Convolutional neural networks (CNNs) play a key role in deep learning applications. However, the large storage overheads and the substantial computational cost of CNNs are problematic in hardware accelerators. Computing-in-memory (CIM) architecture has demonstrated great potential to effectively compute large-scale matrix–vector multiplication. However, the intensive multiply and accumulation (MAC) operations executed on CIM macros remain bottlenecks for further improvement of energy efficiency and throughput. To reduce computational costs, model compression is a widely studied method to shrink the model size. For implementation in a static random access memory (SRAM) CIM–based accelerator, the model compression algorithm must consider the hardware limitations of CIM macros. In this study, a software and hardware co-design approach is proposed to design MARS, a SRAM-based CIM (SRAM CIM)-based CNN accelerator that can utilize multiple SRAM CIM macros as processing units and support a sparse CNN, and an SRAM CIM-aware model compression algorithm that considers a CIM architecture to reduce the number of network parameters. With the proposed hardware software co-designed method, MARS can reach over 700 and 400 FPS for CIFAR-10 and CIFAR-100, respectively. In addition, MARS achieves 52.3 and 88.2 TOPs/W in VGG16 and ResNet18, respectively.

Software Defined Network separates the control plane from network equipment and has great advantage in network management as compared with traditional approaches. With this paradigm, the security issues persist to exist and could become even worse because of the flexibility on handling the packets. In this paper we propose an effective framework by integrating SDN and machine learning to detect and categorize P2P network traffics. This work provides experimental evidence showing that our approach can automatically analyze network traffic and flexibly change flow entries in OpenFlow switches through the SDN controller. This can effectively help the network administrators manage related security problems.

Without STP and LLDP, we propose a new approach to find faster paths between the source and the destination and fully utilize the links between switches under the Software Defined Network paradigm. To find round-trip path with lighter load, we design a mechanism to compose special ARP packets for exploring the routing path and avoid broadcast storm by dropping late or duplicate packets. Meanwhile, the controller can carefully regulate other packets without affecting the behavior of the special packets. The evaluation results with the Clos-like network topology show that our approach can be used to improve the performance of transmission throughput. We believe our method is very suitable, especially, for a dynamic network environment, where links are up and down frequently. Our implementation can be seen as an instance showing the flexibility of SDN.

IP multicast dictates high-end routers and incurs high administrative overhead, which prevent them from being deployed in many video streaming scenarios. In this paper, we study the problem of computing the multipath multicast routes for streaming videos in Software-Defined Networks (SDNs), which adopt less expensive switches and reduce administrative overhead for lower CAPEX/OPEX. The objectives of the considered problem are robustness, load balance, SDN compatibility, and adaptiveness. We formulate this routing problem into a mathematical optimization problem, and propose two algorithms to address this problem. We implement the proposed algorithms on a popular OpenFlow controller to demonstrate its practicality, and we conduct extensive experiments to evaluate the proposed algorithms. The experiment results clearly show the merits of our algorithms over the IP multicast, e.g., we observe: (i) frame loss rate reduction between 19% and 95%, (ii) video quality improvement between 4 dB and 15 dB, (iii) sink throughput increase between 25% and 66%, and (iv) maximal link utilization reduction between 15% and 50%. We also show the tradeoff between optimality and run time of the two proposed algorithms: one of them is more suitable for smaller and more static networks, and the other one is more suitable for larger and more dynamic networks.