Protocol and operating system overheads have become the limiting factor for communications performance on fast networks such as ATM. A large component of these overheads stems fromm the protocol redundancy that arises when layering higher level protocols on top of lower level ones. ATM, for example, requires specific mechanisms for connection management, flow control, congestion avoidance, and segmentation and reassembly. With these mechanisms in place, it is relatively simple and inexpensive to provide for reliable sequenced delivery at the network interface level, making similar functionality in higher-level protocols such as TCP/IP redundant. In this paper we present a protocol architecture specifically tailored for communication over high-bandwith low-latency local or metropolitan ATM networks. Our architecture yields high performance by eliminating protocol redundancy and by exploiting common-case communication behavior. With this approach, we can combine the functionality typically found in four separate layers of the ISO model - data like through session - in a single pass over the data, delivering high throughput and low latency. Our protocol architecture requires minimal hardware support from the network interface and switch fabric, yet effciently provides services such as segmentation and reassembly, flow control, congestion avoidance, and error recovery. We have implemented our protocol architecture on a switch-based ATM network consisting of DECstation 5000/200 workstations running the Mach 3.0 operating system. Our implementation achieves latencies and bandwiths close to the physical limitations imposed by the hardware, yet offers applications a high-level reliable transport interface.