Towards reproducible performance studies of datacenter network architectures using an open-source simulation approach

Abstract

In datacenter network (DCN) research, one key challenge is to reproduce the relative performances of different scalable DCN architectures in an unbiased and transparent way. Adequately addressing this challenge will support the validation of performance studies, and this is fundamental to building a strong research foundation and making wise datacenter investment decisions. In addressing this challenge, this paper presents the NTU-DSI-DCN initiative with a DCN simulation module based on an open-source platform called ns-3, on which the performance models of DCN topologies can be developed and made open source to support independent reproducibility of their performances. Advantages of the framework include the following: (1) it is low cost, (2) it provides transparent performance benchmarks of known DCN architectures and (3) it enables unbiased comparative performance simulations of DCN architectures, without the tedium of developing existing DCN models from scratch. In realizing this NTU-DSI-DCN initiative, the open-source performance models of the Fat tree and the BCube architectures have been implemented on the ns-3 platform. A comparative performance study between Fat tree and BCube is reported, along with a performance reproducibility study of Fat tree. The documentation and source codes for our simulation setups are publicly available at http://code.google.com/p/ntu-dsidcn/. In continually adding new DCN architectural models or their variants in future work, the NTU-DSI-DCN is a promising initiative that can evolve into a well-documented open-source simulation platform supporting quality research in DCNs.