HighSpeed TCP

TCP is limited primarily in the way it controls its window for low bandwidth paths. However, in the face of increasing high speed networks, TCP increasing the window size is not enough. TCP works under Additive Increase Multiplicative Decrease (AIMD) mechanisms which define how the window show grow and shrink when packet losses occur. The primary reason for the window is congestion control. It has been proven that the current congestion avoidance and congestion control algorithms supported by the protocol limit the efficiency in network resource utilization especially on high-speed data paths given the conservative approach adopted in case of congestion.

The varying types of TCP implementation have changed little in terms of how it changes it window size, but instead have relied on extra feedback from the network to achieve better performance. Even though useful, tcp options such as SACKs and ECN are limited by there functionality when dealing with multi Gigabit networks.

Sally Floyd's stack

High Speed TCP [HSTCP] is a modification to TCP's current congestion control mechanisms for high speed links. The current implementations of TCP are limited by the way it controls its congestion windows in the face of losses which result major disruptions in goodput. Based around the inference that a TCP throughput is inversely proportional to the square root of the loss, Sally Floyd proposes changes to the following 'response function' [HSTCP] for tcp;

• w = 1.2/sqrt(p)
  
  Where w is the congestion window size in MSS sized segments and p is the steady state packet drop rate.

Through implementing three parameters Low_Window, High_Window and High_P, she defines a threshold by which TCP will perform differently to standard TCP. When the congestion window of a TCP connection is at most Low_Window, High Speed TCP proposes to use the normal AIMD algorithms; when it is greater than Low_Window, it proposes the use of the HighSpeed response function with alternate values of AIMD. As such, when performing high throughput transfer with HighSpeed TCP, it should be less prone to losses and hence able to maintain a high speed connection. This is directly related to the reduced number of rtts required to the size of the congestion window constant. The direct constants involved with the change in AIMD are related to the values of High_Window and High_P.

One issue with the change to the AIMD algorithms is that HighSpeed TCP may impose a certain degree with unfairness as it does not reduce its transfer rate at much as Standard TCP. Similarly, under congestion control, its slow start can be more aggressive. This is especially pronounced given higher loss rates as HighSpeed TCP will have a larger congestion window and hence able to push more data into the network.

It is our intention to analyse the parameter space of HighSpeed TCP and to perform real life WAN tests over the GridNM infrastructure to judge the advantages of HighSpeed TCP over Standard TCP.

Background

Dynamic Behavior of Slowly-Responsive Congestion Control Algorithms by Deepak Bansal and Hari Balakrishnan. notes.

HighSpeed TCP for Large Congestion Windows by Sally Floyd. notes.
Personal HSTCP Background Details here.

A HighSpeed TCP Study: Characteristics and Deployment Issues by Evandro de Souza and Deb Agarwal

Implementation And Download

We have created a patch against linux 2.4.19 and 2.4.20. Details here. Download here.

Results

Initial Results of UCL->CERN 23 Sept 2002, and UCL->SLAC 27 Sept 2002.

Initial Results on MB-NG

Presentations

Implementing HSTCP @ PPNCG 12th September 2002
Implementing HSTCP via phone conference @ DataTAG meeting 1st October 2002

Papers

TBA

Thu, 7 August, 2003 3:23