Arthur Dick
October 7, 2005
Core routers in use today contain buffers which may be much larger than necessary. This project will use simulation software to analyse what effect changing the buffer sizes in routers would have on the performance of network protocols. Specifically, we will be focusing on TCP performance, as it is the transfer protocol in use for popular applications such as the web, FTP, email, and peer to peer networks.
The widely used rule-of-thumb used to determine Internet router buffer size is that we need at least one round trip time worth of buffering[2]. This rule-of-thumb came to be prior to a thourough study on how much memory is needed before a loss in performance occurs.
The goal of this project is to determine if the current rule-of-thumb used to determine the amount of buffer space in core routers is larger than necessary, using a tool called The Network Simulator[1], or ns-2 to simulate different network conditions and buffer sizes. Since the fast memory required in core routers is expensive, a large reduction in the required memory would reduce the cost of these routers. Additionally, if the memory requirements in a buffer are reduced, an all-optical router becomes feasible[4].
There have been a number of papers written recently discussing the possibility of significantly reducing the amount of buffer space in core routers.
The rule-of-thumb used to determine the size of the buffer in a router is calculated using the formula B = RTT * C, where RTT is the average round trip time, and C is the data rate of the link[5]. A buffer of this size is thought to allow the router to run at a high utilization. However, it has been argued that utilization is not necessarily the right metric[2], as core routers today run well below 100% utilization.
A router with a small buffer promotes a desynchronized TCP flow, and if a TCP flow is desynchronized, a small buffer size is sufficient[3]. This creates a virtuous cycle. TCP's "sawtooth" congestion control algorithm is designed such that it will fill any buffer[2]. Therefore, a router with a small buffer will drop packets once the buffer is full causing the sender to reduce their sending rate, and allow the buffer to empty.
Using a ns-2, we will be implementing simulations to show that the size of buffers in core routers can be significantly reduced with only a small decrease in utilization.
While further research is still required pertaining to TCP protocol specifics and current router designs, the simulations will attempt to cover a broad range of link speeds and buffer sizes. We will be analysing a number of different network topologies, to determine what the effect on a number of TCP flows flowing through would be.
| Date (2005) | Objective |
|---|---|
| Oct 7 | Submit research proposal. |
| Oct 7 - Oct 21 | Conduct research pertaining to TCP protocols and router design. |
| Oct 22 - Nov 11 | Become familiar with ns-2 tool. Begin implementation of router simulations. |
| Nov 11 | Submit interim report. |
| Nov 11 - Dec 8 | Completion and analysis of router simulations. Preparation of final report. |
| Dec 8 | Submit final report. |
| Dec 8/9 | Give final presentation. |
| [1] | The Network Simulator - ns-2. http://www.isi.edu/nsnam/ns/. |
| [2] | Damon Wischik, Nick McKeown. Buffer Sizes for Core Routers. ACM SIGCOMM Computer Communication Review, 35(3):75-78, July 2005. |
| [3] | Gaurav Raina, Don Towsley, Damon Wischik. Control Theory for Buffer Sizing. ACM SIGCOMM Computer Communication Review, 35(3):79-82, July 2005. |
| [4] | Mihaela Enachescu, Yashar Ganjali, Ashish Goel, Nick McKeown, Tim Roughgarden. Routers with Very Small Buffers. ACM SIGCOMM Computer Communication Review, 35(3):83-90, July 2005. |
| [5] | Guido Appenzeller, Isaac Keslassy, Nick McKeown. Sizing Router Buffers. ACM SIGCOMM 2004, Portland, August 2004. |