Decoupled scheduling in Store-and-Forward OCS networks

Abstract

By introducing storage into Optical Circuit Switched (OCS) networks and storing delay-tolerant requests until off-peak hours, the networks can have better performance. Since scheduling a data transfer in SnF networks has to deal with the spatial and temporal arrangement of the request, it is much more complicated than the routing-alone problem. In our prior research, we proposed TS-MLG, a holistic approach to solve this problem, but the computational and spatial complexity of the approach is high, limiting its use in large scale networks. In this paper, we propose a decoupled solution and specific algorithms for both sub-problems. The solutions can reduce spatial and computational complexity from O(N²*L²) to O(N*L) with marginal sacrifice on network performance, in which N and L are respectively the number of network nodes and state changes in the network. Simulation results show that in a moderate sized network with frequent state changes, the computing time of request scheduling can be reduced by 600 times. We conclude that the decoupled SnF scheduling has the potential of analyzing SnF in large networks with frequent state changes.

Introduction

Data with relatively large sizes, often called bulk data, contributes to 90% of overall data traffic over the Internet. The transferring of bulk data takes up much network resources (e.g. link bandwidth) for a long period, leading to increased and sometimes unpredictable delay for data with smaller sizes. Since bulk data is often delay insensitive, the problem can be partially solved by putting off its transferring process. This can be realized by a so-called Store-and-Forward (SnF) mechanism, in which data are stored at intermediate storage nodes when the traversed network is busy, and re-scheduled at times when the network becomes less congested [1].

However, SnF consequently introduces a new challenge - determining where and when the data should be stored, and when data should resume transmission, thus adds one additional domain of complexity into the route-alone problem in multiple wavelength OCS networks. Therefore, request scheduling in SnF OCS networks are in fact the combination of request routing and storage scheduling problems. Since the computational complexity of the original route-alone problem has already been proved to be NP-hard, the combined problem has a very high level of complexity. Therefore, solutions to the SnF problem which offer integrated approaches, which most existing solutions choose to offer, always have high complexity. This leads to restrictions on SnF mechanisms to be used in large-scale networks. To make the concept of SnF realistic in such networks, an easy to use routing and scheduling heuristic tool is necessary. The tool should be able to maintain intuitive presentations and achieve a certain level of overall performance, with a reasonable computational overhead. In this case, decoupling will be a better thought of reducing model complexity.

In this paper, we first propose to decouple the SnF problem into sub-problems of request routing and storage scheduling. We then offer specific solutions for the two sub-problems, by applying offline shortest path algorithms on planar graphs for request routing, and recursion based computation on storage scheduling. The spatial and computational complexity of the SnF problem can reduce from O(N²*L²) to O(N*L), in which N and L are respectively the number of network nodes and state changes, with marginal sacrifice on network performance. Simulation results show that the proposed solution results in a 600 times reduction of request scheduling time. With decoupling, the spatial and computational complexity of the SnF problem can reach a far lower level, indicating that the decoupled SnF scheduling has the potential of analyzing large networks with frequent state changes.

This paper is organized as follows: we discuss the related work in Section II. We decouple SnF problem into request routing and storage scheduling, and propose solutions to both sub-problems in Section III. In Section IV, we compare the proposed solution to the original one, and analyze their advantages and drawbacks. In Section V, we conduct simulations to show and compare the performance improvements that may be achieved by the new solution. We conclude this paper in Section VI.