Daniel Suo

Scientific progress goes 'boink'

Ph.D. Candidate
Princeton University
Department of Computer Science

Coflow: A Networking Abstraction for Cluster Applications

(link)

A coflow is a collection of flows (data over network between two nodes) between two groups of machines with associated semantics and a collective objective (e.g., shuffle in MapReduce).

Three primary patterns: shuffle, broadcast, aggregation. Two primary objectives: minimize completion times and meeting deadlines.

Coflows can interact via sharing, prioritizing, and ordering.

API:

  • create / update a coflow
  • get / put (senders and receivers create pairs on content id)
  • terminate (coflow is done)