I have multiple microservices talking to each other over network. When these microservices are deployed on Kubernetes, my application is experiencing a periodic drop of throughput on loadtesting, where I am pinning one of the microservice to one core of the CPU and saturating it to 100%. Note that all the pods are on the same node. The time series throughput plot is as follows: [Time series throughput of the above setup when load test performed for 10 min]
I have tried 3 setups:
- Running the microservices application on baremetal with communication on localhost
- Running the microservices on different pods but with host Networking but pods on same node
- Running the microservices on different pods but without host Networking but pods on same node
The throughput in the first case is highest. For 2nd case, it is almost 95% of the first case which is acceptable. But 3rd case is the one, for which I am seeing a periodic drop of throughput every few seconds.
What could be the reason for this? Is there some queue that is getting full or is it some configuration issue?
Note: The microservices are simeple client server application built using C++ and cpprestsdk and using redis as a DB. The images of these have a base image of ubuntu.
Cluster information: Kubernetes version: 1.26.3
Cloud being used: Bare-Metal
Installation method: kubeadm
Host OS: Ubuntu 20.02