Feedback-based synchronization for QoS traffic in cluster computing

Many applications in cluster computing require performance predictability. One way to ensure that is to generate global schedules and execute the local portion of the schedules at each network interface. With a local clock at each network interface, it is essential that the clocks share a global notion of time. The task of maintaining a single notion of time is called the synchronization problem and this paper addresses it for cluster computing environments. To solve the synchronization problem, FM-QoS proposed a simple notion of synchronization called FBS (Feedback Based Synchronization) for a single switch network. For the generalization of FBS into any networks, this paper extends the basic notion of FBS to a theoretical framework: (1) to identify a set of network flow control signals for synchrony and formalize it in the form of a synchronizing schedule; and (2) to establish the skew analysis model which measures the synchronization precision. Based on the analysis, numerical results are obtained with flow control parameters of Myrinet-1280/SAN. At every observed point, the skew and a network bandwidth overhead of FBS are less than 1.7 /spl mu/s and 5%, respectively, which shows the usability of FBS. With the proposed framework of FBS, this paper concludes that FBS can be an attractive synchronization mechanism in cluster computing environments.