Data scientists and practitioners of high-performance computing, the target audience for this paper, know well that massively parallel systems generate massive amounts of data, which must be stored and transmitted as well as processed. In the example presented here, a combustion simulation, the bottleneck is input/output (I/O), but there is spare computing power available. Hence, the authors use nearline storage and data compression at the core level to reduce the communications overhead substantially. The paper discusses the file formats and storage schemes in detail, but there is no detail of the compression algorithms used, which future researchers might find helpful for comparison.

It might be interesting to see whether techniques from compressive sensing [1], which could reduce the amount of input data rather than the amount of output data, would be useful in this and similar applications.

By the way, the use of the term “CODEC” in the title is a specialization of the traditional meaning; here it means “COmpression and DECompression.”