Evaluation of TCP Congestion Control Algorithms on the Windows Vista Platform
CTCP, an innovative TCP congestion control algorithm developed by Microsoft, is evaluated and compared to HSTCP and StandardTCP. Tests were performed on the production Internet from Stanford Linear Accelerator Center (SLAC) to various geographically located hosts to give a broad overview of the performances. We find that certain issues were apparent during testing (not directly related to the congestion control algorithms) which may skew results. With this in mind, we find that CTCP performed similarly to HSTCP across a multitude of different network environments. However, to improve the fairness and to reduce the impact of CTCP upon existing StandardTCP traffic, two areas of further research were investigated. Algorithmic additions to CTCP for burst control to reduce the aggressiveness of its cwnd increments demonstrated beneficial improvements in both fairness and throughput over the original CTCP algorithm. Similarly, {gamma} auto-tuning algorithms were investigated to dynamically adapt CTCP flows to their network conditions for optimal performance. While the effects of these auto-tuning algorithms when used in addition to burst control showed little to no benefit to fairness nor throughput for the limited number of network paths tested, one of the auto-tuning algorithms performed such that there was negligible impact upon StandardTCP. With these improvements, CTCP was found to perform better than HSTCP in terms of fairness and similarly in terms of throughput under the production environments tested.
- Research Organization:
- SLAC National Accelerator Lab., Menlo Park, CA (United States)
- Sponsoring Organization:
- USDOE
- DOE Contract Number:
- AC02-76SF00515
- OSTI ID:
- 885508
- Report Number(s):
- SLAC-TN-06-005; TRN: US0604208
- Country of Publication:
- United States
- Language:
- English
Similar Records
Characterization and Evaluation of TCP and UDP-based Transport on Real Networks
Evaluation of Advanced TCP Stacks on Fast Long-Distance Production Networks