High-fidelity measurements in low-latency networks

Myungjin Lee, Purdue University

Abstract

Many data center applications such as web search, advertising, online retail, social collaboration, recommendation systems, and high performance computing (HPC) require low latency from the network. Failure to meet the latency requirements in the network aggravates user experience and leads to potential loss of customers. Thus, detecting, localizing and repairing delay spikes that cause application Service Level Agreement (SLA) violations in low-latency networks is one of most crucial tasks to support these applications properly. Yet, network operators today lack sufficient fine-grain latency measurement tools that can be deployed ubiquitously. On the one hand, solutions such as SNMP and NetFlow with which modern routers are equipped provide little information about latency and jitter that applications experience. On the other hand, high-end latency measurement solutions for stock exchange markets prohibit ubiquitous deployment due to their high cost despite their effectiveness. In this dissertation, to bridge the gap, we propose a comprehensive set of router solutions for obtaining high-fidelity latency measurements. The solutions (1) detect application-specific latency anomalies and localize a responsible router, (2) support arbitrary measurement granularity such as per-packet latencies, and (3) are resilient to phenomena such as packet loss and reordering that happen in IP networks.

Degree

Ph.D.

Advisors

Rao, Purdue University.

Subject Area

Computer Engineering

COinS