Automatic TCP Buffer Tuning
Even when TCP performance enhancements such as RFC 2018 SACK and RFC 1323 large window extensions are used, a TCP connection may still not achieve the optimal performance expected by the user. One reason for suboptimal performance is that the connection may be buffer limited by the sender or the receiver, particularly for large bandwidth*delay paths.
Manual configuration eases the problem somewhat, but requires information that is often not available to the user or system administrator, and is not flexible enough to account for changes in the network or the end system.
We propose an Automatic TCP Buffer Tuning scheme to alleviate these problems. It has several main advantages:
- Each connection gets the best possible performance
- It is transparent to the user and the application
- Memory is fairly shared when many connections are in use, allowing a large number of simultaneous connections for a single host
The following items are available from our research:
- A technical report describing changes to the implementation of the fair share algorithm.
- A technical report describing a kernel monitoring system that was developed to study the effects of autotuning.
- A paper describing autotuning, presenting test results from an actual implementation.
- A slide presentation, given at SIGCOMM ‘98
- An earlier version of the presentation, given at an NCNE online training talk, which contains a small amount of tutorial information not present in the SIGCOMM presentation.
- tar files containing our Autotuning implementation. (The tar file also contains our implementations of SACK and FACK, which can be disabled.)