Hardware Implementation
pFabric: Minimal Near-Optimal Datacenter Transport!Original paper: Mohammad Alizadeh, Shuang Yang, Sachin Katti, Nick McKeown, Balaji Prabakhar, Scott Shenker!
Further work on NetFPGA-based implementation: Yan Michalevsky!Department of Electrical Engineering, Stanford University
pFabric Simulation
Future Plans
Motivation: Providing Near-Optimal Performance for Flows in a Datacenter Network!
- Initial simulation performed using NS-2 network simulator. - Results confirmed by a different team using the Mininet
network emulation framework (Wendy Mu, Bridget Vuong). - pFabric is expected to provide near-optimal flow completion
times and outperform the compared schemes (TCP DropTail, DCTCP, PDQ).
Goals - Provide fairness and high throughput in presence of simultaneous mice and
elephant flows. - A simple scheme that can be practically deployed step-by-step using existing
equipment. Solution Simple packet scheduling logic according to flow priority set by endpoints (based on remaining flow size).
- Based on NetFPGA 10G (netfpga.org) - Step-by-step implementation starting with a simple priority-
queues based scheme.
- Integration with RAMCloud project - Implementation of the full pFabric scheme - Validation of pFabric’s promise in a real high-speed network - How can we approximate remaining flow size when it is not
available in advance?