[rabbitmq-discuss] Performance bottleneck on inter-node connection

[Matthias Radestock]

On 08/04/14 19:44, joseph rouphael wrote:
> I am facing a performance bottleneck on the inter-node connection
> between two nodes in a cluster.
> [...]
> {hipe_compile, true},

Have you tried this w/o hipe?

> {delegate_count, 32},

Why did you set this?

> Queues are located on node1.
>
> When producing/consuming messages of 1KB on the same node (node 1), I am
> able to achieve up to 100K TPS, even more.
>
> If I produce messages on node 2, and consume messages from node 1; I am
> not able to hit more then 15K TPS. Knowing that the CPU load on the two
> nodes is very relaxed (90% idle), and no bottleneck on the network
> bandwidth between the nodes.
>
> The same, if I produce on node 1, and consume on node 2; I am not able
> to hit more then 15K TPS
>
> I am suspecting a limitation on the inter-node connection between the nodes.
>
>
> Is there a way to create multiple links between the nodes to scale up?
>
> If not possible to create multiple links, Is there a way to optimize the
> link performance?
>
> Is there something else I can do to increase the performance between nodes?

I suspect your test is latency sensitive rather than bandwidth 
sensitive, presumably because you are doing something synchronous, e.g. 
publish a messages and wait for it to arrive at the consumer, then send 
the next message, etc.

So please explain what your test actually does. How are you measuring TPS?


Matthias.