MAUI RINKO (Feb 22 2001) by haruhito Reasons net to deploy RED AbstractF In this paper authors discussed the benefits of RED (Random early detection). Authors made a testbed network by 2 Cisco 7500, and we measured the performance. We found 2 results; - RED with small buffers does not improve significantly the performance. - RED has no big impact on the end-to-end performance. We recommended that we should have more research with realistic network to develop a full quantitative understanding of RED. INTRODUCTION - We built a testbed network - Cisco IOS 12.0 RED implementation and Chariot 2.2 load generator - We checked the relationship between RED parameter choice and end-to-end transmission performance. - Parameter choice in RED is hard -However, according to [5], any parameter setting will improve the system performance. EXPERIMENTAL PLATFORM - Chariot load generator - It can simulate real network traffic. (Other simurators cannot) - Magage and synchronize a whole set of traffic conections on different endpoints. - Similates FTP connections, NFS traffic, HTTP traffic, real-time audio/visual traffic. - 80% TCP connections (FTP and HTTP), 20% UDP flows PARAMETER SETTING FOR RED WITH ONE ROUTER - We evaluate performance of RED, and compare with tail drop. - IOS allows to set 4 parameters; - min threshold (minth), max threshold (maxth), max drop probability (maxp), averaging parameter (alpha) - We chenged maxp, the size of dropping interval (maxth-minth) and alpha for evaluation. - We measured 3 valus; - the average throuput, the number of bytes sent, UDP pachets lost rate The maximum drop probability - We changed maxp. - We had 2 experiments; - For the small buffer. size=40packets. This is default value for Cisco. - For the large buffer, size=200packets. - Result - For the small buffer, throughput is not changing significantly. - For the large buffer, TCP performance is changed, but it is not a simple function of the setting of maxp. - UDP drop rate is improved (compere to Tail Drop). Varying the size of the dropping interval (maxth-minth) - We changed (maxth-minth). - - We had 2 experiments; - For the small buffer. size=40packets. This is default value for Cisco. - For the large buffer, size=200packets. - Result - (maxth-minth) does not influence the system performance. (TCP and UDP) - UDP drop rate is improved (compere to Tail Drop). Fairness - We examined the fairness of RED vs TailDrop. - Result - No big difference between RED and TailDrop. - Both are not fair. - In RED, some results were very bad. This might be due to the fact that RED increases the probability that a packets get dropped. If this occurs during the slow start phase it is very hard to recover. Increase the buffer size - We changed the buffer size. - Result - The number TCP packets sent is increasing when we increase the buffer size. - UDP drop rate is improved when We increase the buffer size. - UDP drop rate is much higher with TD then with RED. Impact of the averaging parameter alpha - We examine how different averaging settings impact the performance of RED. - The authors of RED said that the averaging will avoid the bias against bursty traffic. - Result - We got more drops for bursty traffic with RED. This is an opposite result. - The influence of the weighting factor is minimal. THE MULTIPLE ROUTER CASE - In this section we discuss the problems that occur when RED is used in an existing network --- multi router, multi ISP environment. - We added a second router in the testbed network. - We examined the end-to-end performance of the flows traversing two routers. - We had 2 experiments; - For the small buffer. size=40packets. This is default value for Cisco. - For the large buffer, size=200packets. - Result - For the small buffer, the performance is improved by pure RED network. But RED-TD network had worst result. - For large buffer, pure TD network was best. - Parameter tuning for multi RED network is very important. Misconfiguration or bad configuration will degrade the performance of the whole network. esp. UDP throughput. DISCUSSION - RED does not exhibit much better performance than Tail Drop. - The RED parameters have a minor impact on the performance with small buffer. With large buffer, RED can improve the performance. - A static RED cannot provide better results than TD. - We should use more sophisticated active buffer management schemes. - We need to have more experiments on larger testbet network.