Clustering performance regression in ejbremote-repl-sync scenario

We can see performance regression in our stress tests, which test performance of the cluster under increasing load (number of concurrent clients). All tests use a 4-node EAP cluster and 5 nodes that generate load.

Scenario name: stress-ejbremote-repl-sync
Scenario description:

• remote invocations of stateful clustered EJBs
• default cache: repl (replicated): state is replicated to all cluster members
• cache mode: sync: replication is synchronous (i.e., response is only sent to the client when all replicas confirm the update)
• read/write ratio: default: all requests read and write to the session

Regression can be seen in ER4 run compared to ER3 run (only).

See the attachment for graphs with results.

Horizontal axis represents number of concurrent clients generating load.
Vertical axis represents throughput = TPS (number of requests processed per second).

In short, ER4 was able to reach 10000 TPS compared to 13000 TPS in ER3 run.

Links:
EAP 7.0.0.ER4: graph-throughput.png
EAP 7.0.0.ER3: graph-throughput.png
EAP 6.4.0.GA (JDK8): graph-throughput.png