Skip to Main Content
FM-QoS employs a novel communication architecture based on network feedback to provide predictable communication performance (e.g. deterministic latencies and guaranteed bandwidths) for high speed cluster interconnects. Network feedback is combined with self-synchronizing communication schedules to achieve synchrony in the network interfaces (NIs). Based on this synchrony, the network can be scheduled to provide predictable performance without special network QoS hardware. We describe the key element of the FM-QoS approach, feedback-based synchronization (FBS), which exploits network feedback to synchronize senders. We use Petri nets to characterize the set of self-synchronizing communication schedules for which FBS is effective and to describe the resulting synchronization overhead as a function of the clock drift across the network nodes. Analytic modeling suggests that for clocks of quality 300 ppm (such as found in the Myrinet NI), a synchronization overhead less than 1% of the total communication traffic is achievable -- significantly better than previous software-based schemes and comparable to hardware-intensive approaches such as virtual circuits (e.g. ATM). We have built a prototype of FBS for Myricom s Myrinet network (a 1.28 Gbps cluster network) which demonstrates the viability of the approach by sharing network resources with predictable performance. The prototype, which implements the local node schedule in software, achieves predictable latencies of 23 µs for a single-switch, 8-node network and 2 KB packets. In comparison, the best-effort scheme achieves 104 µs for the same network without FBS. While this ratio of over four to one already demonstrates the viability of the approach, it includes nearly 10 µs of overhead due to the software implementation. For hardware implementations of local node scheduling, and for networks with cascaded switches, these ratios should be much larger factors.