A Case for Performance-Centric Network Allocation
Source: University of California
The authors consider the problem of allocating network resources across applications in a private cluster running data-parallel frameworks. Their primary observation is that these applications have different communication requirements and thus require different support from the network to effectively parallelize. They argue that network resources should be shared in a performance-centric fashion that aids parallelism and allows developers to reason about the overall performance of their applications. This paper tries to address the question of whether/how fairness centric proposals relate to a performance-centric approach for different communication patterns common in these frameworks and engages in a quest for a unified mechanism to share the network in such settings.