Aggregated Aggressiveness Control on Groups of TCP Flows Soohyun Cho and Riccardo Bettati Computer Science Department, Texas A&M University, College Station, TX 77843 USA {s0c6496, bettati}@cs.tamu.edu Abstract. The use of multiple concurrent parallel TCP flows is an easy way to achieve higher speed reliable data transfers. However, parallel TCP flows are inherently unfair with respect to single TCP flows. We suggest a new scheme called TCP-P, which controls aggressiveness of a group of parallel TCP flows by regulating their total aggressiveness (or unfairness) to be comparable to a single TCP flow, or any multiple thereof. TCP-P makes a group of N parallel TCP flows appear to other flows like k separable TCP flows - i.e., have strength k - through appro- priate manipulations of increase and decrease behavior of the congestion windows of the TCP flows in the group. We implemented our scheme as part of Linux and experimental results show that the proposed scheme effectively controls aggressiveness of parallel TCP flows. 1 Introduction A widely used scheme to work around the limitations of TCP over high delay- bandwidth product connections is to use multiple parallel TCP connections. The use of parallel TCP flows has several benefits compared to a single TCP flow [1]. If an end-host opens N parallel TCP flows to the same destination, its congestion window recovery and increase are N times faster than a single TCP flow [2]. As a result, the achievable throughput of parallel TCP flows is significantly bigger than that of a single TCP flow given the same packet loss probability. Unfortunately, this increase comes at the expense of the throughput experienced by other, single TCP flows, as sender nodes who open multiple parallel TCP flows will consume unfairly more bandwidth when they compete for the same bottleneck links. With the increased venues for bundling of TCP flows (e.g., overlay networks with TCP splicing [3], large servers with topological aggregation of service deliv- ery, dedicated connections between supercomputers or campuses, etc.,) flexible schemes are needed for the controllable aggregation of large numbers of parallel TCP flows. Naively limiting the number of parallel connections that applica- tions in a node can open concurrently is not appropriate in many situations, as it violates the separation of application design from network resource allocation: Making the number of available connections visible to the application unduly R. Boutaba et al. (Eds.): NETWORKING 2005, LNCS 3462, pp. 586–597, 2005. c  IFIP International Federation for Information Processing 2005