It's quite an obvious point but I'll ask it anyway - are there
differences in the network topology between the nodes exhibiting the
different behaviour? e.g. switches, NICs?

Which OS are you using?

I'm not entirely clear about the topology you are describing - is the
publisher on a separate machine from the consumers?

Robert

2009/10/29 Andy Li <andr.deve...@gmail.com>:
> Hi,
>
> We ran into strange differences of flow control behavior of fanout exchange.
> We use configuration with one publisher and 10 subscribers.
>
> 1)On some machines, publisher throughput will be higher than subscriber
> throughput, and eventually the queue will overflow and throw an exception.
> This is, as I understand, the expected behavior - publisher has entire 1
> Gbit bandwidth to itself, whereas subscribers share it, so they have lower
> theoretical limit. We've coded a simple flow control scheme to manage this.
>
> 2)However, on other machines we installed the broker, publisher and each
> subscriber throughput seem to move in sync. There is little buffering in the
> queue, and the overflow never happens. While the lack of overflow is
> convenient, bandwidth utilization is 2-4 times lower compared to broker
> machines exibiting behavior (1). In (1), combined output bandwidth utilized
> by subscribers is 80-90% of 1Gbit available. In (2), it's 20-40%, depending
> on the machine and latency to client.
>
> All broker machines use the same OS version and broker 0.5 binaries with
> default broker config. We can't figure out what is causing this difference
> in behavior. We did seem to induce behavior (2) on machines (1) by running
> unrelated workload concurrently that strongly competed for bandwidth.
>
> Has anyone else run into the same issue? Or has any idea about what might
> cause this? Any help or suggestions would be greatly appreciated.
>
> Thanks,
> Andy
>
> Client Details:
>
> The publisher and subscriber clients are the same in both cases and run on a
> separate machine. We're using asynchronous send and sync() every 1000
> messages. Listeners use ACCEPT_MODE_NONE and FlowControl::unlimited().
> Message size is random, linearly distributed between 500-1500 bytes.
>

---------------------------------------------------------------------
Apache Qpid - AMQP Messaging Implementation
Project:      http://qpid.apache.org
Use/Interact: mailto:users-subscr...@qpid.apache.org

Reply via email to