of 5 GB,
>> i.e., 2.5 GB
>>
>> * According to 1.b VH1 using 0MB does not get any additional
>> memory while VH2 gets the full of the remainder of the 5 GB
>> totaling 7.5 GB.
>>
>> * The Queues on VH1 don't have messages on them so the
>> VirtualHost falls back to allocating them equal shares: 250 MB
>> each.
>>
>> * On VH2 the total current memory usage is 9*10 MB + 100 MB =
>> 190 MB so the smaller Queues receive 10/190 * 7.5 GB = 395 MB
>> while the large Queue receives 100/190 * 7.5 GB = 3950 MB.
>>
>> * In total we allocated 10 * 250 MB + 9 * 395 MB + 1 * 3950 MB
>> totaling 10 GB (within bounds of rounding errors).
>>
>>
>>
>> On 19/12/16 20:48, Ramayan Tiwari wrote:
>>
>> Hi Rob,
>>
>> I did another exhaustive performance test using the
>> MultiQueueConsumer feature with 6.0.5 (and the patch). The
>> broker CPU issues has been resolved and we no longer have the
>> problem message prefetch (caused by long running message).
>>
>> Fairness among queue is also great (not as perfect as 0.32
>> broker though, see attached graphs). Everything looks great,
>> except for:
>>
>> 1. 6.0.5 delivered around 4.6% less messages. Flow to disk
>> triggered aggressively in 6.0.5 but I don't see any flow to
>> disk happening in 0.32 (looking for QUE-1014). This might be
>> the reason for lesser message delivery.
>>
>> 2. Direct memory utilization in the new broker does not make
>> sense to us. We did 2 tests: 1 millions and 2 million messages
>> (220 Byte average message size), however, the direct memory
>> utilization never exceeded 500MB (see attached graph), even
>> when we are allocating 8GB for direct memory. Because there is
>> a 1KB heap overhead with each message, heap utilization looks
>> same for both 0.32 and 6.0.5. For our setup, this essentially
>> means that, we are cutting our memory capacity by half,
>> because now are allocating half of the available RAM to direct
>> memory, but will be limited by heap anyway.
>>
>> These tests were performed using 16GB RAM, where 8GB was
>> allocated to heap and 8GB for Direct memory. I also changed
>> flowToDiskThreshold to 60%. This is one of our biggest concern
>> with the new broker, since our average message size in
>> production is less than 1KB. Currently we allocate all the
>> available RAM to heap, which will be reduced in half with the
>> new broker.
>>
>> What is the recommendation for memory allocation (heap vs dm)
>> in our use case?
>>
>> Thanks
>> Ramaayn
>>
>> On Fri, Oct 28, 2016 at 5:37 AM, Keith W <keith.wall@gmail.com
>> <ma...@gmail.com> <mailto:keith.wall@gmail.com
>>
>> <ma...@gmail.com>>> wrote:
>>
>> Hi Ramayan
>>
>> QPID-7462 is a new (experimental) feature, so we don't
>> consider this
>> appropriate for inclusion in the 6.0.5 defect release We
>> follow a
>> Semantic Versioning[1] strategy.
>>
>> The underlying issue is your testing has uncovered is poor
>> performance
>> with large numbers of consumers. QPID-7462 effectively
>> side steps the
>> problem (by introducing alternative consumer behaviour)
>> but does not
>> address the root cause. We continue to consider how best
>> to resolve
>> the problem completely, but don't yet have timelines for
>> this change.
>> It is something that will be getting attention in what
>> remains of this
>> year. We will keep you posted.
>>
>> In the meanwhile, I understand this causes you a problem.
>> If you
>> cannot adopt 6.1 (there should be another RC out soon),
>> you could
>> consider applying the patch (attached to the JIRA) to
>> 6.0.x branch and
>> building yourself.
>>
>> Kind regards, Keith.
>>
>>
>> [1] http://semver.org
>>
>>
>> On 27 October 2016 at 23:19, Ramayan Tiwari
>> <ramayan.tiwari@gmail.com
>> <ma...@gmail.com>
>> <mailto:ramayan.tiwari@gmail.com
>> <ma...@gmail.com>>> wrote:
>> > Hi Rob,
>> >
>> > I have the truck code which I am testing with, I haven't
>> finished the test
>> > runs yet. I was hoping that once I validate the change,
>> I can simply
>> > release 6.0.5.
>> >
>> > Thanks
>> > Ramayan
>> >
>> > On Thu, Oct 27, 2016 at 12:41 PM, Rob Godfrey
>> <rob.j.godfrey@gmail.com <ma...@gmail.com>
>> <mailto:rob.j.godfrey@gmail.com <mailto:rob.j.godfrey@gmail.com
>> >>>
>> > wrote:
>> >
>> >> Hi Ramayan,
>> >>
>> >> did you verify that the change works for you? You said
>> you were
>> going to
>> >> test with the trunk code...
>> >>
>> >> I'll discuss with the other developers tomorrow about
>> whether
>> we can put
>> >> this change into 6.0.5.
>> >>
>> >> Cheers,
>> >> Rob
>> >>
>> >> On 27 October 2016 at 20:30, Ramayan Tiwari
>> <ramayan.tiwari@gmail.com
>> <ma...@gmail.com>
>> <mailto:ramayan.tiwari@gmail.com
>> <ma...@gmail.com>>>
>> >> wrote:
>> >>
>> >> > Hi Rob,
>> >> >
>> >> > I looked at the release notes for 6.0.5 and it doesn't
>> include the fix
>> >> for
>> >> > large consumers issues [1]. The fix is marked for
>> 6.1, which
>> will not
>> >> have
>> >> > JMX and for us to use this version requires major
>> changes in our
>> >> monitoring
>> >> > framework. Could you please include the fix in 6.0.5
>> release?
>> >> >
>> >> > Thanks
>> >> > Ramayan
>> >> >
>> >> > [1]. https://issues.apache.org/jira/browse/QPID-7462
>> <https://issues.apache.org/jira/browse/QPID-7462>
>> <https://issues.apache.org/jira/browse/QPID-7462
>> <https://issues.apache.org/jira/browse/QPID-7462>>
>> >> >
>> >> > On Wed, Oct 19, 2016 at 4:49 PM, Helen Kwong
>> <helenkwong@gmail.com <ma...@gmail.com>
>> <mailto:helenkwong@gmail.com <ma...@gmail.com>>>
>> >> wrote:
>> >> >
>> >> > > Hi Rob,
>> >> > >
>> >> > > Again, thank you so much for answering our
>> questions and
>> providing a
>> >> > patch
>> >> > > so quickly :) One more question I have: would it be
>> possible to include
>> >> > > test cases involving many queues and listeners (in
>> the order of
>> >> thousands
>> >> > > of queues) for future Qpid releases, as part of
>> standard
>> perf testing
>> >> of
>> >> > > the broker?
>> >> > >
>> >> > > Thanks,
>> >> > > Helen
>> >> > >
>> >> > > On Tue, Oct 18, 2016 at 10:40 AM, Ramayan Tiwari <
>> >> > ramayan.tiwari@gmail.com
>> <ma...@gmail.com>
>> <mailto:ramayan.tiwari@gmail.com
>> <ma...@gmail.com>>
>> >> > > > wrote:
>> >> > >
>> >> > >> Thanks so much Rob, I will test the patch against
>> trunk
>> and will
>> >> update
>> >> > >> you with the outcome.
>> >> > >>
>> >> > >> - Ramayan
>> >> > >>
>> >> > >> On Tue, Oct 18, 2016 at 2:37 AM, Rob Godfrey
>> <rob.j.godfrey@gmail.com <ma...@gmail.com>
>> <mailto:rob.j.godfrey@gmail.com <ma...@gmail.com>>
>> >> >
>> >> > >> wrote:
>> >> > >>
>> >> > >>> On 17 October 2016 at 21:50, Rob Godfrey
>> <rob.j.godfrey@gmail.com <ma...@gmail.com>
>> <mailto:rob.j.godfrey@gmail.com <mailto:rob.j.godfrey@gmail.com
>> >>>
>> >> > >>> wrote:
>> >> > >>>
>> >> > >>> >
>> >> > >>> >
>> >> > >>> > On 17 October 2016 at 21:24, Ramayan Tiwari <
>> >> > ramayan.tiwari@gmail.com
>> <ma...@gmail.com>
>> <mailto:ramayan.tiwari@gmail.com
>>
>> <ma...@gmail.com>>>
>>
>> >> > >>> > wrote:
>> >> > >>> >
>> >> > >>> >> Hi Rob,
>> >> > >>> >>
>> >> > >>> >> We are certainly interested in testing the "multi
>> queue consumers"
>> >> > >>> >> behavior
>> >> > >>> >> with your patch in the new broker. We would
>> like to know:
>> >> > >>> >>
>> >> > >>> >> 1. What will the scope of changes, client or
>> broker or
>> both? We
>> >> are
>> >> > >>> >> currently running 0.16 client, so would like
>> to make
>> sure that we
>> >> > will
>> >> > >>> >> able
>> >> > >>> >> to use these changes with 0.16 client.
>> >> > >>> >>
>> >> > >>> >>
>> >> > >>> > There's no change to the client. I can't
>> remember what
>> was in the
>> >> > 0.16
>> >> > >>> > client... the only issue would be if there are
>> any bugs
>> in the
>> >> > parsing
>> >> > >>> of
>> >> > >>> > address arguments. I can try to test that out tmr.
>> >> > >>> >
>> >> > >>>
>> >> > >>>
>> >> > >>> OK - with a little bit of care to get round the
>> address
>> parsing
>> >> issues
>> >> > in
>> >> > >>> the 0.16 client... I think we can get this to
>> work. I've
>> created the
>> >> > >>> following JIRA:
>> >> > >>>
>> >> > >>> https://issues.apache.org/jira/browse/QPID-7462
>> <https://issues.apache.org/jira/browse/QPID-7462>
>> <https://issues.apache.org/jira/browse/QPID-7462
>> <https://issues.apache.org/jira/browse/QPID-7462>>
>> >> > >>>
>> >> > >>> and attached to it are a patch which applies against
>> trunk, and a
>> >> > >>> separate
>> >> > >>> patch which applies against the 6.0.x branch (
>> >> > >>>
>> https://svn.apache.org/repos/asf/qpid/java/branches/6.0.x
>> <https://svn.apache.org/repos/asf/qpid/java/branches/6.0.x>
>> <https://svn.apache.org/repos/asf/qpid/java/branches/6.0.x
>> <https://svn.apache.org/repos/asf/qpid/java/branches/6.0.x>> -
>> this is
>> >> > >>> 6.0.4
>> >> > >>> plus a few other fixes which we will soon be
>> releasing as
>> 6.0.5)
>> >> > >>>
>> >> > >>> To create a consumer which uses this feature (and
>> multi queue
>> >> > >>> consumption)
>> >> > >>> for the 0.16 client you need to use something
>> like the
>> following as
>> >> the
>> >> > >>> address:
>> >> > >>>
>> >> > >>> queue_01 ; {node : { type : queue }, link : {
>> x-subscribes : {
>> >> > >>> arguments : { x-multiqueue : [ queue_01, queue_02,
>> queue_03 ],
>> >> > >>> x-pull-only : true }}}}
>> >> > >>>
>> >> > >>>
>> >> > >>> Note that the initial queue_01 has to be a name of an
>> actual queue on
>> >> > >>> the virtual host, but otherwise it is not
>> actually used
>> (if you were
>> >> > >>> using a 0.32 or later client you could just use ''
>> here). The actual
>> >> > >>> queues that are consumed from are in the list value
>> associated with
>> >> > >>> x-multiqueue. For my testing I created a list
>> with 3000
>> queues here
>> >> > >>> and this worked fine.
>> >> > >>>
>> >> > >>> Let me know if you have any questions / issues,
>> >> > >>>
>> >> > >>> Hope this helps,
>> >> > >>> Rob
>> >> > >>>
>> >> > >>>
>> >> > >>> >
>> >> > >>> >
>> >> > >>> >> 2. My understanding is that the "pull vs push"
>> change
>> is only with
>> >> > >>> respect
>> >> > >>> >> to broker and it does not change our architecture
>> where we use
>> >> > >>> >> MessageListerner to receive messages
>> asynchronously.
>> >> > >>> >>
>> >> > >>> >
>> >> > >>> > Exactly - this is only a change within the
>> internal broker
>> >> threading
>> >> > >>> > model. The external behaviour of the broker
>> remains
>> essentially
>> >> > >>> unchanged.
>> >> > >>> >
>> >> > >>> >
>> >> > >>> >>
>> >> > >>> >> 3. Once I/O refactoring is completely, we would be
>> able to go back
>> >> > to
>> >> > >>> use
>> >> > >>> >> standard JMS consumer (Destination), what is the
>> timeline and
>> >> broker
>> >> > >>> >> release version for the completion of this work?
>> >> > >>> >>
>> >> > >>> >
>> >> > >>> > You might wish to continue to use the "multi
>> queue" model,
>> >> depending
>> >> > on
>> >> > >>> > your actual use case, but yeah once the I/O work is
>> complete I
>> >> would
>> >> > >>> hope
>> >> > >>> > that you could use the thousands of consumers model
>> should you
>> >> wish.
>> >> > >>> We
>> >> > >>> > don't have a schedule for the next phase of I/O
>> rework
>> right now -
>> >> > >>> about
>> >> > >>> > all I can say is that it is unlikely to be complete
>> this year. I'd
>> >> > >>> need to
>> >> > >>> > talk with Keith (who is currently on vacation)
>> as to
>> when we think
>> >> we
>> >> > >>> may
>> >> > >>> > be able to schedule it.
>> >> > >>> >
>> >> > >>> >
>> >> > >>> >>
>> >> > >>> >> Let me know once you have integrated the patch
>> and I
>> will re-run
>> >> our
>> >> > >>> >> performance tests to validate it.
>> >> > >>> >>
>> >> > >>> >>
>> >> > >>> > I'll make a patch for 6.0.x presently (I've been
>> working on a
>> >> change
>> >> > >>> > against trunk - the patch will probably have to
>> change
>> a bit to
>> >> apply
>> >> > >>> to
>> >> > >>> > 6.0.x).
>> >> > >>> >
>> >> > >>> > Cheers,
>> >> > >>> > Rob
>> >> > >>> >
>> >> > >>> > Thanks
>> >> > >>> >> Ramayan
>> >> > >>> >>
>> >> > >>> >> On Sun, Oct 16, 2016 at 3:30 PM, Rob Godfrey <
>> >> > rob.j.godfrey@gmail.com
>> <ma...@gmail.com>
>> <mailto:rob.j.godfrey@gmail.com <ma...@gmail.com>>
>>
>>
>> >> > >>> >
>> >> > >>> >> wrote:
>> >> > >>> >>
>> >> > >>> >> > OK - so having pondered / hacked around a
>> bit this
>> weekend, I
>> >> > think
>> >> > >>> to
>> >> > >>> >> get
>> >> > >>> >> > decent performance from the IO model in 6.0
>> for your
>> use case
>> >> > we're
>> >> > >>> >> going
>> >> > >>> >> > to have to change things around a bit.
>> >> > >>> >> >
>> >> > >>> >> > Basically 6.0 is an intermediate step on our
>> IO /
>> threading
>> >> model
>> >> > >>> >> journey.
>> >> > >>> >> > In earlier versions we used 2 threads per
>> connection
>> for IO (one
>> >> > >>> read,
>> >> > >>> >> one
>> >> > >>> >> > write) and then extra threads from a pool to
>> "push"
>> messages
>> >> from
>> >> > >>> >> queues to
>> >> > >>> >> > connections.
>> >> > >>> >> >
>> >> > >>> >> > In 6.0 we move to using a pool for the IO
>> threads,
>> and also
>> >> > stopped
>> >> > >>> >> queues
>> >> > >>> >> > from "pushing" to connections while the IO
>> threads
>> were acting
>> >> on
>> >> > >>> the
>> >> > >>> >> > connection. It's this latter fact which is
>> screwing up
>> >> > performance
>> >> > >>> for
>> >> > >>> >> > your use case here because what happens is
>> that on
>> each network
>> >> > >>> read we
>> >> > >>> >> > tell each consumer to stop accepting pushes
>> from the
>> queue until
>> >> > >>> the IO
>> >> > >>> >> > interaction has completed. This is causing
>> lots of
>> loops over
>> >> > your
>> >> > >>> 3000
>> >> > >>> >> > consumers on each session, which is eating
>> up a lot
>> of CPU on
>> >> > every
>> >> > >>> >> network
>> >> > >>> >> > interaction.
>> >> > >>> >> >
>> >> > >>> >> > In the final version of our IO refactoring
>> we want
>> to remove the
>> >> > >>> >> "pushing"
>> >> > >>> >> > from the queue, and instead have the consumers
>> "pull" - so that
>> >> > the
>> >> > >>> only
>> >> > >>> >> > threads that operate on the queues (outside of
>> housekeeping
>> >> tasks
>> >> > >>> like
>> >> > >>> >> > expiry) will be the IO threads.
>> >> > >>> >> >
>> >> > >>> >> > So, what we could do (and I have a patch
>> sitting on
>> my laptop
>> >> for
>> >> > >>> this)
>> >> > >>> >> is
>> >> > >>> >> > to look at using the "multi queue consumers"
>> work I
>> did for you
>> >> > guys
>> >> > >>> >> > before, but augmenting this so that the
>> consumers
>> work using a
>> >> > >>> "pull"
>> >> > >>> >> model
>> >> > >>> >> > rather than the push model. This will guarantee
>> strict fairness
>> >> > >>> between
>> >> > >>> >> > the queues associated with the consumer
>> (which was
>> the issue you
>> >> > had
>> >> > >>> >> with
>> >> > >>> >> > this functionality before, I believe).
>> Using this
>> model you'd
>> >> > only
>> >> > >>> >> need a
>> >> > >>> >> > small number (one?) of consumers per
>> session. The
>> patch I have
>> >> is
>> >> > >>> to
>> >> > >>> >> add
>> >> > >>> >> > this "pull" mode for these consumers
>> (essentially
>> this is a
>> >> > preview
>> >> > >>> of
>> >> > >>> >> how
>> >> > >>> >> > all consumers will work in the future).
>> >> > >>> >> >
>> >> > >>> >> > Does this seem like something you would be
>> interested in
>> >> pursuing?
>> >> > >>> >> >
>> >> > >>> >> > Cheers,
>> >> > >>> >> > Rob
>> >> > >>> >> >
>> >> > >>> >> > On 15 October 2016 at 17:30, Ramayan Tiwari <
>> >> > >>> ramayan.tiwari@gmail.com
>> <ma...@gmail.com>
>> <mailto:ramayan.tiwari@gmail.com
>>
>> <ma...@gmail.com>>>
>> >> > >>> >> > wrote:
>> >> > >>> >> >
>> >> > >>> >> > > Thanks Rob. Apologies for sending this
>> over weekend :(
>> >> > >>> >> > >
>> >> > >>> >> > > Are there are docs on the new threading
>> model? I
>> found this on
>> >> > >>> >> > confluence:
>> >> > >>> >> > >
>> >> > >>> >> > >
>> https://cwiki.apache.org/confluence/display/qpid/IO+
>> <https://cwiki.apache.org/confluence/display/qpid/IO+>
>> <https://cwiki.apache.org/confluence/display/qpid/IO+
>> <https://cwiki.apache.org/confluence/display/qpid/IO+>>
>> >> > >>> >> > Transport+Refactoring
>> >> > >>> >> > >
>> >> > >>> >> > > We are also interested in understanding the
>> threading model a
>> >> > >>> little
>> >> > >>> >> > better
>> >> > >>> >> > > to help us figure our its impact for our usage
>> patterns. Would
>> >> > be
>> >> > >>> very
>> >> > >>> >> > > helpful if there are more
>> docs/JIRA/email-threads
>> with some
>> >> > >>> details.
>> >> > >>> >> > >
>> >> > >>> >> > > Thanks
>> >> > >>> >> > >
>> >> > >>> >> > > On Sat, Oct 15, 2016 at 9:21 AM, Rob Godfrey <
>> >> > >>> rob.j.godfrey@gmail.com
>> <ma...@gmail.com>
>> <mailto:rob.j.godfrey@gmail.com <ma...@gmail.com>>
>>
>>
>> >> > >>> >> >
>> >> > >>> >> > > wrote:
>> >> > >>> >> > >
>> >> > >>> >> > > > So I *think* this is an issue because of the
>> extremely large
>> >> > >>> number
>> >> > >>> >> of
>> >> > >>> >> > > > consumers. The threading model in v6
>> means that
>> whenever a
>> >> > >>> network
>> >> > >>> >> > read
>> >> > >>> >> > > > occurs for a connection, it iterates
>> over the
>> consumers on
>> >> > that
>> >> > >>> >> > > connection
>> >> > >>> >> > > > - obviously where there are a large
>> number of
>> consumers this
>> >> > is
>> >> > >>> >> > > > burdensome. I fear addressing this may
>> not be a
>> trivial
>> >> > >>> change...
>> >> > >>> >> I
>> >> > >>> >> > > shall
>> >> > >>> >> > > > spend the rest of my afternoon pondering
>> this...
>> >> > >>> >> > > >
>> >> > >>> >> > > > - Rob
>> >> > >>> >> > > >
>> >> > >>> >> > > > On 15 October 2016 at 17:14, Ramayan
>> Tiwari <
>> >> > >>> >> ramayan.tiwari@gmail.com
>> <ma...@gmail.com>
>> <mailto:ramayan.tiwari@gmail.com
>> <ma...@gmail.com>>>
>> >> > >>> >> > > > wrote:
>> >> > >>> >> > > >
>> >> > >>> >> > > > > Hi Rob,
>> >> > >>> >> > > > >
>> >> > >>> >> > > > > Thanks so much for your response. We use
>> transacted
>> >> sessions
>> >> > >>> with
>> >> > >>> >> > > > > non-persistent delivery. Prefetch size
>> is 1
>> and every
>> >> > message
>> >> > >>> is
>> >> > >>> >> same
>> >> > >>> >> > > > size
>> >> > >>> >> > > > > (200 bytes).
>> >> > >>> >> > > > >
>> >> > >>> >> > > > > Thanks
>> >> > >>> >> > > > > Ramayan
>> >> > >>> >> > > > >
>> >> > >>> >> > > > > On Sat, Oct 15, 2016 at 2:59 AM, Rob
>> Godfrey <
>> >> > >>> >> > rob.j.godfrey@gmail.com
>> <ma...@gmail.com>
>> <mailto:rob.j.godfrey@gmail.com
>>
>> <ma...@gmail.com>>>
>>
>> >> > >>> >> > > > > wrote:
>> >> > >>> >> > > > >
>> >> > >>> >> > > > > > Hi Ramyan,
>> >> > >>> >> > > > > >
>> >> > >>> >> > > > > > this is interesting... in our
>> testing (which
>> admittedly
>> >> > >>> didn't
>> >> > >>> >> > cover
>> >> > >>> >> > > > the
>> >> > >>> >> > > > > > case of this many queues /
>> listeners) we saw
>> the 6.0.x
>> >> > >>> broker
>> >> > >>> >> using
>> >> > >>> >> > > > less
>> >> > >>> >> > > > > > CPU on average than the 0.32
>> broker. I'll
>> have a look
>> >> > this
>> >> > >>> >> weekend
>> >> > >>> >> > > as
>> >> > >>> >> > > > to
>> >> > >>> >> > > > > > why creating the listeners is
>> slower. On
>> the dequeing,
>> >> > can
>> >> > >>> you
>> >> > >>> >> > give
>> >> > >>> >> > > a
>> >> > >>> >> > > > > > little more information on the usage
>> pattern
>> - are you
>> >> > using
>> >> > >>> >> > > > > transactions,
>> >> > >>> >> > > > > > auto-ack or client ack? What
>> prefetch size
>> are you
>> >> using?
>> >> > >>> How
>> >> > >>> >> > large
>> >> > >>> >> > > > are
>> >> > >>> >> > > > > > your messages?
>> >> > >>> >> > > > > >
>> >> > >>> >> > > > > > Thanks,
>> >> > >>> >> > > > > > Rob
>> >> > >>> >> > > > > >
>> >> > >>> >> > > > > > On 14 October 2016 at 23:46, Ramayan
>> Tiwari <
>> >> > >>> >> > > ramayan.tiwari@gmail.com
>> <ma...@gmail.com>
>> <mailto:ramayan.tiwari@gmail.com
>>
>> <ma...@gmail.com>>>
>>
>> >> > >>> >> > > > > > wrote:
>> >> > >>> >> > > > > >
>> >> > >>> >> > > > > > > Hi All,
>> >> > >>> >> > > > > > >
>> >> > >>> >> > > > > > > We have been validating the new Qpid
>> broker (version
>> >> > >>> 6.0.4)
>> >> > >>> >> and
>> >> > >>> >> > > have
>> >> > >>> >> > > > > > > compared against broker version
>> 0.32 and
>> are seeing
>> >> > major
>> >> > >>> >> > > > regressions.
>> >> > >>> >> > > > > > > Following is the summary of our
>> test setup and
>> >> results:
>> >> > >>> >> > > > > > >
>> >> > >>> >> > > > > > > *1. Test Setup *
>> >> > >>> >> > > > > > > *a). *Qpid broker runs on a
>> dedicated
>> host (12
>> >> cores,
>> >> > >>> 32 GB
>> >> > >>> >> > RAM).
>> >> > >>> >> > > > > > > *b).* For 0.32, we allocated 16
>> GB heap.
>> For 6.0.6
>> >> > >>> broker,
>> >> > >>> >> we
>> >> > >>> >> > use
>> >> > >>> >> > > > 8GB
>> >> > >>> >> > > > > > > heap and 8GB direct memory.
>> >> > >>> >> > > > > > > *c).* For 6.0.4, flow to disk
>> has been
>> configured at
>> >> > >>> 60%.
>> >> > >>> >> > > > > > > *d).* Both the brokers use BDB
>> host type.
>> >> > >>> >> > > > > > > *e).* Brokers have around 6000
>> queues
>> and we create
>> >> 16
>> >> > >>> >> listener
>> >> > >>> >> > > > > > > sessions/threads spread over 3
>> connections, where each
>> >> > >>> >> session is
>> >> > >>> >> > > > > > listening
>> >> > >>> >> > > > > > > to 3000 queues. However, messages
>> are only
>> enqueued
>> >> and
>> >> > >>> >> processed
>> >> > >>> >> > > > from
>> >> > >>> >> > > > > 10
>> >> > >>> >> > > > > > > queues.
>> >> > >>> >> > > > > > > *f).* We enqueue 1 million messages
>> across 10
>> >> > different
>> >> > >>> >> queues
>> >> > >>> >> > > > > (evenly
>> >> > >>> >> > > > > > > divided), at the start of the test.
>> Dequeue only
>> >> starts
>> >> > >>> once
>> >> > >>> >> all
>> >> > >>> >> > > the
>> >> > >>> >> > > > > > > messages have been enqueued. We
>> run the
>> test for 2
>> >> hours
>> >> > >>> and
>> >> > >>> >> > > process
>> >> > >>> >> > > > as
>> >> > >>> >> > > > > > > many messages as we can. Each
>> message runs
>> for around
>> >> > 200
>> >> > >>> >> > > > milliseconds.
>> >> > >>> >> > > > > > > *g).* We have used both 0.16 and
>> 6.0.4
>> clients for
>> >> > these
>> >> > >>> >> tests
>> >> > >>> >> > > > (6.0.4
>> >> > >>> >> > > > > > > client only with 6.0.4 broker)
>> >> > >>> >> > > > > > >
>> >> > >>> >> > > > > > > *2. Test Results *
>> >> > >>> >> > > > > > > *a).* System Load Average (read
>> notes
>> below on how
>> >> we
>> >> > >>> >> compute
>> >> > >>> >> > > it),
>> >> > >>> >> > > > > for
>> >> > >>> >> > > > > > > 6.0.4 broker is 5x compared to 0.32
>> broker. During
>> >> start
>> >> > >>> of
>> >> > >>> >> the
>> >> > >>> >> > > test
>> >> > >>> >> > > > > > (when
>> >> > >>> >> > > > > > > we are not doing any dequeue), load
>> average is normal
>> >> > >>> (0.05
>> >> > >>> >> for
>> >> > >>> >> > > 0.32
>> >> > >>> >> > > > > > broker
>> >> > >>> >> > > > > > > and 0.1 for new broker), however,
>> while we are
>> >> dequeuing
>> >> > >>> >> > messages,
>> >> > >>> >> > > > the
>> >> > >>> >> > > > > > load
>> >> > >>> >> > > > > > > average is very high (around 0.5
>> consistently).
>> >> > >>> >> > > > > > >
>> >> > >>> >> > > > > > > *b). *Time to create listeners
>> in new
>> broker has
>> >> gone
>> >> > >>> up by
>> >> > >>> >> > 220%
>> >> > >>> >> > > > > > compared
>> >> > >>> >> > > > > > > to 0.32 broker (when using 0.16
>> client).
>> For old
>> >> broker,
>> >> > >>> >> creating
>> >> > >>> >> > > 16
>> >> > >>> >> > > > > > > sessions each listening to 3000 queues
>> takes 142
>> >> seconds
>> >> > >>> and
>> >> > >>> >> in
>> >> > >>> >> > new
>> >> > >>> >> > > > > > broker
>> >> > >>> >> > > > > > > it took 456 seconds. If we use 6.0.4
>> client, it took
>> >> > even
>> >> > >>> >> longer
>> >> > >>> >> > at
>> >> > >>> >> > > > > 524%
>> >> > >>> >> > > > > > > increase (887 seconds).
>> >> > >>> >> > > > > > > *I).* The time to create
>> consumers
>> increases as
>> >> we
>> >> > >>> create
>> >> > >>> >> > more
>> >> > >>> >> > > > > > > listeners on the same connections.
>> We have
>> 20 sessions
>> >> > >>> (but
>> >> > >>> >> end
>> >> > >>> >> > up
>> >> > >>> >> > > > > using
>> >> > >>> >> > > > > > > around 5 of them) on each
>> connection and
>> we create
>> >> about
>> >> > >>> 3000
>> >> > >>> >> > > > consumers
>> >> > >>> >> > > > > > and
>> >> > >>> >> > > > > > > attach MessageListener to it. Each
>> successive session
>> >> > >>> takes
>> >> > >>> >> > longer
>> >> > >>> >> > > > > > > (approximately linear increase) to
>> setup
>> same number
>> >> of
>> >> > >>> >> consumers
>> >> > >>> >> > > and
>> >> > >>> >> > > > > > > listeners.
>> >> > >>> >> > > > > > >
>> >> > >>> >> > > > > > > *3). How we compute System Load
>> Average *
>> >> > >>> >> > > > > > > We query the Mbean
>> SysetmLoadAverage and
>> divide it by
>> >> > the
>> >> > >>> >> value
>> >> > >>> >> > of
>> >> > >>> >> > > > > MBean
>> >> > >>> >> > > > > > > AvailableProcessors. Both of these
>> MBeans are
>> >> available
>> >> > >>> under
>> >> > >>> >> > > > > > > java.lang.OperatingSystem.
>> >> > >>> >> > > > > > >
>> >> > >>> >> > > > > > > I am not sure what is causing these
>> regressions and
>> >> > would
>> >> > >>> like
>> >> > >>> >> > your
>> >> > >>> >> > > > > help
>> >> > >>> >> > > > > > in
>> >> > >>> >> > > > > > > understanding it. We are aware
>> about the
>> changes with
>> >> > >>> respect
>> >> > >>> >> to
>> >> > >>> >> > > > > > threading
>> >> > >>> >> > > > > > > model in the new broker, are there any
>> design docs
>> >> that
>> >> > >>> we can
>> >> > >>> >> > > refer
>> >> > >>> >> > > > to
>> >> > >>> >> > > > > > > understand these changes at a high
>> level?
>> Can we tune
>> >> > some
>> >> > >>> >> > > parameters
>> >> > >>> >> > > > > to
>> >> > >>> >> > > > > > > address these issues?
>> >> > >>> >> > > > > > >
>> >> > >>> >> > > > > > > Thanks
>> >> > >>> >> > > > > > > Ramayan
>> >> > >>> >> > > > > > >
>> >> > >>> >> > > > > >
>> >> > >>> >> > > > >
>> >> > >>> >> > > >
>> >> > >>> >> > >
>> >> > >>> >> >
>> >> > >>> >>
>> >> > >>> >
>> >> > >>> >
>> >> > >>>
>> >> > >>
>> >> > >>
>> >> > >
>> >> >
>> >>
>>
>> -----------------------------
>> ----------------------------------------
>> To unsubscribe, e-mail: users-unsubscribe@qpid.apache.org
>> <ma...@qpid.apache.org>
>> <mailto:users-unsubscribe@qpid.apache.org
>> <ma...@qpid.apache.org>>
>> For additional commands, e-mail:
>> users-help@qpid.apache.org <ma...@qpid.apache.org>
>> <mailto:users-help@qpid.apache.org
>> <ma...@qpid.apache.org>>
>>
>>
>>
>>
>> ------------------------------------------------------------
>> ---------
>> To unsubscribe, e-mail: users-unsubscribe@qpid.apache.org
>> <ma...@qpid.apache.org>
>> For additional commands, e-mail: users-help@qpid.apache.org
>> <ma...@qpid.apache.org>
>>
>>
>>
>> ---------------------------------------------------------------------
>> To unsubscribe, e-mail: users-unsubscribe@qpid.apache.org
>> <ma...@qpid.apache.org>
>> For additional commands, e-mail: users-help@qpid.apache.org
>> <ma...@qpid.apache.org>
>>
>>
>>
>>
>> ---------------------------------------------------------------------
>> To unsubscribe, e-mail: users-unsubscribe@qpid.apache.org
>> For additional commands, e-mail: users-help@qpid.apache.org
>>
>
>
>
> ---------------------------------------------------------------------
> To unsubscribe, e-mail: users-unsubscribe@qpid.apache.org
> For additional commands, e-mail: users-help@qpid.apache.org
>
>
Re: Qpid broker 6.0.4 performance issues
Posted by Rob Godfrey <ro...@gmail.com>.
Hi Ramayan,
On 4 January 2017 at 23:54, Ramayan Tiwari <ra...@gmail.com> wrote:
> Hi Lorenz,
>
> Happy new year to everyone, hope you guys had fun!
>
> I am doing performance test runs to figure out a reasonable threshold for
> direct memory, considering our use case of small message payloads. I have a
> few more questions:
>
> 1. I assume there is no way to disable direct memory (so that we always use
> heap to store messages), however, I was wondering if this would be
> something that can be offered as part of broker config (unless it a
> significant change and unfeasible to do it).
>
Unfortunately it's not that simple (as I guess you might have expected).
Moving to using direct memory wasn't something we did as an active
choice... what actually happened is that we discovered that when we moved
to java nio socket channels rather than the blocking io, that the JVM
libraries were actually taking our heap buffers and then allocating and
copying the contents into direct memory. Because of issues with regard to
how these (JVM library managed) direct byte buffers were being cached, we
saw much more frequent out of memory conditions, and from code that we
really had no way to control... So the decision to start using direct
memory was essentially forced upon us by our move to NIO. For the same
reason it doesn't really make sense to offer a "use heap memory" option, as
you'll still end up using direct memory (because of the Java library) but
in a much harder to control way.
>
> 2. We monitor Qpid broker memory and stop enqueue at a certain threshold to
> protect broker heap. Now that message payload is stored in DM, we would
> like to track that as well. However, DM size reported by Mbean doesn't
> correlate directly with the sum of message payload (possibly because the
> byte buffers are not allocated/de-allocated that frequently). So to get
> around that, we are considering using the sum queue sizes to get actual DM
> usage.
>
>
Yes - the way we use direct memory is to try to cache the allocated buffers
and re-use them rather than constantly creating and GCing them (because
there are issues with how the JVM does (or rather doesn't) GC direct
buffers). The broker's own flow control / flow-to-disk algorithms are
based around aggregating the sizes reported by the queues, so this is
certainly a viable approach. As you know we are deprecating the use of
JMX, but you do bring up an issue that we are not exposing (through any
management channel) a way of doing deep inspection in how direct memory is
being managed. Potentially we should be adding a mechanism to enquire how
much direct memory is allocated and in use vs. how much is in the cache,
but actually "free".
-- Rob
> Thanks
> Ramayan
>
> On Wed, Dec 21, 2016 at 5:17 AM, Lorenz Quack <qu...@gmail.com>
> wrote:
>
> > Hi,
> >
> > Regarding 0.32 behaviour, it checked to see whether to flow a message
> > to disk when putting a message on the Queue the same way Qpid 6 does.
> > In that sense 6 is not more or less aggressive. However, the
> > algorithm behind the decision whether or not to flow to disk has
> > changed. This change was done as part of a larger effort to isolate
> > VirtualHosts from each other. I would have to go back and check how
> > the algorithm worked previously but I would assume that it just
> > considered the total (estimated) amount of memory used and did no per
> > VH or per Queue allocation. This means 6 effectively lowers the
> > threshold on individual Queues especially in the case where some
> > VirtualHosts and/or Queues are used less than others. On the upside,
> > the broker is fairer in its resource management and a single
> > VirtualHost can no longer use up all available memory. How exactly
> > that trade-off between fairness and efficient use of available memory
> > is made is debatable but I don't think we want to go back to the pre-6
> > model of just lumping everything together.
> >
> > Given your numbers (1 VH, 6000 Qs) each Queue would initially be
> > allocated 1/6000th of 60% of 8 GB ≈ 1 MB. In then end state the full
> > Queues should end up with approximately 780 MB but as you noticed the
> > threshold is only recalculated periodically during housekeeping (by
> > default every 30 s) or when a VH or Queue is added or deleted. If you
> > have DEBUG logging you should see periodic messages like "Allocating
> > target size to queues [...]" if not then I am afraid you won't be
> > able to tell the current thresholds because they are only reported
> > once when flowToDisk becomes active/inactive.
> >
> > So I think your analysis is probably correct that the revision of the
> > threshold is always "behind" the publishing, raising it on every
> > revision but never far enough to prevent flowToDisk. This is not
> > ideal. We will have to address this. However, I am afraid that in
> > the current release there is no way to influence the algorithm other
> > than setting the available memory and broker.flowToDiskThreshold.
> >
> > Regarding the MemoryStore, the algorithm triggering flowToDisk is the
> > same for all stores, just the implementation of the actual writing
> > messages to disk differs. For the MemoryStore it is a noop, i.e., the
> > message is not flown to disk and remains in memory. Performancewise
> > we do not do a lot of testing with the MemoryStore because it is not a
> > typical use-case and mainly used for unit and system testing. I would
> > assume that the better distribution you are seeing is coincidental
> > since that part of the code should be relatively independent of the
> > store type. Unfortunately, I cannot see any of your graphs. I
> > believe the mailing list strips all attachments.
> >
> > Regarding a recommendation of how to configure your DM vs Heap I would
> > like to refer you to our documentation [1], especially section
> > "9.11.6. Memory Tuning the Broker". There we provide formulas to
> > estimate the memory consumption of the broker for both DM and Heap.
> > Note that these are estimates and you should test your chosen settings
> > under a typical peak workload. Given that your messages are small you
> > will probably want to favour Heap over DM but I am reluctant to make
> > an explicit recommendation.
> >
> > Kind regards,
> > Lorenz
> >
> > P.S.: I am going on a 2 day vacation later today but feel free to
> > continue this conversation with others on this list.
> >
> > [1] https://qpid.apache.org/releases/qpid-java-6.1.0/java-broker
> > /book/Java-Broker-Runtime-Memory.html
> >
> > On 20/12/16 20:37, Ramayan Tiwari wrote:
> >
> >> Hi Lorenz,
> >>
> >> Thanks a lot for your response and explaining the flow to disk algorithm
> >> in detail. I have described the test setup in detail in the first email
> of
> >> this thread, to summarize the points again:
> >> a) There is only one virtual host.
> >> b) There are 6000 queues in this virtual host, but messages are only
> >> enqueued to 10 queues.
> >> c) Every queue gets equal number of messages (100k) at the start of the
> >> test (we do not start dequeue till all the 1 million messages are
> enqueued).
> >> d) Heap and DM memory are equal (8GB each) and DM flow to disk threshold
> >> is 60%.
> >>
> >> I looked at QUE-1014/15 log lines and following is what I notice:
> >> a) These log lines are not present in 0.32 broker's log, which means
> that
> >> its not doing any flow to disk. Is flow to disk behavior different in
> the
> >> two brokers, it looks like 6.0.x is a lot more aggressive in this
> regard.
> >>
> >> b) Since all the 1 million messages are enqueued at the start of test
> >> (takes about 7 mins to enqueue), flow to disk threshold revisions
> performed
> >> by the housekeeping task are not able to catch up. Or the rate with
> which
> >> thresholds are revised can not catch up with the rate of enqueue. In my
> >> test, revisions once happened twice (4 seconds and 5 mins after test
> start)
> >> and then on, the threshold was not revised for the queues.
> >>
> >> To make sure that we are not getting penalized by writing to disk, I
> also
> >> did a test using Memory store type and compared the result with BDB
> store
> >> type. Apparently, BDB store is slightly more efficient (2.7%) in terms
> of
> >> number of messages delivered. Memory store also takes more broker CPU
> (3%
> >> more on average), but its better in terms of distributing messages in a
> >> round robin manner from all the queues. See the attached graphs for
> details.
> >>
> >> I do notice that flow to disk behavior is almost exactly same
> >> (QUE-1014/15 log lines are present) when running with Memory store. I am
> >> wondering what does flow to disk does when we use Memory store?
> >>
> >> Since our average messages size is less than 1KB, I am really looking
> >> forward to some recommendation around the % allocation for DM vs Heap.
> >>
> >> Thanks
> >> Ramayan
> >>
> >>
> >> On Tue, Dec 20, 2016 at 4:02 AM, Lorenz Quack <quack.lorenz@gmail.com
> >> <ma...@gmail.com>> wrote:
> >>
> >> Hello Ramayan,
> >>
> >> glad to hear that the patch is (mostly) working for you.
> >> To address your points:
> >>
> >> 1. If indeed in one case flow to disk is kicking in while in
> >> the other one it is not, then I am not surprised that
> >> there is a 5% difference. The question is whether the
> >> flow to disk is expected or not which leads to
> >>
> >> 2. The direct memory utilization not exceeding a certain
> >> value is a strong indication that flow to disk is active.
> >> Could you verify that by checking the logs (QUE-1014/15)?
> >> If the flow to disk limit is exceeded then it is expected
> >> that 2 million messages consume the same amount of direct
> >> memory as 1 million messages. Could you share a little
> >> more about the test setup? How many VirtualHost are
> >> running on the broker? How many Queues are on each
> >> VirtualHost? What is the Queue depth of those Queues?
> >> All of those factors influence the actual flow to disk
> >> threshold. This is to ensure some fairness between
> >> VirtualHosts as far as memory consumption is concerned.
> >> Below I explain how threshold allocation is currently
> >> performed. We are considering changing the algorithm in
> >> the future or making it tunable. Your ideas, requirements,
> >> and input on this would certainly be of interest to us.
> >>
> >> Looking forward to hearing from you.
> >>
> >> Kind regards,
> >> Lorenz
> >>
> >>
> >> Algorithm for flow to disk threshold:
> >>
> >> 1. Take the total amount of the broker.flowToDiskThreshold and
> >> divide it amongst all active VirtualHosts as follows
> >>
> >> a. Half of broker.flowToDiskThreshold is evenly devided
> >> amongst the VHs to ensure a minimum amount is available to
> >> each VH.
> >>
> >> b. The remaining half is allocated proportional to the current
> >> usage pattern. For example, if VH1 is currently using 3
> >> MB, VH2 is using 1 MB and VH3 is using 0 MB, then of the
> >> remaining half 3/4 will be allocated to VH1, 1/4 to VH2,
> >> and nothing to VH3. If all VHs are empty distribute this
> >> half evenly like in 1.a.
> >>
> >> 2. The VirtualHosts allocate their available memory to their
> >> Queues in a proportional fashion as explained above (1.b).
> >>
> >>
> >> Example:
> >>
> >> * The broker.flowToDiskThreshold is set to 10 GB.
> >>
> >> * Two Virtual Hosts with 10 Queues each.
> >>
> >> * VH1 all 10 Queues are empty.
> >>
> >> * VH2 all Queues contain 10 MB except of one Queue that
> >> contains 100 MB.
> >>
> >> * According to 1.a each VirtualHost is allocated half of 5 GB,
> >> i.e., 2.5 GB
> >>
> >> * According to 1.b VH1 using 0MB does not get any additional
> >> memory while VH2 gets the full of the remainder of the 5 GB
> >> totaling 7.5 GB.
> >>
> >> * The Queues on VH1 don't have messages on them so the
> >> VirtualHost falls back to allocating them equal shares: 250 MB
> >> each.
> >>
> >> * On VH2 the total current memory usage is 9*10 MB + 100 MB =
> >> 190 MB so the smaller Queues receive 10/190 * 7.5 GB = 395 MB
> >> while the large Queue receives 100/190 * 7.5 GB = 3950 MB.
> >>
> >> * In total we allocated 10 * 250 MB + 9 * 395 MB + 1 * 3950 MB
> >> totaling 10 GB (within bounds of rounding errors).
> >>
> >>
> >>
> >> On 19/12/16 20:48, Ramayan Tiwari wrote:
> >>
> >> Hi Rob,
> >>
> >> I did another exhaustive performance test using the
> >> MultiQueueConsumer feature with 6.0.5 (and the patch). The
> >> broker CPU issues has been resolved and we no longer have the
> >> problem message prefetch (caused by long running message).
> >>
> >> Fairness among queue is also great (not as perfect as 0.32
> >> broker though, see attached graphs). Everything looks great,
> >> except for:
> >>
> >> 1. 6.0.5 delivered around 4.6% less messages. Flow to disk
> >> triggered aggressively in 6.0.5 but I don't see any flow to
> >> disk happening in 0.32 (looking for QUE-1014). This might be
> >> the reason for lesser message delivery.
> >>
> >> 2. Direct memory utilization in the new broker does not make
> >> sense to us. We did 2 tests: 1 millions and 2 million messages
> >> (220 Byte average message size), however, the direct memory
> >> utilization never exceeded 500MB (see attached graph), even
> >> when we are allocating 8GB for direct memory. Because there is
> >> a 1KB heap overhead with each message, heap utilization looks
> >> same for both 0.32 and 6.0.5. For our setup, this essentially
> >> means that, we are cutting our memory capacity by half,
> >> because now are allocating half of the available RAM to direct
> >> memory, but will be limited by heap anyway.
> >>
> >> These tests were performed using 16GB RAM, where 8GB was
> >> allocated to heap and 8GB for Direct memory. I also changed
> >> flowToDiskThreshold to 60%. This is one of our biggest concern
> >> with the new broker, since our average message size in
> >> production is less than 1KB. Currently we allocate all the
> >> available RAM to heap, which will be reduced in half with the
> >> new broker.
> >>
> >> What is the recommendation for memory allocation (heap vs dm)
> >> in our use case?
> >>
> >> Thanks
> >> Ramaayn
> >>
> >> On Fri, Oct 28, 2016 at 5:37 AM, Keith W <keith.wall@gmail.com
> >> <ma...@gmail.com> <mailto:keith.wall@gmail.com
> >>
> >> <ma...@gmail.com>>> wrote:
> >>
> >> Hi Ramayan
> >>
> >> QPID-7462 is a new (experimental) feature, so we don't
> >> consider this
> >> appropriate for inclusion in the 6.0.5 defect release We
> >> follow a
> >> Semantic Versioning[1] strategy.
> >>
> >> The underlying issue is your testing has uncovered is poor
> >> performance
> >> with large numbers of consumers. QPID-7462 effectively
> >> side steps the
> >> problem (by introducing alternative consumer behaviour)
> >> but does not
> >> address the root cause. We continue to consider how best
> >> to resolve
> >> the problem completely, but don't yet have timelines for
> >> this change.
> >> It is something that will be getting attention in what
> >> remains of this
> >> year. We will keep you posted.
> >>
> >> In the meanwhile, I understand this causes you a problem.
> >> If you
> >> cannot adopt 6.1 (there should be another RC out soon),
> >> you could
> >> consider applying the patch (attached to the JIRA) to
> >> 6.0.x branch and
> >> building yourself.
> >>
> >> Kind regards, Keith.
> >>
> >>
> >> [1] http://semver.org
> >>
> >>
> >> On 27 October 2016 at 23:19, Ramayan Tiwari
> >> <ramayan.tiwari@gmail.com
> >> <ma...@gmail.com>
> >> <mailto:ramayan.tiwari@gmail.com
> >> <ma...@gmail.com>>> wrote:
> >> > Hi Rob,
> >> >
> >> > I have the truck code which I am testing with, I haven't
> >> finished the test
> >> > runs yet. I was hoping that once I validate the change,
> >> I can simply
> >> > release 6.0.5.
> >> >
> >> > Thanks
> >> > Ramayan
> >> >
> >> > On Thu, Oct 27, 2016 at 12:41 PM, Rob Godfrey
> >> <rob.j.godfrey@gmail.com <ma...@gmail.com>
> >> <mailto:rob.j.godfrey@gmail.com <mailto:rob.j.godfrey@gmail.com
> >> >>>
> >> > wrote:
> >> >
> >> >> Hi Ramayan,
> >> >>
> >> >> did you verify that the change works for you? You said
> >> you were
> >> going to
> >> >> test with the trunk code...
> >> >>
> >> >> I'll discuss with the other developers tomorrow about
> >> whether
> >> we can put
> >> >> this change into 6.0.5.
> >> >>
> >> >> Cheers,
> >> >> Rob
> >> >>
> >> >> On 27 October 2016 at 20:30, Ramayan Tiwari
> >> <ramayan.tiwari@gmail.com
> >> <ma...@gmail.com>
> >> <mailto:ramayan.tiwari@gmail.com
> >> <ma...@gmail.com>>>
> >> >> wrote:
> >> >>
> >> >> > Hi Rob,
> >> >> >
> >> >> > I looked at the release notes for 6.0.5 and it doesn't
> >> include the fix
> >> >> for
> >> >> > large consumers issues [1]. The fix is marked for
> >> 6.1, which
> >> will not
> >> >> have
> >> >> > JMX and for us to use this version requires major
> >> changes in our
> >> >> monitoring
> >> >> > framework. Could you please include the fix in 6.0.5
> >> release?
> >> >> >
> >> >> > Thanks
> >> >> > Ramayan
> >> >> >
> >> >> > [1]. https://issues.apache.org/jira/browse/QPID-7462
> >> <https://issues.apache.org/jira/browse/QPID-7462>
> >> <https://issues.apache.org/jira/browse/QPID-7462
> >> <https://issues.apache.org/jira/browse/QPID-7462>>
> >> >> >
> >> >> > On Wed, Oct 19, 2016 at 4:49 PM, Helen Kwong
> >> <helenkwong@gmail.com <ma...@gmail.com>
> >> <mailto:helenkwong@gmail.com <ma...@gmail.com>>>
> >> >> wrote:
> >> >> >
> >> >> > > Hi Rob,
> >> >> > >
> >> >> > > Again, thank you so much for answering our
> >> questions and
> >> providing a
> >> >> > patch
> >> >> > > so quickly :) One more question I have: would it be
> >> possible to include
> >> >> > > test cases involving many queues and listeners (in
> >> the order of
> >> >> thousands
> >> >> > > of queues) for future Qpid releases, as part of
> >> standard
> >> perf testing
> >> >> of
> >> >> > > the broker?
> >> >> > >
> >> >> > > Thanks,
> >> >> > > Helen
> >> >> > >
> >> >> > > On Tue, Oct 18, 2016 at 10:40 AM, Ramayan Tiwari <
> >> >> > ramayan.tiwari@gmail.com
> >> <ma...@gmail.com>
> >> <mailto:ramayan.tiwari@gmail.com
> >> <ma...@gmail.com>>
> >> >> > > > wrote:
> >> >> > >
> >> >> > >> Thanks so much Rob, I will test the patch against
> >> trunk
> >> and will
> >> >> update
> >> >> > >> you with the outcome.
> >> >> > >>
> >> >> > >> - Ramayan
> >> >> > >>
> >> >> > >> On Tue, Oct 18, 2016 at 2:37 AM, Rob Godfrey
> >> <rob.j.godfrey@gmail.com <ma...@gmail.com>
> >> <mailto:rob.j.godfrey@gmail.com <mailto:rob.j.godfrey@gmail.com
> >>
> >> >> >
> >> >> > >> wrote:
> >> >> > >>
> >> >> > >>> On 17 October 2016 at 21:50, Rob Godfrey
> >> <rob.j.godfrey@gmail.com <ma...@gmail.com>
> >> <mailto:rob.j.godfrey@gmail.com <mailto:rob.j.godfrey@gmail.com
> >> >>>
> >> >> > >>> wrote:
> >> >> > >>>
> >> >> > >>> >
> >> >> > >>> >
> >> >> > >>> > On 17 October 2016 at 21:24, Ramayan Tiwari <
> >> >> > ramayan.tiwari@gmail.com
> >> <ma...@gmail.com>
> >> <mailto:ramayan.tiwari@gmail.com
> >>
> >> <ma...@gmail.com>>>
> >>
> >> >> > >>> > wrote:
> >> >> > >>> >
> >> >> > >>> >> Hi Rob,
> >> >> > >>> >>
> >> >> > >>> >> We are certainly interested in testing the
> "multi
> >> queue consumers"
> >> >> > >>> >> behavior
> >> >> > >>> >> with your patch in the new broker. We would
> >> like to know:
> >> >> > >>> >>
> >> >> > >>> >> 1. What will the scope of changes, client or
> >> broker or
> >> both? We
> >> >> are
> >> >> > >>> >> currently running 0.16 client, so would like
> >> to make
> >> sure that we
> >> >> > will
> >> >> > >>> >> able
> >> >> > >>> >> to use these changes with 0.16 client.
> >> >> > >>> >>
> >> >> > >>> >>
> >> >> > >>> > There's no change to the client. I can't
> >> remember what
> >> was in the
> >> >> > 0.16
> >> >> > >>> > client... the only issue would be if there are
> >> any bugs
> >> in the
> >> >> > parsing
> >> >> > >>> of
> >> >> > >>> > address arguments. I can try to test that out
> tmr.
> >> >> > >>> >
> >> >> > >>>
> >> >> > >>>
> >> >> > >>> OK - with a little bit of care to get round the
> >> address
> >> parsing
> >> >> issues
> >> >> > in
> >> >> > >>> the 0.16 client... I think we can get this to
> >> work. I've
> >> created the
> >> >> > >>> following JIRA:
> >> >> > >>>
> >> >> > >>> https://issues.apache.org/jira/browse/QPID-7462
> >> <https://issues.apache.org/jira/browse/QPID-7462>
> >> <https://issues.apache.org/jira/browse/QPID-7462
> >> <https://issues.apache.org/jira/browse/QPID-7462>>
> >> >> > >>>
> >> >> > >>> and attached to it are a patch which applies
> against
> >> trunk, and a
> >> >> > >>> separate
> >> >> > >>> patch which applies against the 6.0.x branch (
> >> >> > >>>
> >> https://svn.apache.org/repos/asf/qpid/java/branches/6.0.x
> >> <https://svn.apache.org/repos/asf/qpid/java/branches/6.0.x>
> >> <https://svn.apache.org/repos/asf/qpid/java/branches/6.0.x
> >> <https://svn.apache.org/repos/asf/qpid/java/branches/6.0.x>> -
> >> this is
> >> >> > >>> 6.0.4
> >> >> > >>> plus a few other fixes which we will soon be
> >> releasing as
> >> 6.0.5)
> >> >> > >>>
> >> >> > >>> To create a consumer which uses this feature (and
> >> multi queue
> >> >> > >>> consumption)
> >> >> > >>> for the 0.16 client you need to use something
> >> like the
> >> following as
> >> >> the
> >> >> > >>> address:
> >> >> > >>>
> >> >> > >>> queue_01 ; {node : { type : queue }, link : {
> >> x-subscribes : {
> >> >> > >>> arguments : { x-multiqueue : [ queue_01, queue_02,
> >> queue_03 ],
> >> >> > >>> x-pull-only : true }}}}
> >> >> > >>>
> >> >> > >>>
> >> >> > >>> Note that the initial queue_01 has to be a name of
> an
> >> actual queue on
> >> >> > >>> the virtual host, but otherwise it is not
> >> actually used
> >> (if you were
> >> >> > >>> using a 0.32 or later client you could just use ''
> >> here). The actual
> >> >> > >>> queues that are consumed from are in the list value
> >> associated with
> >> >> > >>> x-multiqueue. For my testing I created a list
> >> with 3000
> >> queues here
> >> >> > >>> and this worked fine.
> >> >> > >>>
> >> >> > >>> Let me know if you have any questions / issues,
> >> >> > >>>
> >> >> > >>> Hope this helps,
> >> >> > >>> Rob
> >> >> > >>>
> >> >> > >>>
> >> >> > >>> >
> >> >> > >>> >
> >> >> > >>> >> 2. My understanding is that the "pull vs push"
> >> change
> >> is only with
> >> >> > >>> respect
> >> >> > >>> >> to broker and it does not change our
> architecture
> >> where we use
> >> >> > >>> >> MessageListerner to receive messages
> >> asynchronously.
> >> >> > >>> >>
> >> >> > >>> >
> >> >> > >>> > Exactly - this is only a change within the
> >> internal broker
> >> >> threading
> >> >> > >>> > model. The external behaviour of the broker
> >> remains
> >> essentially
> >> >> > >>> unchanged.
> >> >> > >>> >
> >> >> > >>> >
> >> >> > >>> >>
> >> >> > >>> >> 3. Once I/O refactoring is completely, we would
> be
> >> able to go back
> >> >> > to
> >> >> > >>> use
> >> >> > >>> >> standard JMS consumer (Destination), what is the
> >> timeline and
> >> >> broker
> >> >> > >>> >> release version for the completion of this work?
> >> >> > >>> >>
> >> >> > >>> >
> >> >> > >>> > You might wish to continue to use the "multi
> >> queue" model,
> >> >> depending
> >> >> > on
> >> >> > >>> > your actual use case, but yeah once the I/O work
> is
> >> complete I
> >> >> would
> >> >> > >>> hope
> >> >> > >>> > that you could use the thousands of consumers
> model
> >> should you
> >> >> wish.
> >> >> > >>> We
> >> >> > >>> > don't have a schedule for the next phase of I/O
> >> rework
> >> right now -
> >> >> > >>> about
> >> >> > >>> > all I can say is that it is unlikely to be
> complete
> >> this year. I'd
> >> >> > >>> need to
> >> >> > >>> > talk with Keith (who is currently on vacation)
> >> as to
> >> when we think
> >> >> we
> >> >> > >>> may
> >> >> > >>> > be able to schedule it.
> >> >> > >>> >
> >> >> > >>> >
> >> >> > >>> >>
> >> >> > >>> >> Let me know once you have integrated the patch
> >> and I
> >> will re-run
> >> >> our
> >> >> > >>> >> performance tests to validate it.
> >> >> > >>> >>
> >> >> > >>> >>
> >> >> > >>> > I'll make a patch for 6.0.x presently (I've been
> >> working on a
> >> >> change
> >> >> > >>> > against trunk - the patch will probably have to
> >> change
> >> a bit to
> >> >> apply
> >> >> > >>> to
> >> >> > >>> > 6.0.x).
> >> >> > >>> >
> >> >> > >>> > Cheers,
> >> >> > >>> > Rob
> >> >> > >>> >
> >> >> > >>> > Thanks
> >> >> > >>> >> Ramayan
> >> >> > >>> >>
> >> >> > >>> >> On Sun, Oct 16, 2016 at 3:30 PM, Rob Godfrey <
> >> >> > rob.j.godfrey@gmail.com
> >> <ma...@gmail.com>
> >> <mailto:rob.j.godfrey@gmail.com <mailto:rob.j.godfrey@gmail.com
> >>
> >>
> >>
> >> >> > >>> >
> >> >> > >>> >> wrote:
> >> >> > >>> >>
> >> >> > >>> >> > OK - so having pondered / hacked around a
> >> bit this
> >> weekend, I
> >> >> > think
> >> >> > >>> to
> >> >> > >>> >> get
> >> >> > >>> >> > decent performance from the IO model in 6.0
> >> for your
> >> use case
> >> >> > we're
> >> >> > >>> >> going
> >> >> > >>> >> > to have to change things around a bit.
> >> >> > >>> >> >
> >> >> > >>> >> > Basically 6.0 is an intermediate step on our
> >> IO /
> >> threading
> >> >> model
> >> >> > >>> >> journey.
> >> >> > >>> >> > In earlier versions we used 2 threads per
> >> connection
> >> for IO (one
> >> >> > >>> read,
> >> >> > >>> >> one
> >> >> > >>> >> > write) and then extra threads from a pool to
> >> "push"
> >> messages
> >> >> from
> >> >> > >>> >> queues to
> >> >> > >>> >> > connections.
> >> >> > >>> >> >
> >> >> > >>> >> > In 6.0 we move to using a pool for the IO
> >> threads,
> >> and also
> >> >> > stopped
> >> >> > >>> >> queues
> >> >> > >>> >> > from "pushing" to connections while the IO
> >> threads
> >> were acting
> >> >> on
> >> >> > >>> the
> >> >> > >>> >> > connection. It's this latter fact which is
> >> screwing up
> >> >> > performance
> >> >> > >>> for
> >> >> > >>> >> > your use case here because what happens is
> >> that on
> >> each network
> >> >> > >>> read we
> >> >> > >>> >> > tell each consumer to stop accepting pushes
> >> from the
> >> queue until
> >> >> > >>> the IO
> >> >> > >>> >> > interaction has completed. This is causing
> >> lots of
> >> loops over
> >> >> > your
> >> >> > >>> 3000
> >> >> > >>> >> > consumers on each session, which is eating
> >> up a lot
> >> of CPU on
> >> >> > every
> >> >> > >>> >> network
> >> >> > >>> >> > interaction.
> >> >> > >>> >> >
> >> >> > >>> >> > In the final version of our IO refactoring
> >> we want
> >> to remove the
> >> >> > >>> >> "pushing"
> >> >> > >>> >> > from the queue, and instead have the consumers
> >> "pull" - so that
> >> >> > the
> >> >> > >>> only
> >> >> > >>> >> > threads that operate on the queues (outside of
> >> housekeeping
> >> >> tasks
> >> >> > >>> like
> >> >> > >>> >> > expiry) will be the IO threads.
> >> >> > >>> >> >
> >> >> > >>> >> > So, what we could do (and I have a patch
> >> sitting on
> >> my laptop
> >> >> for
> >> >> > >>> this)
> >> >> > >>> >> is
> >> >> > >>> >> > to look at using the "multi queue consumers"
> >> work I
> >> did for you
> >> >> > guys
> >> >> > >>> >> > before, but augmenting this so that the
> >> consumers
> >> work using a
> >> >> > >>> "pull"
> >> >> > >>> >> model
> >> >> > >>> >> > rather than the push model. This will
> guarantee
> >> strict fairness
> >> >> > >>> between
> >> >> > >>> >> > the queues associated with the consumer
> >> (which was
> >> the issue you
> >> >> > had
> >> >> > >>> >> with
> >> >> > >>> >> > this functionality before, I believe).
> >> Using this
> >> model you'd
> >> >> > only
> >> >> > >>> >> need a
> >> >> > >>> >> > small number (one?) of consumers per
> >> session. The
> >> patch I have
> >> >> is
> >> >> > >>> to
> >> >> > >>> >> add
> >> >> > >>> >> > this "pull" mode for these consumers
> >> (essentially
> >> this is a
> >> >> > preview
> >> >> > >>> of
> >> >> > >>> >> how
> >> >> > >>> >> > all consumers will work in the future).
> >> >> > >>> >> >
> >> >> > >>> >> > Does this seem like something you would be
> >> interested in
> >> >> pursuing?
> >> >> > >>> >> >
> >> >> > >>> >> > Cheers,
> >> >> > >>> >> > Rob
> >> >> > >>> >> >
> >> >> > >>> >> > On 15 October 2016 at 17:30, Ramayan Tiwari <
> >> >> > >>> ramayan.tiwari@gmail.com
> >> <ma...@gmail.com>
> >> <mailto:ramayan.tiwari@gmail.com
> >>
> >> <ma...@gmail.com>>>
> >> >> > >>> >> > wrote:
> >> >> > >>> >> >
> >> >> > >>> >> > > Thanks Rob. Apologies for sending this
> >> over weekend :(
> >> >> > >>> >> > >
> >> >> > >>> >> > > Are there are docs on the new threading
> >> model? I
> >> found this on
> >> >> > >>> >> > confluence:
> >> >> > >>> >> > >
> >> >> > >>> >> > >
> >> https://cwiki.apache.org/confluence/display/qpid/IO+
> >> <https://cwiki.apache.org/confluence/display/qpid/IO+>
> >> <https://cwiki.apache.org/confluence/display/qpid/IO+
> >> <https://cwiki.apache.org/confluence/display/qpid/IO+>>
> >> >> > >>> >> > Transport+Refactoring
> >> >> > >>> >> > >
> >> >> > >>> >> > > We are also interested in understanding the
> >> threading model a
> >> >> > >>> little
> >> >> > >>> >> > better
> >> >> > >>> >> > > to help us figure our its impact for our
> usage
> >> patterns. Would
> >> >> > be
> >> >> > >>> very
> >> >> > >>> >> > > helpful if there are more
> >> docs/JIRA/email-threads
> >> with some
> >> >> > >>> details.
> >> >> > >>> >> > >
> >> >> > >>> >> > > Thanks
> >> >> > >>> >> > >
> >> >> > >>> >> > > On Sat, Oct 15, 2016 at 9:21 AM, Rob
> Godfrey <
> >> >> > >>> rob.j.godfrey@gmail.com
> >> <ma...@gmail.com>
> >> <mailto:rob.j.godfrey@gmail.com <mailto:rob.j.godfrey@gmail.com
> >>
> >>
> >>
> >> >> > >>> >> >
> >> >> > >>> >> > > wrote:
> >> >> > >>> >> > >
> >> >> > >>> >> > > > So I *think* this is an issue because of
> the
> >> extremely large
> >> >> > >>> number
> >> >> > >>> >> of
> >> >> > >>> >> > > > consumers. The threading model in v6
> >> means that
> >> whenever a
> >> >> > >>> network
> >> >> > >>> >> > read
> >> >> > >>> >> > > > occurs for a connection, it iterates
> >> over the
> >> consumers on
> >> >> > that
> >> >> > >>> >> > > connection
> >> >> > >>> >> > > > - obviously where there are a large
> >> number of
> >> consumers this
> >> >> > is
> >> >> > >>> >> > > > burdensome. I fear addressing this may
> >> not be a
> >> trivial
> >> >> > >>> change...
> >> >> > >>> >> I
> >> >> > >>> >> > > shall
> >> >> > >>> >> > > > spend the rest of my afternoon pondering
> >> this...
> >> >> > >>> >> > > >
> >> >> > >>> >> > > > - Rob
> >> >> > >>> >> > > >
> >> >> > >>> >> > > > On 15 October 2016 at 17:14, Ramayan
> >> Tiwari <
> >> >> > >>> >> ramayan.tiwari@gmail.com
> >> <ma...@gmail.com>
> >> <mailto:ramayan.tiwari@gmail.com
> >> <ma...@gmail.com>>>
> >> >> > >>> >> > > > wrote:
> >> >> > >>> >> > > >
> >> >> > >>> >> > > > > Hi Rob,
> >> >> > >>> >> > > > >
> >> >> > >>> >> > > > > Thanks so much for your response. We use
> >> transacted
> >> >> sessions
> >> >> > >>> with
> >> >> > >>> >> > > > > non-persistent delivery. Prefetch size
> >> is 1
> >> and every
> >> >> > message
> >> >> > >>> is
> >> >> > >>> >> same
> >> >> > >>> >> > > > size
> >> >> > >>> >> > > > > (200 bytes).
> >> >> > >>> >> > > > >
> >> >> > >>> >> > > > > Thanks
> >> >> > >>> >> > > > > Ramayan
> >> >> > >>> >> > > > >
> >> >> > >>> >> > > > > On Sat, Oct 15, 2016 at 2:59 AM, Rob
> >> Godfrey <
> >> >> > >>> >> > rob.j.godfrey@gmail.com
> >> <ma...@gmail.com>
> >> <mailto:rob.j.godfrey@gmail.com
> >>
> >> <ma...@gmail.com>>>
> >>
> >> >> > >>> >> > > > > wrote:
> >> >> > >>> >> > > > >
> >> >> > >>> >> > > > > > Hi Ramyan,
> >> >> > >>> >> > > > > >
> >> >> > >>> >> > > > > > this is interesting... in our
> >> testing (which
> >> admittedly
> >> >> > >>> didn't
> >> >> > >>> >> > cover
> >> >> > >>> >> > > > the
> >> >> > >>> >> > > > > > case of this many queues /
> >> listeners) we saw
> >> the 6.0.x
> >> >> > >>> broker
> >> >> > >>> >> using
> >> >> > >>> >> > > > less
> >> >> > >>> >> > > > > > CPU on average than the 0.32
> >> broker. I'll
> >> have a look
> >> >> > this
> >> >> > >>> >> weekend
> >> >> > >>> >> > > as
> >> >> > >>> >> > > > to
> >> >> > >>> >> > > > > > why creating the listeners is
> >> slower. On
> >> the dequeing,
> >> >> > can
> >> >> > >>> you
> >> >> > >>> >> > give
> >> >> > >>> >> > > a
> >> >> > >>> >> > > > > > little more information on the usage
> >> pattern
> >> - are you
> >> >> > using
> >> >> > >>> >> > > > > transactions,
> >> >> > >>> >> > > > > > auto-ack or client ack? What
> >> prefetch size
> >> are you
> >> >> using?
> >> >> > >>> How
> >> >> > >>> >> > large
> >> >> > >>> >> > > > are
> >> >> > >>> >> > > > > > your messages?
> >> >> > >>> >> > > > > >
> >> >> > >>> >> > > > > > Thanks,
> >> >> > >>> >> > > > > > Rob
> >> >> > >>> >> > > > > >
> >> >> > >>> >> > > > > > On 14 October 2016 at 23:46, Ramayan
> >> Tiwari <
> >> >> > >>> >> > > ramayan.tiwari@gmail.com
> >> <ma...@gmail.com>
> >> <mailto:ramayan.tiwari@gmail.com
> >>
> >> <ma...@gmail.com>>>
> >>
> >> >> > >>> >> > > > > > wrote:
> >> >> > >>> >> > > > > >
> >> >> > >>> >> > > > > > > Hi All,
> >> >> > >>> >> > > > > > >
> >> >> > >>> >> > > > > > > We have been validating the new Qpid
> >> broker (version
> >> >> > >>> 6.0.4)
> >> >> > >>> >> and
> >> >> > >>> >> > > have
> >> >> > >>> >> > > > > > > compared against broker version
> >> 0.32 and
> >> are seeing
> >> >> > major
> >> >> > >>> >> > > > regressions.
> >> >> > >>> >> > > > > > > Following is the summary of our
> >> test setup and
> >> >> results:
> >> >> > >>> >> > > > > > >
> >> >> > >>> >> > > > > > > *1. Test Setup *
> >> >> > >>> >> > > > > > > *a). *Qpid broker runs on a
> >> dedicated
> >> host (12
> >> >> cores,
> >> >> > >>> 32 GB
> >> >> > >>> >> > RAM).
> >> >> > >>> >> > > > > > > *b).* For 0.32, we allocated 16
> >> GB heap.
> >> For 6.0.6
> >> >> > >>> broker,
> >> >> > >>> >> we
> >> >> > >>> >> > use
> >> >> > >>> >> > > > 8GB
> >> >> > >>> >> > > > > > > heap and 8GB direct memory.
> >> >> > >>> >> > > > > > > *c).* For 6.0.4, flow to disk
> >> has been
> >> configured at
> >> >> > >>> 60%.
> >> >> > >>> >> > > > > > > *d).* Both the brokers use BDB
> >> host type.
> >> >> > >>> >> > > > > > > *e).* Brokers have around 6000
> >> queues
> >> and we create
> >> >> 16
> >> >> > >>> >> listener
> >> >> > >>> >> > > > > > > sessions/threads spread over 3
> >> connections, where each
> >> >> > >>> >> session is
> >> >> > >>> >> > > > > > listening
> >> >> > >>> >> > > > > > > to 3000 queues. However, messages
> >> are only
> >> enqueued
> >> >> and
> >> >> > >>> >> processed
> >> >> > >>> >> > > > from
> >> >> > >>> >> > > > > 10
> >> >> > >>> >> > > > > > > queues.
> >> >> > >>> >> > > > > > > *f).* We enqueue 1 million
> messages
> >> across 10
> >> >> > different
> >> >> > >>> >> queues
> >> >> > >>> >> > > > > (evenly
> >> >> > >>> >> > > > > > > divided), at the start of the test.
> >> Dequeue only
> >> >> starts
> >> >> > >>> once
> >> >> > >>> >> all
> >> >> > >>> >> > > the
> >> >> > >>> >> > > > > > > messages have been enqueued. We
> >> run the
> >> test for 2
> >> >> hours
> >> >> > >>> and
> >> >> > >>> >> > > process
> >> >> > >>> >> > > > as
> >> >> > >>> >> > > > > > > many messages as we can. Each
> >> message runs
> >> for around
> >> >> > 200
> >> >> > >>> >> > > > milliseconds.
> >> >> > >>> >> > > > > > > *g).* We have used both 0.16 and
> >> 6.0.4
> >> clients for
> >> >> > these
> >> >> > >>> >> tests
> >> >> > >>> >> > > > (6.0.4
> >> >> > >>> >> > > > > > > client only with 6.0.4 broker)
> >> >> > >>> >> > > > > > >
> >> >> > >>> >> > > > > > > *2. Test Results *
> >> >> > >>> >> > > > > > > *a).* System Load Average (read
> >> notes
> >> below on how
> >> >> we
> >> >> > >>> >> compute
> >> >> > >>> >> > > it),
> >> >> > >>> >> > > > > for
> >> >> > >>> >> > > > > > > 6.0.4 broker is 5x compared to 0.32
> >> broker. During
> >> >> start
> >> >> > >>> of
> >> >> > >>> >> the
> >> >> > >>> >> > > test
> >> >> > >>> >> > > > > > (when
> >> >> > >>> >> > > > > > > we are not doing any dequeue), load
> >> average is normal
> >> >> > >>> (0.05
> >> >> > >>> >> for
> >> >> > >>> >> > > 0.32
> >> >> > >>> >> > > > > > broker
> >> >> > >>> >> > > > > > > and 0.1 for new broker), however,
> >> while we are
> >> >> dequeuing
> >> >> > >>> >> > messages,
> >> >> > >>> >> > > > the
> >> >> > >>> >> > > > > > load
> >> >> > >>> >> > > > > > > average is very high (around 0.5
> >> consistently).
> >> >> > >>> >> > > > > > >
> >> >> > >>> >> > > > > > > *b). *Time to create listeners
> >> in new
> >> broker has
> >> >> gone
> >> >> > >>> up by
> >> >> > >>> >> > 220%
> >> >> > >>> >> > > > > > compared
> >> >> > >>> >> > > > > > > to 0.32 broker (when using 0.16
> >> client).
> >> For old
> >> >> broker,
> >> >> > >>> >> creating
> >> >> > >>> >> > > 16
> >> >> > >>> >> > > > > > > sessions each listening to 3000
> queues
> >> takes 142
> >> >> seconds
> >> >> > >>> and
> >> >> > >>> >> in
> >> >> > >>> >> > new
> >> >> > >>> >> > > > > > broker
> >> >> > >>> >> > > > > > > it took 456 seconds. If we use 6.0.4
> >> client, it took
> >> >> > even
> >> >> > >>> >> longer
> >> >> > >>> >> > at
> >> >> > >>> >> > > > > 524%
> >> >> > >>> >> > > > > > > increase (887 seconds).
> >> >> > >>> >> > > > > > > *I).* The time to create
> >> consumers
> >> increases as
> >> >> we
> >> >> > >>> create
> >> >> > >>> >> > more
> >> >> > >>> >> > > > > > > listeners on the same connections.
> >> We have
> >> 20 sessions
> >> >> > >>> (but
> >> >> > >>> >> end
> >> >> > >>> >> > up
> >> >> > >>> >> > > > > using
> >> >> > >>> >> > > > > > > around 5 of them) on each
> >> connection and
> >> we create
> >> >> about
> >> >> > >>> 3000
> >> >> > >>> >> > > > consumers
> >> >> > >>> >> > > > > > and
> >> >> > >>> >> > > > > > > attach MessageListener to it. Each
> >> successive session
> >> >> > >>> takes
> >> >> > >>> >> > longer
> >> >> > >>> >> > > > > > > (approximately linear increase) to
> >> setup
> >> same number
> >> >> of
> >> >> > >>> >> consumers
> >> >> > >>> >> > > and
> >> >> > >>> >> > > > > > > listeners.
> >> >> > >>> >> > > > > > >
> >> >> > >>> >> > > > > > > *3). How we compute System Load
> >> Average *
> >> >> > >>> >> > > > > > > We query the Mbean
> >> SysetmLoadAverage and
> >> divide it by
> >> >> > the
> >> >> > >>> >> value
> >> >> > >>> >> > of
> >> >> > >>> >> > > > > MBean
> >> >> > >>> >> > > > > > > AvailableProcessors. Both of these
> >> MBeans are
> >> >> available
> >> >> > >>> under
> >> >> > >>> >> > > > > > > java.lang.OperatingSystem.
> >> >> > >>> >> > > > > > >
> >> >> > >>> >> > > > > > > I am not sure what is causing these
> >> regressions and
> >> >> > would
> >> >> > >>> like
> >> >> > >>> >> > your
> >> >> > >>> >> > > > > help
> >> >> > >>> >> > > > > > in
> >> >> > >>> >> > > > > > > understanding it. We are aware
> >> about the
> >> changes with
> >> >> > >>> respect
> >> >> > >>> >> to
> >> >> > >>> >> > > > > > threading
> >> >> > >>> >> > > > > > > model in the new broker, are there
> any
> >> design docs
> >> >> that
> >> >> > >>> we can
> >> >> > >>> >> > > refer
> >> >> > >>> >> > > > to
> >> >> > >>> >> > > > > > > understand these changes at a high
> >> level?
> >> Can we tune
> >> >> > some
> >> >> > >>> >> > > parameters
> >> >> > >>> >> > > > > to
> >> >> > >>> >> > > > > > > address these issues?
> >> >> > >>> >> > > > > > >
> >> >> > >>> >> > > > > > > Thanks
> >> >> > >>> >> > > > > > > Ramayan
> >> >> > >>> >> > > > > > >
> >> >> > >>> >> > > > > >
> >> >> > >>> >> > > > >
> >> >> > >>> >> > > >
> >> >> > >>> >> > >
> >> >> > >>> >> >
> >> >> > >>> >>
> >> >> > >>> >
> >> >> > >>> >
> >> >> > >>>
> >> >> > >>
> >> >> > >>
> >> >> > >
> >> >> >
> >> >>
> >>
> >> -----------------------------
> >> ----------------------------------------
> >> To unsubscribe, e-mail: users-unsubscribe@qpid.apache.org
> >> <ma...@qpid.apache.org>
> >> <mailto:users-unsubscribe@qpid.apache.org
> >> <ma...@qpid.apache.org>>
> >> For additional commands, e-mail:
> >> users-help@qpid.apache.org <ma...@qpid.apache.org>
> >> <mailto:users-help@qpid.apache.org
> >> <ma...@qpid.apache.org>>
> >>
> >>
> >>
> >>
> >> ------------------------------------------------------------
> >> ---------
> >> To unsubscribe, e-mail: users-unsubscribe@qpid.apache.org
> >> <ma...@qpid.apache.org>
> >> For additional commands, e-mail: users-help@qpid.apache.org
> >> <ma...@qpid.apache.org>
> >>
> >>
> >>
> >> ------------------------------------------------------------
> ---------
> >> To unsubscribe, e-mail: users-unsubscribe@qpid.apache.org
> >> <ma...@qpid.apache.org>
> >> For additional commands, e-mail: users-help@qpid.apache.org
> >> <ma...@qpid.apache.org>
> >>
> >>
> >>
> >>
> >> ---------------------------------------------------------------------
> >> To unsubscribe, e-mail: users-unsubscribe@qpid.apache.org
> >> For additional commands, e-mail: users-help@qpid.apache.org
> >>
> >
> >
> >
> > ---------------------------------------------------------------------
> > To unsubscribe, e-mail: users-unsubscribe@qpid.apache.org
> > For additional commands, e-mail: users-help@qpid.apache.org
> >
> >
>