EndRE: An End-System Redundancy Elimination Service for Enterprises

  • Bhavish Aggarwal ,
  • Aditya Akella ,
  • Ashok Anand ,
  • Athula Balachandran ,
  • Pushkar V. Chitnis ,
  • Chitra Muthukrishnan ,
  • ,
  • George Varghese

Symposium on Networked Systems Design and Implementation (NSDI) |

Published by USENIX

In many enterprises today, middleboxes called WAN optimizers are being deployed across WAN access links in order to eliminate redundancy in network traffic and reduce WAN access costs. In this paper, we present the design and implementation of EndRE, an alternate approach where redundancy elimination is provided as an end system service. Unlike middleboxes, such an approach benefits both end-to-end encrypted traffic as well as traffic on last-hop wireless links to mobile devices.

EndRE needs to be fast, adaptive and parsimonious in memory usage in order to opportunistically leverage resources on end hosts. Thus, we design a new fingerprinting scheme called SampleByte that is much faster than Rabin fingerprinting while delivering similar compression gains. Unlike Rabin, SampleByte can also adapt its CPU usage depending on server load. Further, we introduce optimizations to reduce server memory footprint by 33-75% compared to prior approaches. Using several terabytes of network traffic traces from 11 enterprise sites, testbed experiments and a pilot deployment, we show that EndRE delivers 26% bandwidth savings on average, processes payloads at speeds of 1.5-4Gbps, reduces end-to-end latencies by up to 30%, and translates bandwidth savings into equivalent energy savings on mobile smartphone.