[
https://issues.apache.org/jira/browse/QPID-3460?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=13095527#comment-13095527
]
[email protected] commented on QPID-3460:
-----------------------------------------------------
bq. On 2011-09-01 14:44:26, Alan Conway wrote:
bq. > /trunk/qpid/cpp/include/qpid/messaging/Tracker.h, line 46
bq. > <https://reviews.apache.org/r/1687/diff/1/?file=36886#file36886line46>
bq. >
bq. > Remove track(Sender). Send and receive should be symmetric and we
may add more event types later.
bq. >
bq. > Applications using this API will themselves be server-like and
getting their own async events that trigger the need to send messages. If any
session gets capacity they'll want to check if they have an event for that
session. Apps that know in advance about a set of messages to send on a session
won't use this API.
bq. >
bq. > If there's an easy way for the user to tell which sessions are
interesting, they can implement that in their own loop, no need to push that
responsibility onto the Tracker.
bq.
bq. Gordon Sim wrote:
bq. I don't think send and receive *are* symmetric. The only reason you
care about send in this context is where you have messages to send but cannot
do so due to capacity limits in the sender. What I think you therefore want is
to be able to track when a sender is 'writable'. However in general most
senders on session will be writable. So while for incoming messages the session
is a good level to track, for outgoing messages sender level control is far
more important in my view.
bq.
bq. Alan Conway wrote:
bq. We can't know in general whether there will be more sending or
receiving in a given application. Applying your argument to sys::Poller, do you
think we should have had a different API for reads and writes? Consider the way
the broker IO works - driven by readabilty/writabiliy. Applications using this
API are going to be server like applications driven by
sendability/receivability. A sending application will be queuing up its own
events and needs to know when they can be sent. If the senders generally do
have capacity the applications queue will be empty, but whatever is driving
the application may temporarily exceed capacity and the the app needs to be
notified when there is capacity.
bq.
bq. I'm not arguing against a finer granularity, but it should be
symmetric between read and write. If you can watch individual Senders you
should also be able to watch individual receivers.
bq.
bq. The asymetric API is unclear: does track(Session) mean track all its
senders and receivers or just receivers? Does remove(Session) also remove
senders that were added by track(Sender)?
It's not about whether there is more sending or receiving, its what is the most
common need in each case. As I mentioned in the description, I can certainly
see extending this to support tracking individual receivers and also tracking
all the senders for a session. However the most common cases are that you care
about specific senders, but about all sessions, so thats where I started.
I agree that the asymmetry is less than ideal though.
bq. On 2011-09-01 14:44:26, Alan Conway wrote:
bq. > /trunk/qpid/cpp/include/qpid/messaging/Tracker.h, line 49
bq. > <https://reviews.apache.org/r/1687/diff/1/?file=36886#file36886line49>
bq. >
bq. > I would suggest something more like sys::Poller with a single
wait() function that returns an Event. That's a pretty common pattern.
(seelect, poll, epoll etc)
bq. >
bq. > I'd also suggest replacing track/remove
bq. > with:
bq. >
bq. > void monitor(Session, incoming, outgoing)
bq. >
bq. > So you can control monitoring per-session. remove becomes:
monitor(session, false, false)
bq. >
bq. > I think the 2 booleans probably want to be replaced by an enum or
set of flags to be
bq. > more extensible.
bq. >
bq.
bq. Gordon Sim wrote:
bq. There is a wait() function as well that can be used to wait until
there is an incoming message on one of the tracked sessions and/or until an
outgoing message can be sent on one of the tracked senders. As above I don't
think the session provides sufficient granularity for tracking ability to send
without blocking; I think that needs to be done on specific senders.
bq.
bq. Alan Conway wrote:
bq. In that case we need the finer granularity for receivers too.
Fair enough, but I think that is driven more by the desire for symmetry than by
real world needs.
- Gordon
-----------------------------------------------------------
This is an automatically generated e-mail. To reply, visit:
https://reviews.apache.org/r/1687/#review1718
-----------------------------------------------------------
On 2011-08-31 18:01:43, Gordon Sim wrote:
bq.
bq. -----------------------------------------------------------
bq. This is an automatically generated e-mail. To reply, visit:
bq. https://reviews.apache.org/r/1687/
bq. -----------------------------------------------------------
bq.
bq. (Updated 2011-08-31 18:01:43)
bq.
bq.
bq. Review request for Alan Conway and Darryl Pierce.
bq.
bq.
bq. Summary
bq. -------
bq.
bq. This allows multiple sessions (from multiple connections) to be managed
with a single thread. It also allows non-blocking control over senders that
reach their capacity (e.g. due to flow control).
bq.
bq. It does not yet cover tracking of message settlement and reconnection is
still blocking.
bq.
bq. The API changes consist of a new class, Tracker - not at all keen on that
name but have yet to come up with something I like - through which a set of
sessions and/or senders can be tracked for 'readability' or 'writability'
respectively. This is somewhat asymmetric. The reason for this choice was that
in general you only care about the writability (i.e. available capacity) of
certain senders (those you have data to send out on at present), whereas
generally you care about any incoming messages. I could see this being extended
to support both tracking specific receivers and to track all senders on a
session but those seem like more special cases and not as critical at first.
bq.
bq. The internal implementation on the 0-10 codebase leaves a lot to be
desired. At present it is certainly not the most optimal, but it has virtually
no effect on users who don't use the new class. The threading of the 0-10
client is driven by its use of the old client API underneath it. That and the
locking in the messaging API layer means that a lot of checking actually needs
to occur on application threads rather than the IO thread. Its this reason that
I haven't tried at this stage to e.g. make a file handle readable when the
tracker has a next() session available. Since it would in any case require a
separate thread, little is gained at this stage. I envisage the 1.0
implementation being able to handle that case much better, being architected
from the start with a more ideal threading.
bq.
bq. So while this patch is still fairly rudimentary I thought it was worth
sharing for some wider comment from interested parties.
bq.
bq.
bq. This addresses bug QPID-3460.
bq. https://issues.apache.org/jira/browse/QPID-3460
bq.
bq.
bq. Diffs
bq. -----
bq.
bq. /trunk/qpid/cpp/src/qpid/messaging/SessionImpl.h 1163236
bq. /trunk/qpid/cpp/src/qpid/sys/BlockingQueue.h 1163236
bq. /trunk/qpid/cpp/src/qpid/messaging/Session.cpp 1163236
bq. /trunk/qpid/cpp/examples/messaging/Makefile.am 1163236
bq. /trunk/qpid/cpp/examples/messaging/extra_dist/Makefile 1163236
bq. /trunk/qpid/cpp/examples/messaging/non_blocking.cpp PRE-CREATION
bq. /trunk/qpid/cpp/include/qpid/messaging/Session.h 1163236
bq. /trunk/qpid/cpp/include/qpid/messaging/Tracker.h PRE-CREATION
bq. /trunk/qpid/cpp/src/CMakeLists.txt 1163236
bq. /trunk/qpid/cpp/src/Makefile.am 1163236
bq. /trunk/qpid/cpp/src/qpid/client/SessionImpl.h 1163236
bq. /trunk/qpid/cpp/src/qpid/client/SessionImpl.cpp 1163236
bq. /trunk/qpid/cpp/src/qpid/client/amqp0_10/SenderImpl.cpp 1163236
bq. /trunk/qpid/cpp/src/qpid/client/amqp0_10/SessionImpl.h 1163236
bq. /trunk/qpid/cpp/src/qpid/client/amqp0_10/SessionImpl.cpp 1163236
bq. /trunk/qpid/cpp/src/qpid/client/amqp0_10/SessionTracker.h PRE-CREATION
bq. /trunk/qpid/cpp/src/qpid/client/amqp0_10/SessionTracker.cpp PRE-CREATION
bq.
bq. Diff: https://reviews.apache.org/r/1687/diff
bq.
bq.
bq. Testing
bq. -------
bq.
bq. Minimal ad hoc testing with the new example contained in this patch.
bq.
bq.
bq. Thanks,
bq.
bq. Gordon
bq.
bq.
> Better support for non-blocking usage in messaging API
> ------------------------------------------------------
>
> Key: QPID-3460
> URL: https://issues.apache.org/jira/browse/QPID-3460
> Project: Qpid
> Issue Type: Improvement
> Components: C++ Client
> Affects Versions: 0.12
> Reporter: Gordon Sim
> Assignee: Gordon Sim
>
> In particular:
> * remove the requirement for a thread-per-session when processing incoming
> messages
> * remove the need to either block or poll when a sender reaches its capacity
> * remove the need to poll for changes in message settlement (i.e. completion
> of sends and acknowledgements)
> * allow non-blocking reconnection
> * allow integration with polling loops etc
--
This message is automatically generated by JIRA.
For more information on JIRA, see: http://www.atlassian.com/software/jira
---------------------------------------------------------------------
Apache Qpid - AMQP Messaging Implementation
Project: http://qpid.apache.org
Use/Interact: mailto:[email protected]
