When the various testing environments were created and configured, Dell Technologies and Intel engineers ran a series of performance tests specific to Apache Cassandra. A closed-loop tuning approach was used where only one parameter was varied for any given test. This methodology isolates the effect of any one configuration change and is more deterministic than making several tuning changes in parallel.
The performance tests used the cassandra-stress tool, a Java-based utility for benchmarking and load testing Apache Cassandra. The tool is used to measure the Apache Cassandra cluster’s total throughput by adding all the client operations per second. The engineers also measured the SLA, which is the average latency required to achieve maximum throughput. The SLA for these scenarios is the average of all client operations per second throughput at 99th percentile latency. The engineers can vary the number of client threads until the clients achieve throughput at or below the target SLA. For more information about cassandra-stress, see the DSE 6.8 Developer Guide.
The engineers ran each performance benchmark for 10 minutes each. To provide predictable performance results, the dataset was compacted after each run and the caches were cleared before starting the next benchmark.
The Dell Technologies and Intel engineers ran all performance tests from three separate client machines. Overall, this configuration represents a typical hybrid or multicloud deployment where applications or clients that use the Apache Cassandra run from a remote location.
For specific information about the Apache Cassandra settings, database build commands and benchmark steps, see Appendix A.
