Networks are the well-known bottlenecks for distributed deep learning (DDL) jobs. The DDL jobs require topologies matching their communication patterns as well as high and stable bandwidth in network transmission, which makes the traditional optimization methods in the field of electrical packet switching (EPS) perform poorly when adapting to the DDL jobs. In contrast, optical circuit switching (OCS) has the natural advantages of topology reconstruction and high bandwidth, which can well fulfill the network requirements and reduce the in-network delay of the DDL jobs. Consequently, in the present work, we propose Optical Slicing provisioning in support of DDL jobs (OSDL). Practically, over an EPS + OCS hybrid network, to match the network topology and the traffic patterns of DDL jobs, we comprehensively design and propose corresponding algorithms for job placement, and scheduling of both OCS and EPS. We first evaluate OSDL with large-scale networks and high-density DDL jobs, by simulations. The simulation results show that OSDL outperforms multiple well-known scheduling methods. More specifically, the speedup of OSDL achieves up to 3.31 times and 10.48 times, over a homogeneous network and a heterogeneous network, respectively. Additionally, we conduct experiments with a relatively small-scale network and relatively low-density DDL jobs, we still observe up to 1.74 times of the speedup. The simulations and the experiments verify the effectiveness of OSDL in accelerating DDL jobs.

The continuous growth of the traffic at a rapid pace raises challenges related to the efficient use of resources and saving of energy in the transport network. The new traffic pattern is characterized by faster growth of metro traffic compared with core, due in part to video content replication. In this paper, we compare the resource requirements of three different transport technologies (optical packet switching, opaque circuit switching and transparent circuit switching) within two architecture scenarios (the current hub-and-spoke architecture and the next generation distributed architecture). The comparison study takes into consideration the predicted traffic growth from 2014 to 2019 and the potential traffic decentralization due to content replication. For the optical packet switching, we focus on Time-domain Wavelength Interleaved Network (TWIN) solution which achieves passive switching at the intermediate nodes thanks to the precomputed optical packet emission schedule. Results show that TWIN outperforms the other transport technologies in terms of number of transmitters and receivers with a high ability to cope with scalability issues.

ITU-T Recommendation G.694.1 defines normative DWDM frequency grids, each being a reference set of values that correspond to allowed nominal central frequencies, obtained using a fixed channel spacing (e.g., 12.5 GHz, 25 GHz, 50 GHz or 100 GHz). This rigid, grid-based approach does not seem well adapted for data rates beyond 100 Gb/s, is particularly inefficient when a whole wavelength is assigned to a lower rate optical path, and is not flexible enough for multi-rate systems . Consequently, the next generation of optical networks will require a flexible, highly efficient and adaptive management of the optical spectrum, along with advanced optical modulation schemes that efficiently use allocated spectrum slots, and recent progress on optical network technology justifies research on both new optical transmission systems as well as the applicability of control and management frameworks to such networks. We design and deploy a GMPLS control plane for flexible optical networks with coherent optical orthogonal frequency-division multiplexing (O-OFDM) transmission; we detail its functional architecture, which combines a centralized entity that performs path routing and modulation assignment, with a distributed spectrum allocation. The centralized entity (i.e., a path computation element) uses pre-configured static path characterizations, based on exhaustive OFDM transmission simulations, when performing dynamic path computation in line with GMPLS constrained shortest path mechanisms. The distributed spectrum allocation assigns frequency ranges (slots) to connection requests, by using dynamic signaling procedures and applying slot assignment policies. We summarize the control plane protocol extensions involved in the main functional aspects: routing and topology dissemination, path computation, signaling and resource reservation. We experimentally validate and evaluate the integrated centralized PCE and GMPLS control plane in a control plane testbed, obtaining key performance indicators such as path setup latency and blocking probability for different frequency slot assignment policies.

A testbed of OBS/GMPLS interworking network is established with the supporting of the BUPT-KDDI cooperation project. A dedicated GMPLS border controller with necessary GMPLS extensions and some OBS extensions are introduced in this network to achieve a dynamic, efficient and transparent inter-domain