Multidimensional and Reconfigurable Optical Interconnects for High-Performance Computing (HPC) Systems

The increasing demand for higher communication bandwidth, reduced power consumption, and increased reliability combined with fundamental electrical signalling limitations is leading the drive for optics as an interconnect technology of choice for high-performance computing (HPC) systems. However, failure in any optical link can completely disrupt communication by isolating processing nodes in HPC systems. Moreover, while static allocation of wavelengths (channels) provides every node with equal opportunity for communication, it can also lead to network congestion for nonuniform traffic patterns. In this paper, we propose a multidimensional optoelectronic architecture, called nD-reconfigurable, all-photonic interconnect for distributed and parallel systems (ndimensional-RAPID) where n can be 1, 2, or 3. nD-RAPID exploits optical architecture and technology design space that simultaneously tackles both fault-tolerance and dynamic bandwidth reallocation (DBR) of system architecture. Fault-tolerance in nD-RAPID is enabled through a multidimensional architecture. DBR is implemented by the row-column switching matrix using silicon-on-insulator (SOI)-based microring resonators that adapts to changes in communication patterns at runtime. Simulation results indicate that nD-RAPID outperformed other electrical networks for most traffic patterns. Results on DBR show that the proposed row-column switch organization significantly improves throughput and latency with a slight increase in electrical power consumption (~ 0.4% for the worst case traffic).     

A practical approach for statistical matching of output queueing

We study a practical approach to match the performance of an output-queued switch statistically. For this purpose, we propose a novel switching architecture called a multiple input/output-queued (MIOQ) switch that requires no speedup for providing sufficient switching bandwidth. To operate an MIOQ switch in a practical manner, we also propose a multitoken-based arbiter which schedules the switch at a high operation rate and a virtual first-in first-out queueing scheme which guarantees the departure order of cells belonging to the same traffic flow at output. Additionally, we show that the proposed switch can naturally provide asymmetric bandwidth for inputs and outputs, which may be important in dealing with the links with different bandwidth demands. Finally, we compare the performance of an MIOQ switch with that of an output-queued switch and discuss the design criteria to match the performance of an output-queued switch.     

Multi-stage switching system using optical WDM grouped links based on dynamic bandwidth sharing

A three-stage Clos switch architecture is attractive because of its scalability. From an implementation point of view, it allows us to relax the cooling limitation, but there is a problem interconnecting different stages. This article presents a three-stage switching system that uses optical WDM grouped links and dynamic bandwidth sharing. We call it a WDM grouped-link switch. The introduction of WDM makes the number of cables used in the system proportional to the switch size. Dynamic bandwidth sharing among WDM grouped links prevents the statistical multiplexing gain offered by WDM from falling even if the switching system becomes large. The WDM grouped-link switch uses cell-by-cell wavelength routing. A performance evaluation confirms the scalability and cost-effectiveness of this switch. An implementation of the WDM grouped link and a compact planar lightwave circuit platform is described. This architecture allows us to expand the throughput of the switching system up to 5 Tb/s.     

Sunshine: a high performance self-routing broadband packet switcharchitecture

The authors present a high-performance self-routing packet switch architecture, called Sunshine, that can support a wide range of services having diverse performance objectives and traffic characteristics. Sunshine is based on Batcher-banyan networks and achieves high performance by utilizing both internal and output queuing techniques within a single architecture. This queuing strategy results in an extremely robust and efficient architecture suitable for a wide range of services. An enhanced architecture allowing the bandwidth from an arbitrary set of transmission links to be aggregated into trunk groups to create high bandwidth pipes is also presented. Trunk groups appear as a single logical port on the switch and can be used to increase the efficiency of the switch in an extremely bursty environment or to increase the access bandwidth for selected high-bandwidth terminations. Simulation results are presented     

Tunable optical filters for dense WDM networks

WDM is currently taking over as the leading technology in point-to-point transmission links. For optical implementation of WDM networks, logical functionalities such as wavelength (channel) selection should be carried out in the wavelength domain; thus, the development of dynamic optical devices is required. One key device is a tunable optical filter. Important features of such a filter include low insertion loss, narrow bandwidth, high sidelobe suppression, large dynamic range, fast tuning speed, a simple control mechanism, small size, and cost effectiveness. Here, an extensive overview of the different technologies used to produce tunable optical filters is presented. Among them, fiber filters such as fiber Bragg gratings and fiber Fabry Perot are the most commercialized, yet inherently limited in their dynamic speeds. For high demanding dynamics, micro-machined and acousto-optic filters can offer a good solution for microsecond tuning speeds. Faster tunable devices, in nanosecond tuning speeds, might emerge out of microresonators, electrooptic filters, and active DBR filters     

A fast packet switch for the integrated services backbone network

With the projected growth in demand for bandwidth and telecommunication services will come the requirement for a multiservice backbone network of far greater efficiency, capacity, and flexibility than ISDN (integrated-services digital network) is able to satisfy. This class of network has been termed the broadband ISDN, and the design of the switching nodes of such a network is the subject of much research. The author investigates one possible solution. The design and performance, for multiservice traffic, is presented for a fast packet switch based on a nonbuffered, multistage interconnection network. It is shown that for an implementation in current CMOS technology, operating at 50 MHz, switches with a total traffic capacity of up to 150 Gb/s can be constructed. Furthermore, if the reserved service traffic load is limited on each input port to a maximum of 80% of switch port saturation, then a maximum delay across the switch of on the order of 100 μs can be guaranteed, for 99% of the reserved service traffic, regardless of the unreserved service traffic load     

Pure frequency modulation of a multielectrodedistributed-Bragg-reflector (DBR) laser diode

Intensity-modulation suppressed frequency modulation (FM) is demonstrated over a 600-MHz bandwidth by using a three-electrode distributed-Bragg-reflector (DBR) laser diode (LD). A nonuniform carrier-density distribution in the active region is realized by two separated electrodes, whose push-pull modulation yields only a 0.7% intensity fluctuation per 1-GHz frequency deviation. This corresponds to a 20-dB FM purity improvement. A third electrode on the DBR region makes it possible to tune the wavelength simultaneously. Because of its wider modulation bandwidth, the LD is suitable for the lightwave oscillator in frequency-shift-keyed (FSK) transmissions and for linewidth reduction by electrical feedback     

FEBA: A Bandwidth Allocation Algorithm for Service Differentiation in IEEE 802.16 Mesh Networks

In wireless mesh networks, the end-to-end throughput of traffic flows depends on the path length, i.e., the higher the number of hops, the lower becomes the throughput. In this paper, a fair end-to-end bandwidth allocation (FEBA) algorithm is introduced to solve this problem. FEBA is implemented at the medium access control (MAC) layer of single-radio, multiple channels IEEE 802.16 mesh nodes, operated in a distributed coordinated scheduling mode. FEBA negotiates bandwidth among neighbors to assign a fair share proportional to a specified weight to each end-to-end traffic flow. This way traffic flows are served in a differentiated manner, with higher priority traffic flows being allocated more bandwidth on the average than the lower priority traffic flows. In fact, a node requests/grants bandwidth from/to its neighbors in a round-robin fashion where the amount of service depends on both the load on its different links and the priority of currently active traffic flows. If multiple channels are available, they are all shared evenly in order to increase the network capacity due to frequency reuse. The performance of FEBA is evaluated by extensive simulations. It is shown that wireless resources are shared fairly among best-effort traffic flows, while multimedia streams are provided with a differentiated service that enables quality of service.     

A self-timed real-time sorting network

High-speed networks are expected to carry traffic classes with diverse quality of service (QoS) guarantees. For efficient utilization of resources, sophisticated scheduling protocols are needed; however, these must be implemented without sacrificing the maximum possible bandwidth. This paper presents the architecture and implementation of a self-timed real-time sorting network to be used in packet switches that support a diverse mix of traffic. The sorting network receives packets with appropriately assigned priorities and schedules the packets for departure in a highest-priority-first manner. The circuit implementation uses zero-overhead, self-timed, and self-precharging domino logic to minimize the circuit latency. An experimental sorting network chip has been designed using the techniques described in this paper to support 10 Gb/s links with ATM-size packets     

High-speed signal wavelength conversion using aunidirectional-output wavelength conversion device withasymmetric-κ DBR structure

The wavelength conversion device whose frequency response is evaluated analytically and experimentally in this work has a saturable absorber region, which acts as an optical gate, within the active region. The 3-dB bandwidth of this device was measured to be 800 MHz and found to be limited by the frequency response of the saturable absorber region. To speed up this device, the electrodes of the saturable absorber and the gain regions were connected electrically so that the device would act as a unidirectional-output distributed Bragg reflector (DBR) laser. When the active region (saturable absorber and the gain regions) of this DBR laser were biased above threshold condition, the 3-dB bandwidth was more than 10 GHz, and clear eye patterns were observed when the input TM-polarized light intensity was modulated by a 10-Gbit/s pseudo-random NRZ signal. These results show that high-speed wavelength conversion can be achieved when the device is operated as a laser diode     

Courtesy piggybacking: supporting differentiated services in multihop mobile ad hoc networks

Due to the salient characteristics such as the time-varying and error-prone wireless links, the dynamic and limited bandwidth, the time-varying traffic pattern and user locations, and the energy constraints, it is a challenging task to efficiently support heterogeneous traffic with different quality of service (CoS) requirements in multihop mobile ad hoc networks. In the last few years, many channel-dependent mechanisms are proposed to address this issue based on the cross-layer design philosophy. However, a lot of problems remain before more efficient solutions are found. One of the problems is how to alleviate the conflict between throughput and fairness for different prioritized traffic, especially how to avoid the bandwidth starvation problem for low-priority traffic when the high-priority traffic load is very high. In this paper, we propose a novel scheme named Courtesy Piggybacking to address this problem. With the recognition of interlayer coupling, our Courtesy Piggybacking scheme exploits the channel dynamics and stochastic traffic features to alleviate the conflict. The basic idea is to let the high-priority traffic help the low-priority traffic by sharing unused residual bandwidth with courtesy. Another noteworthy feature of the proposed scheme is its implementation simplicity: The scheme is easy to implement and is applicable in networks using either reservation-based or contention-based MAC protocols.     

Evaluation and characterization of available bandwidth probing techniques

The packet pair mechanism has been shown to be a reliable method to measure the bottleneck link capacity on a network path, but its use for measuring available bandwidth is more challenging. In this paper, we use modeling, measurements, and simulations to better characterize the interaction between probing packets and the competing network traffic. We first construct a simple model to understand how competing traffic changes the probing packet gap for a single-hop network. The gap model shows that the initial probing gap is a critical parameter when using packet pairs to estimate available bandwidth. Based on this insight, we present two available bandwidth measurement techniques, the initial gap increasing (IGI) method and the packet transmission rate (PTR) method. We use extensive Internet measurements to show that these techniques estimate available bandwidth faster than existing techniques such as Pathload, with comparable accuracy. Finally, using both Internet measurements and ns simulations, we explore how the measurement accuracy of active probing is affected by factors such as the probing packet size, the length of probing packet train, and the competing traffic on links other than the tight link.     

Guaranteed frame rate: a better service for TCP/IP in ATM networks

Guaranteed frame rate, approved by the ATM Forum, is expected to become an important service category to efficiently support TCP/IP traffic in ATM networks. We first describe the GFR traffic contract in detail. We then present different types of switch implementations that have been proposed to support GFR. We analyze the performance of three of these switch implementations by simulations in two different network environments. These simulations show that the scheduler-based implementations provide a much better performance than the simple switch implementation. However, we also show that coupling an active packet discard mechanism to a scheduler-based switch implementation does not produce a performance gain when many TCP connections are multiplexed inside one ATM VC     

Deficit round-robin scheduling for input-queued switches

We address the problem of fair scheduling of packets in Internet routers with input-queued switches. The goal is to ensure that packets of different flows leave a router in proportion to their reservations under heavy traffic. First, we examine the problem when fair queuing is applied only at output link of a router, and verify that this approach is ineffective. Second, we propose a flow-based iterative deficit-round-robin (iDRR) fair scheduling algorithm for the crossbar switch that supports fair bandwidth distribution among flows, and achieves asymptotically 100% throughput under uniform traffic. Since the flow-based algorithm is hard to implement in hardware, we finally propose a port-based version of iDRR (called iPDRR) and describe its hardware implementation.     

Minimizing Internal Speedup for Performance Guaranteed Switches With Optical Fabrics

We consider traffic scheduling in an N times N packet switch with an optical switch fabric, where the fabric requires a reconfiguration overhead to change its switch configurations. To provide 100% throughput with bounded packet delay, a speedup in the switch fabric is necessary to compensate for both the reconfiguration overhead and the inefficiency of the scheduling algorithm. In order to reduce the implementation cost of the switch, we aim at minimizing the required speedup for a given packet delay bound. Conventional Birkhoff-von Neumann traffic matrix decomposition requires N2 - 2N + 2 configurations in the schedule, which lead to a very large packet delay bound. The existing DOUBLE algorithm requires a fixed number of only 2N configurations, but it cannot adjust its schedule according to different switch parameters. In this paper, we first design a generic approach to decompose a traffic matrix into an arbitrary number of Ns (N² - 2N + 2 > N_S > N) configurations. Then, by taking the reconfiguration overhead into account, we formulate a speedup function. Minimizing the speedup function results in an efficient scheduling algorithm ADAPT. We further observe that the algorithmic efficiency of ADAPT can be improved by better utilizing the switch bandwidth. This leads to a more efficient algorithm SRF (scheduling residue first). ADAPT and SRF can automatically adjust the number of configurations in a schedule according to different switch parameters. We show that both algorithms outperform the existing DOUBLE algorithm.     

Blocking probability of heterogeneous traffic in a multiratemulticast switch

A multirate multicast switch, which can provide a single uniform switching function for vastly different classes of traffic, is crucial for the successful deployment of integrated broadband networks. We analyze the call blocking probability of heterogeneous circuit switched traffic in a multirate multicast switch using the arrival modulation technique. For the analysis, we introduce two simple traffic models: fan-out heterogeneous traffic and bandwidth heterogeneous traffic. The analysis of two simplified traffic models shows that calls of different characteristics interact in complex yet subtle ways. Our results on fan-out heterogeneous traffic show that the performance gap between large fan-out traffic and small fan-out traffic increases when two types of traffic are mixed. On the other hand, in bandwidth heterogeneous traffic, the performance gap between traffic of different bandwidth decreases. A further investigation indicates that the two types of interactions coexist in general heterogeneous traffic. We conclude the paper introducing several methods to improve the performance of multirate multicast traffic     

Dynamic inter-SLA resource sharing in path-oriented differentiated services networks

This paper proposes novel resource sharing schemes for differentiated services (DiffServ) networks, to achieve both high resource utilization and quality of service (QoS) guarantee. Service level agreements (SLAs) are negotiated at network boundaries and supported by path-oriented resource mapping within the network. The recently proposed SLA management scheme based on virtual partitioning (Bouillet et al., 2002) allows overloaded SLAs to exploit the spare capacity of underloaded SLAs for efficient resource utilization, however, at the the cost of possible SLA violation of the underloaders. In the bandwidth borrowing scheme proposed here, the dedicated bandwidth for underloaded SLAs is determined and adaptively adjusted at network boundaries according to the actual traffic load and QoS policies; the available spare capacity is then properly distributed to related links for lending to others. On the other hand, the traffic flows admitted with borrowed bandwidth are tagged and may be preempted later when the original bandwidth owner needs to claim back the resources. Through a detailed implementation design and extensive computer simulation results we show that, by bandwidth borrowing, both SLA compliance and high resource utilization can be achieved in various load conditions, with some side benefits such as call-level service differentiation, small admission overhead, and convenience for policy-based management. In addition, we propose a distributed bandwidth pushing scheme that can dynamically adjust the spare bandwidth distribution over the network. Combining bandwidth pushing with bandwidth borrowing, the resource utilization can be further improved.     

Design and evaluation of an efficient traffic control scheme for integrated voice,video, and data over asynchronous transfer mode networks: Explicit allowed rate algorithm

Asychronous transfer mode (ATM) networks are high‐speed networks with guaranteed quality of service. The main cause of congestion in ATM networks is over utilization of physical bandwidth. Unlike constant bit‐rate (CBR) traffic, the bandwidth reserved by variable bit‐rate (VBR) traffic is not fully utilized at all instances. Hence, this unused bandwidth is allocated to available bit‐rate (ABR) traffic. As the bandwidth used by VBR traffic changes, available bandwidth for ABR traffic varies; i.e., available bandwidth for ABR traffic is inversely proportional to the bandwidth used by the VBR traffic. Based on this fact, a rate‐based congestion control algorithm, Explicit Allowed Rate Algorithm (EARA), is presented in this paper. EARA is compared with Proportional Rate Control Algorithm (PRCA) and Explicit Rate Indication Congestion Avoidance Algorithm (ERICA), in both LAN and WAN environments. Simulations of all three algorithms are conducted under both congestion and fairness configurations with simultaneous generation of CBR, rt‐VBR, nrt‐VBR and ABR traffic. The results show that, with very small over‐head on the switch, EARA significantly decreases the required buffer space and improves the network throughput. Copyright © 1999 John Wiley & Sons, Ltd.