Performance and energy optimizations for online, data-intensive (OLDI) applications and network packet classification

Balajee Vamanan, Purdue University

Abstract

The growth of Internet data coupled with the emergence of interactive Web applications pose unique challenges to performance and energy-efficiency of datacenter and enterprise networks. As applications evolve from offline, throughput-oriented jobs to interactive Web applications, they demand low and predictable latency from the network. Further, because of their tight response time requirements, the interactive applications complicate energy management. On the other hand, as networks become large and functionally-rich, they introduce newer challenges for packet classification. My research addresses performance (i.e., low-latency) and energy optimizations for a class of interactive Web applications, called Online Data-Intensive applications (OLDIs), and packet classification, and it spans from network hardware to transport protocols and applications. Fine-grained network management (e.g., software-defined networking) increases the number of flow-rules which complicates both lookups (i.e., search) and updates in packet classification at the switch hardware. I proposed EffiCuts and TreeCAM that address lookups and updates, respectively. EffiCuts and TreeCAM effectively decouple lookups from updates and employ distinct data structures for lookups and updates. EffiCuts reduces the memory overhead for lookups by two orders of magnitude. TreeCAM reduces the update effort by a factor of 30. Modern interactive Web applications like Web Search operate under soft-real-time constraints (e.g., 300 ms latency) which imply (1) deadlines for network communication and (2) saving energy is hard. For predictable latency, I proposed Deadline-aware Datacenter TCP (D2TCP), a transport protocol that achieves deadline-based prioritization of network flows. I evaluated D2TCP in Google's datacenters and showed that D2TCP reduces missed deadlines by 50% as compared to existing transport protocols. For energy-efficiency, I proposed TimeTrader which exploits slack in sub-critical replies (i.e., replies that do not fall in the tail of the latency distribution). TimeTrader exploits slack from both network and compute layers. Timetrader achieves significant energy savings over existing schemes - 15% at peak load and 40% at 30% load.

Degree

Ph.D.

Advisors

Vijaykumar, Purdue University.

Subject Area

Computer Engineering|Computer science

Off-Campus Purdue Users:
To access this dissertation, please log in to our
proxy server
.

Share

COinS