Router CPU load is a significant concern in the operation of an ISP. Long periods of high CPU load can increase route convergence times, and has some correlation with router instability. The BGP routing protocol has a potential for significantly impacting CPU load due to the nature of the protocol. Several actions need to be taken upon receiving a route announcement, and on average more route announcements are seen by an ISP today than ever before. In this study, we examine the impact of BGP activity on operational routers in the Sprint IP network. We find that on average, BGP processes tend to consume over 60% of a router's non-idle CPU cycles. During short time scales (5 seconds), we have observed BGP processes contributing almost 100% CPU load. During longer time scales (1 — 5 minutes), we see a weaker correlation. During normal network operation, we find that there is some correlation between increased BGP activity and router CPU load, but the impact is small. During an abnormal network event that lasted over 10 hours, we find a correlation. However, the increase in CPU load was under 20% for most routers. BGP processes tend to run frequently for very short intervals, during which the CPU can reach the maximum utilization. Due to this, BGP consumes the majority of CPU cycles over a very long period of time (weeks). The quantity of BGP messages received during a particular cycle can increase the CPU load. However, this increase is not consistently large enough to cause concern about the operation of the router. A possibility is that certain kinds of BGP messages may increase the load more than others. However, without a detailed understanding of the specific implementation of BGP in these routers, we cannot speculate on such specific behavior. Thus we conclude that during normal operation, CPU load is not significantly impacted by BGP activity in the time scale of minutes. Short term impact in the time scale of seconds in not likely to significantly impact convergence times or router stability. During abnormal events of the magnitude of the SQL Slammer worm, router CPU is not likely to increase significantly.
展开▼