Is your feature request related to a problem? Please describe.
We used 7 machines each with 128 cores and 320GB to set up a new production environment. In performing the RabbitMQ performance validation, we deployed a single replica of rabbitmqcluster with 8 cores and 8GB on one of the nodes. With 10 quorum queues and 30 producers, the incoming messages could reach 60-70K/s.
We then deployed another single replica of rabbitmqcluster with the same configuration on that node and stress-tested all instances together, but the incoming rate for the quorum queues remained at 60-70K/s.
After deploying 5 more single replicas of rabbitmqcluster and conducting a stress test on all instances together, the total production/consumption was only around 60-70K/s, showing no improvement.
Through multiple rounds of testing, we discovered that the CPU utilization of cores 0-7 was consistently >90%, while all other CPU cores remained <10%.
Describe the solution you'd like
We set erlang +stbt config to unbound and redo stress test, the load spread around all the cores, and the total production came to 270K/s，though the throughput was not very stable.
In Kubernetes, it is quite common to deploy multiple RabbitMQ pods on the same node. If the default value of +stbt is kept as "db" (the actual value is 'tnnps'), it can significantly limit the overall performance of RabbitMQ.
Besides, the unbound config is not listed in rabbitmq doc, which i think shoud be added too.
Is your feature request related to a problem? Please describe.
We used 7 machines each with 128 cores and 320GB to set up a new production environment. In performing the RabbitMQ performance validation, we deployed a single replica of rabbitmqcluster with 8 cores and 8GB on one of the nodes. With 10 quorum queues and 30 producers, the incoming messages could reach 60-70K/s.
We then deployed another single replica of rabbitmqcluster with the same configuration on that node and stress-tested all instances together, but the incoming rate for the quorum queues remained at 60-70K/s.
After deploying 5 more single replicas of rabbitmqcluster and conducting a stress test on all instances together, the total production/consumption was only around 60-70K/s, showing no improvement.
Through multiple rounds of testing, we discovered that the CPU utilization of cores 0-7 was consistently >90%, while all other CPU cores remained <10%.
Describe the solution you'd like
We set erlang +stbt config to unbound and redo stress test, the load spread around all the cores, and the total production came to 270K/s，though the throughput was not very stable.
Config as below:
Additional context
In Kubernetes, it is quite common to deploy multiple RabbitMQ pods on the same node. If the default value of +stbt is kept as "db" (the actual value is 'tnnps'), it can significantly limit the overall performance of RabbitMQ.
Besides, the unbound config is not listed in rabbitmq doc, which i think shoud be added too.
Doc refer: