Fixstars recently ran a series of benchmarks that compared Cassandra and GridDB on Microsoft Azure using 1, 8, 16, and 32-node clusters with an equal number of YCSB clients and both in-memory (4M records per node) and out-of-memory (12M records per node) data sets.
As the following graphs show, GridDB’s hybrid storage architecture, in-memory-oriented architecture, outperforms Cassandra both in-memory and in operations required using out-of-memory storage. GridDB accomplishes this while maintaining the same reliability and consistency through twenty-four hours of operation.
The internode communication of GridDB scales significantly better than Cassandra’s decentralized peer-to-peer system, at least up through 32 nodes. GridDB’s performance increases by nearly the same factor as the number of nodes added; Cassandra is only able to scale at 50% of that same factor with this particular Azure instance type.
In update-intensive workloads such as Workload A, Cassandra’s initial results are quite favorable as its log based architecture allows it to quickly mark a row as deleted and then append the new value to the end of the log. Fixstars noticed that over time Cassandra began to slow down.
Fixstars configured an 8-node cluster and loaded 4M and 12M records per node and set operationcount to 2^32-1 and let the test run for twenty-four hours. In both tests, Cassandra’s throughput is less than 50% of what it was in the twenty fourth hour versus the first. Meanwhile GridDB’s performance was stable when doing both in and out of memory operations.
The full report can be downloaded here: https://griddb.net/en/docs/Fixstars_NoSQL_Benchmarks.pdf
The GridDB Community Edition (v3.0.1) can be downloaded from GitHub.