تجزیه و تحلیل عملکرد مقایسه ای کنترل جریان هدایت شده برای زمان واقعی SCI

The distributed nature of routing and flow control in a register-insertion ring topology complicates priority enforcement for real-time systems. Two divergent approaches for priority enforcement for ring-based networks are reviewed: a node-oriented scheme called preemptive priority queue and a ring-wide arbitration approach dubbed TRAIN. This paper introduces a hybrid protocol named directed flow control that combines node- and ring-oriented flow control to yield greater performance. A functional comparison of the three protocols as implemented on the scalable coherent interface is presented, followed by performance results obtained through high-fidelity modeling and simulation.

The information revolution has created a world in which data is a valuable commodity. In the past, the worth of a network was judged by its ability to deliver data reliably and rapidly. With the emergence of a commercial market for video- and audio-on-demand and an increasing reliance on computer control of mission-critical applications in the military and industrial sectors, networks must also deliver data within a guaranteed maximum response time. As the number of users in the global network grows, the capacity and performance requirements for such real-time networks are rapidly increasing. The combination of a high-bandwidth, low-latency network protocol with real-time capabilities is necessary to meet these needs. The protocols in this study implement real-time capabilities on a ring-based network. In addition to the cost-effectiveness of ring-based topologies for large-scale networks, chip-multiprocessor architectures frequently include rings as the on-chip interconnect [5]. Though the protocols in this paper are generically applicable to any ring-based network, the scalable coherent interface (SCI) is used as the underlying transport [8]. As IEEE Standard 1596-1992, SCI provides a bus-like interface over unidirectional, point-to-point links to implement distributed shared memory. The protocol calls for link speeds of 1 GB/s and sub-microsecond latencies over distances of tens of meters. However, the base SCI protocol does not attempt to address real-time issues. Providing real-time capability in a network requires enforcement of packet priorities. Such a priority scheme is required for implementation of algorithms such as rate-monotonic scheduling (RMS) [10]. To support RMS, all packets are assigned a priority and the network must enforce transmission of higher-priority packets ahead of those with lower priorities. Two protocols were originally proposed to provide real-time capabilities for SCI: preemptive priority queue (PPQ) and TRAIN. Simulation shows neither solution to be optimal in sufficient numbers of test cases to be acceptable for a generalized standard. This paper presents a third solution to the SCI/RT problem called directed flow control (DFC). DFC is a hybrid that combines selected aspects of the PPQ and TRAIN protocols to provide performance greater than either protocol alone. Traditionally, real-time protocols for ring-based networks have taken a token-passing approach [2] and [16]. However, such a scheme is unable to capitalize on the concurrent bandwidth that inherently exists in a register-insertion ring network. Other flow-control schemes allow the concurrent bandwidth to be used efficiently but do not adequately address real-time performance needs [11] and [12]. The real-time protocols presented in this paper provide for both concurrent communication to make effective use of available ring bandwidth and priority-level enforcement for real-time communications. This paper begins with an overview of the requirements for a real-time extension to SCI. Next, the three protocols are presented from a functional standpoint, illustrating the novel ways in which each protocol addresses priority inversions in a ring-oriented network. The performance of the protocols is then compared using simulation results from several common test cases.

In this paper, we have presented three protocols that provide priority-based, real-time networking capabilities to SCI. The PPQ protocol takes a node-oriented approach to flow control while the TRAIN protocol employs ring-wide arbitration. The DFC protocol joins the two methods by providing a lightweight form of ring-oriented arbitration with state-based, node-centric decisions for flow control. High-fidelity simulation has shown the DFC protocol to have low latencies comparable to those of base SCI and the PPQ protocol, coupled with the highest throughputs of any of the approaches in a variety of case studies. To improve the performance of the PPQ and TRAIN protocols, several modifications have recently been proposed. These additions include mechanisms to improve the throughput of PPQ under high loadings or long IQST delays by limiting the busy-retries, and a “gearshift” mechanism for the TRAIN protocol to allow a TRAIN system to operate exactly like base SCI under low loadings and switch into arbitration mode as needed. The design, specification and analysis of these and other potential enhancements are the subject of future study. While the DFC protocol has superior performance, the hardware complexity and storage space required to implement the protocol may make a modified version of the PPQ protocol a more attractive option. Also, the ways in which DFC can operate in a switched network environment or interoperate with base SCI are as yet not clear. The PPQ protocol offers a more straightforward approach for switching and base SCI interoperability. Future research should also consider other factors that would affect the real-time performance. Example areas of interest include larger ring sizes, non-ideal link and forwarding delays, and extensions to support more complex topologies such as direct networks based on multi-dimensional tori and indirect networks involving switches. Integrating real-time nodes with traditional SCI hardware as well as other real-time networks is also an avenue for future investigation. As often happens in the real world, the optimum solution is probably not to be found in either extreme of a strictly node-oriented protocol such as PPQ or in an arbitrated ring-oriented protocol like TRAIN. Instead, a mixture of the two approachs can be taken with careful attention to minimizing overhead to yield a superior solution. One such promising approach is DFC. Acknowledgements This work was supported in part by the Naval Air Warfare Center (NAWC) and the Office of Naval Research (ONR), in concert with the IEEE P1596.6 Working Group. In addition, the authors acknowledge Mr. Ralph Lachenmaier at NAWC for insightful ideas and feedback in the development of the DFC protocol and the anonymous reviewers for helping improve the presentation of the content in this paper.