Have you tried tracing the workload on the 100% and 40% nodes for comparison? There just isn't enough detail in your question to help predict what should be happening with the cluster workload. For a starting point, please identify your design goals. It's easy to get confused by advice that seeks to help you do something you don't want to do.
Some things to think about include how the stream workload is composed. How should/would this work if there were only one node? How should behavior change as nodes are added to the topology and the test is repeated? Gedanken: what if the data streamer is doing some really expensive operations as it feeds the data into the stream, but the nodes can very cheaply put the processed data into their cache partitions? In this case, for example, the expensive operations should be refactored into a stream transformer that will move the workload from the stream sender to the stream receivers. https://ignite.apache.org/docs/latest/data-streaming#stream-transformer Also gedanken: what if the data distribution is skewed such that one node gets more data than 2x the data sent to other partitions because of affinity? In this case, for example, changes to affinity/colocation design or changes to cluster topology (more nodes with greater CPU to RAM ratio?) can help distribute the load so that no single node becomes a bottleneck. On Tue, Feb 28, 2023 at 9:27 AM John Smith <java.dev....@gmail.com> wrote: > Hi I'm using the data streamer to insert into a 3 cluster node. I have > noticed that 1 node is pegging at 100% cpu while the others are at 40ish %. > > Is that normal? > > >
