Ip layer per-flow queueing and credit flow control

In this thesis we propose a scheme to improve quality of service over the Internet by a new flow control option in the IP layer--FCIP. FCIP routers do per-flow queueing and credit-based flow control. This allows round-robin scheduling and hop-by-hop flow control to be widely available due to the ubiquity of IP. FCIP routers dynamically set up flows and maintain flow soft-state. This implies that FCIP can be transparent to end users. An FCIP test bed has been implemented in a NetBSD UNIX kernel for protocol verification and performance evaluation. Experimental results demonstrate that FCIP reduces data loss, lowers throughput variation, and provides fair bandwidth sharing and queueing delay among flows that use various transport protocols. Adaptive buffer allocation, which was originally designed for ATM, is incorporated in FCIP to allow statistical multiplexing of buffer space. Detailed kernel instrumentation in network modules reveals CPU cost estimation of credit flow control when a general purpose microprocessor is used. Results suggest that an FCIP PC router is capable of handling the load at campus Internet gateways. This conclusion is based on analysis of load, simultaneous flow count, and flow arrival rate.