Joint provisioning of lightpaths and storage in store-and-transfer wavelength-division multiplexing networks

摘要

So-called store-and-transfer wavelengthdivision multiplexing networks (STWNs) can store data and provision lightpaths at an optimal time when the wavelengths are clear of conflicts. If the wavelength utilization of the wavelength-division multiplexing network is not uniform, the embedded storage nodes can store data and the lightpath is segmented so conflicts along separate segments of the lightpath can be resolved independently. Consequently, blocked requests can be reduced, and the resource utilization of the network can be improved. In this work, we investigate three extensions of the sliding scheduled traffic model, which is a variant of advance reservations. First, we segment a lightpath with embedded storage nodes, split the transfer window into smaller ones, and sequentially assign a smaller window to each segment of the lightpath. Each segment of the lightpath has fewer hops, and thus, overall blocking can be decreased. Second, we allocate fixed time slots in continuous frames over multiple wavelengths for a time-division multiplexing (TDM) mode large data transfer. The resource allocations are more flexible in resolving conflicts, and thus, blocking can be decreased. Third, we use segmented lightpaths and TDM mode transfer along each segment of the lightpath. Since the first two extensions are orthogonal, the combination can decrease the blocking further. We propose four heuristic algorithms to evaluate the extensions. We also propose two parallel implementations of the four heuristic algorithms. The simulations show the following: the provisioning algorithms for the first two extensions have similar performances, and both decrease blocking by 20%. If the service time is long, the second one performs better. The provisioning algorithm for the last extension decreases blocking by 40%. The first extension requires 10% more storage, and the second extension requires 20% less storage. Under a large load and with 8 CPUs, the parallel implementations improve the speed by three times. It also shows that the number of time slots per frame influences the performance; the influence is not monotonic; and, if the value increases too much, the blocking rate increases by 20% and the storage size increases by 30%.