Benchmarking RabbitMQ on Raspberry Pi

2 minute read

A while ago I have written about how to set up rabbitmq on a Raspberry. Now I was interested to get some performance data about this setup. There is a great set of benchmarks on the high end spectrum - Google and Pivotal show a benchmark at 1 million messages per second. And there is a set of [older benchmark results from 2012] (https://www.rabbitmq.com/blog/2012/04/17/rabbitmq-performance-measurements-part-1/) with corresponding toolset from the makers of rabbitmq as well.

I have re-run part of these benchmarks on both generations or Raspberry Pi, the Raspberry 1 and Raspberry 2. Just to see how a 35$ computer compares.

Benchmark results

The first set of tests is a simple publish/consume cycle with one publisher and one publisher in parallel. Here is how it looks on the old armv61 CPU of the Raspberry 1:

Summary 1->1 armv61

More than 700 per sec, I am already impressed. Based on my work earlier I was not expecting such a high throughput from a single-core CPU armv61.

Consume 1->1 armv61

compare this with the same test run on the arm71 CPU of the Raspberry Model 2:

Summary 1->1 armv71

Almost 4x the throughput of the smaller brother. This is going to be fun. I need to start looking for use cases which require this throughput at a 35$ budget.

Consume 1->1 armv71

In both cases you can see that the publisher by far outpaces the consumer and gets throttled when the queue gets longer. This results also in the consumer still running for 10-20 sec after the publisher finished. Accordingly the latency of the messages is given by the queuing in the rabbitmq. With a slower publisher the latency would go down significantly - this could be shown in a future test where we ramp up the publisher speed to show the impact on latency.

But the total impact of this effect is vastly different:

  • the old armv61 consumes a surprisingly good 700 messages per second, and the publisher has to wait every second or so.
  • but on the new armv71 thanks to 4 CPU cores the output is much higher and allows more regular input.
  • as a result the old armv61 needs 20 sec to drain the queue at the end of the test (and the latency of the messages accordingly is at 20 sec) while the multi.core armv71 only needs 10 sec to drain the queue.

Ramping up

One observation from the initial test was that the producer gets throttled by the consumer. Adding a second consumer process should help? Here are the results:

Summary 1->2 armv61 Summary 1->2 armv71

Absolutely neglegible improvement, Lets look at the details:

Consume 1->2 armv61 Consume 1->2 armv71

Actually, the latency on the old Raspberry gets even longer, and the producer can only send in short, sharp bursts.

Summary

When stretching to the limit then something is going to be near breaking point - in this case it is the latency induced by the queuing. The publisher only gets throttled when a certain queue length s reached.

Overall still absolutely impressive numbers and the system keeps up with the traffic by throttling the publisher so that the system internally can keep delivering.

What’s next?

I plan to elaborate in a next post how the setup was done for this test and could also dive deeper into different test scenarios to isolate the bottlenecks of the Raspberry (slow CPU, slow disk, slow Network) and expose then in separate tests.

Leave a Comment