Understanding Network Failures in Data Centers: Measurement, Analysis, and Implications
Proceedings of the ACM Special Interest Group on Data Communications (SIGCOMM '11) |
Published by ACM
We present the first large-scale analysis of failures in a data center network. Through our analysis, we seek to answer several fundamental questions: which devices/links are most unreliable, what causes failures, how do failures impact network traffic and how effective is network redundancy? We answer these questions using multiple data sources commonly collected by network operators. The key findings of our study are that (1) data center networks show high reliability, (2) commodity switches such as ToRs and AggS are highly reliable, (3) load balancers dominate in terms of failure occurrences with many short-lived software related faults, (4) failures have potential to cause loss of many small packets such as keep alive messages and ACKs, and (5) network redundancy is only 40% effective in reducing the median impact of failure.