Most current Light Field Salient Object Detection (LFSOD) methods require full supervision with labor-intensive pixel-level annotations. Unsupervised Light Field Salient Object Detection (ULFSOD) has gained attention due to this limitation. However, existing methods use traditional handcrafted techniques to generate noisy pseudo-labels, which degrades the performance of models trained on them. To mitigate this issue, we present a novel learning-based approach to synthesize labels for ULFSOD. We introduce a prominent focal stack identification module that utilizes light field information (focal stack, depth map, and RGB color image) to generate high-quality pixel-level pseudo-labels, aiding network training. Additionally, we propose a novel model architecture for LFSOD, combining a multi-scale spatial attention module for focal stack information with a cross fusion module for RGB and focal stack integration. Through extensive experiments, we demonstrate that our pseudo-label generation method significantly outperforms existing methods in label quality. Our proposed model, trained with our labels, shows significant improvement on ULFSOD, achieving new state-of-the-art scores across public benchmarks.

Simulation is crucial to performance evaluation of switches. Currently stochastic simulation is the predominant approach, which has two drawbacks: few workload patterns and long simulation times. In this paper, we propose a novel switch simulation method that is based on two techniques: exhaustive workload pattern generation and smoothed periodic input generation. The exhaustive workload pattern generation can produce a huge number of nonuniform workload patterns, which far exceeds the traditional few ones; conclusions based on such truly extensive simulations are much more convincing. The periodic input or cell arrival pattern outperforms the stochastic counterpart in terms of easy repetition and fast convergence of simulation; in particular, among periodic inputs, the smoothed periodic input is most favorable to switch scheduling, and hence switches performing poor with it probably performs worse under others. Combining these two techniques together can systematically identify lots of stuck states at which some switches show poor performances such as low throughput. Specifically, this method discovers that the throughput of iSLIP, FIRM and DRRM, each with one iteration, may be lower than 60% under certain nonuniform and smoothed periodic traffic pattern.

Compared with a bufferless crossbar switch, a combined input-crosspoint-queued (CICQ) switch has better scalability owing to its distributed scheduling. Although the previously proposed round-robin algorithms achieve 100% throughput asymptotically under uniform traffic, these algorithms do not provide a satisfactory performance under nonuniform traffic. In this paper, we propose an efficient round-robin algorithm for a CICQ switch with one-cell cross point buffers. With our algorithm, each input arbiter is associated with dual round-robin pointers. Unlike the existing round-robin algorithms, our algorithm has distinctive round-robin pointer updating rules which are powerful to cope with nonuniform traffic patterns. Extensive simulation results show that our algorithm achieves a satisfactory performance under both uniform and a broad class of nonuniform traffic patterns.

Input queued switches have been well studied in the recent past. A scheduling algorithm is required to schedule the transfer of packets through the switch fabric (e.g. crossbar) at every time slot. The maximum weight matching (MWM) algorithm is known to deliver 100% throughput under any admissible traffic. However, MWM is not practical for its high computational complexity O(N/sup 3/). In this paper, we study a class of approximation algorithms to MWM from the point of view of local search. Notably we observe that: (a) Local search is well suitable for parallel computation. (b) Each line card of high performance router has at least one processor. Based on the two important observations, a parallelized greedy scheduling algorithm is proposed. The proposed algorithm is proved to be rate stable under any admissible traffic. Simulation results show that our algorithm outperforms other randomized approximations to MWM.

Scalability considerations drive the switch fabric design to evolve from output queueing to input queueing and further to combined input and crosspoint queueing (CICQ). However, few CICQ switches are known with guaranteed quality of service, and credit-based flow control induces a scalability bottleneck. In this paper, we propose a novel CICQ switch called the smoothed buffered crossbar or sBUX, based on a new design objective of smoothness and on a new rate-based flow control scheme called the smoothed multiplexer or sMUX. It is proved that with a buffer of just four cells at each crosspoint, sBUX can utilize 100% of the switch capacity to provide deterministic guarantees of bandwidth and fairness, delay and jitter bounds for each flow. In particular, neither credit-based flow control nor speedup is used, and arbitrary fabric-internal latency is allowed between line cards and the switch core.

Input queued switching architectures have become predominant in high speed switches and routers. In this paper, we change the point of view from weight-based matching to weight-based service, and propose a distributed load adaptive scheduling (DLAS) algorithm. In DLAS, the round robin arbiters are used to find a matching between the input ports and output ports. Once the matching between an input-output pair is established, the scheduler will keep it for a certain period, which is a function of the number of cells queued in the corresponding VOQ. Simulation results show that our scheme achieves high throughput and low delay under admissible traffic. For uniform Bernoulli i.i.d. traffic, it achieves 100% throughput, and for nonuniform traffic, its throughput is almost 100%.