The Cost and Benefits of Coordination Programming: Two Case Studies in Concurrent Collection and S-Net

Abstract

This is an evaluation study of the expressiveness provided and the performance delivered by the coordination language S-NET in comparison to Intel’s Concurrent Collections (CnC). An S-NET application is a network of black-box compute components connected through anonymous data streams, with the standard input and output streams linking the application to the environment. Our case study is based on two applications: a face detection algorithm implemented as a pipeline of feature classifiers and a numerical algorithm from the linear algebra domain, namely Cholesky decomposition. The selected applications are representative and have been selected by Intel researchers as evaluation testbeds for CnC in the past. We implement various versions of both algorithms in S-NET and compare them with equivalent CnC implementations, both with and without tuning, previously published by the CnC community. Our experiments on a large-scale server system demonstrate that S-Net delivers very similar scalability and absolute performance on the studied examples as tuned CnC codes do, even without specific tuning. At the same time, S-Net does achieve a much more complete separation of concerns between compute and coordination layers than CnC even intends to.