Design, implementation, and performance of a native mode ATM transport layer

We describe the design, implementation, and performance tuning of a transport layer targeted specifically for ATM networks. The layer has been built from scratch to minimize overhead in the critical path and take advantage of ATM adaptation layer 5 functionality. It provides reliable or unreliable data delivery with feedback or leaky-bucket flow control. These services can be combined to create a customized transport service. Our work is novel in that it is the first end-to-end ATM transport service to provide reliable, flow controlled data transfer. We describe the mechanisms and the operating system support needed to provide these services. A detailed performance measurement allows us to determine the bottlenecks in our system and do tune our implementation. With this tuning, we are able to achieve a user-to-user throughput of 55 Mbps between two 66 MHz Intel 80486 personal computers with Fore Systems' HPA-200 EISA-bus host adaptors. The user-to-user latency for small messages is around 720 /spl mu/s. These figures compare favorably with the performance from far more expensive workstations and validate the correctness of our design choices.