Size-based scheduling vs fairness for datacenter flows

Contrary to the conclusions of a recent body work where approximate shortest remaining processing time first (SRPT) flow scheduling is advocated for datacenter networks, this paper aims demonstrate that per-flow fairness remains preferable objective. We evaluate abstract queuing models by analysis and simulation illustrate non-optimality SRPT under reasonable assumptions flows occur in batches bursts not, as usually assumed, individually at instants Poisson process. Results these have significant implications design bandwidth sharing strategies networks. In particular, we propose novel "virtual fair scheduling" algorithm enforces between arguably simple enough be implemented high speed devices.