Search

The systolic array architecture is one of the most popular choices for convolutional neural network hardware accelerators. The biggest advantage of the systolic array architecture is its simple and efficient design principle. Without complicated control and dataflow, hardware accelerators with the systolic array can calculate traditional convolution very efficiently. However, this advantage also brings new challenges to the systolic array. When computing special types of convolution, such as the small-scale convolution or depthwise convolution, the processing element (PE) utilization rate of the array decreases sharply. The main reason is that the simple architecture design limits the flexibility of the systolic array. In this article, we design a configurable multi-directional systolic array (CMSA) to address these issues. First, we added a data path to the systolic array. It allows users to split the systolic array through configuration to speed up the calculation of small-scale convolution. Second, we redesigned the PE unit so that the array has multiple data transmission modes and dataflow strategies. This allows users to switch the dataflow of the PE array to speed up the calculation of depthwise convolution. In addition, unlike other works, we only make a few changes and modifications to the existing systolic array architecture. It avoids additional hardware overheads and can be easily deployed in application scenarios that require small systolic arrays such as mobile terminals. Based on our evaluation, CMSA can increase the PE utilization rate by up to 1.6 times compared to the typical systolic array when running the last layers of ResNet-18. When running depthwise convolution in MobileNet, CMSA can increase the utilization rate by up to 14.8 times. At the same time, CMSA and the traditional systolic arrays are similar in area and energy consumption.

With the rapid increase of the size of applications and the complexity of the supercomputer architecture, topology-aware process mapping becomes increasingly important. High communication cost has become a dominant constraint of the performance of applications running on the supercomputer. To avoid a bad mapping strategy which can lead to terrible communication performance, we propose an optimized heuristic topology-aware mapping algorithm (OHTMA). The algorithm attempts to minimize the hop-byte metric that we use to measure the mapping results. OHTMA incorporates a new greedy heuristic method and pair-exchange-based optimization. It reduces the number of long-distance communications and effectively enhances the locality of the communication. Experimental results on the Tianhe-3 exascale supercomputer prototype indicate that OHTMA can significantly reduce the communication costs.

Punctuation restoration in speech recognition has a wide range of application scenarios. Despite the widespread success of neural networks methods at performing punctuation restoration for English, there have been only limited attempts for Chinese punctuation restoration. Due to the differences between Chinese and English in terms of grammar and basic semantic units, existing methods for English is not suitable for Chinese punctuation restoration. To tackle this problem, we propose a hybrid model combining the kernel of Bidirectional Encoder Representations from Transformers (BERT), Convolution Neural Network (CNN) and Recurrent Neural Network (RNN). This model employs a flexible structure and special CNN design which can extract word-level features for Chinese language. We compared the performance of the hybrid model with five widely-used punctuation restoration models on the public dataset. Experimental results demonstrate that our hybrid model is simple and efficient. It outperforms other models and achieves an accuracy of 69.1%.

Modern vehicular communication systems are facing significant challenges since the rapid development of intelligent transportation systems and the ever-growing vehicular applications have activated intense information exchange among vehicles, infrastructures, and pedestrians. Without the support of powerful communication, many vehicular services will just stay in conceptual phase and cannot be put into practice. To provide low-latency services to vehicles on roads, vehicular fog computing (VFC) is recently introduced as a promising solution, which utilizes enormous vehicles on roads as communication and computation resources and extends cloud computing service to an edge network. On the other hand, the publish/subscribe (pub/sub) paradigm provides a loosely coupled and scalable communication which can facilitate flexible and dynamic vehicular network services. Motivated by the merits of these two research fields, in this paper, we propose a novel joint design of pub/sub communication model based on VFC architecture, which employs fog nodes as the data platform for messages aggregation. Specifically, we describe a method to measure vehicle mobility and construct a stable VFC on roads. Then, a message dissemination approach is designed based on our communication model. Finally, the experimental results confirm the efficiency of our proposed model in various urban scenarios.

The quality of the mesh is one of the most critical aspects for solving partial differential equations (PDEs) in applications of Computational Fluid Dynamics. Many geometry criteria have been proposed and are widely used in business preprocessing software like ICEM CFD, PointWise, Gambit. However, these traditional geometry criteria fail to recognize some quality features that seriously affect the accuracy of numerical calculations, such as density and distribution of mesh elements. These quality features are usually evaluated based on engineering experience, which heavily increases the pre-processing cost and requires extensive engineering experience. In this paper, we introduce a deep learning model to solve the mentioned issues by offline learning. The proposed model is small and fast and can be embedded in pre-processing software. Experiment results show that the derived model is capable of performing the quality evaluating task and achieve an accuracy of 93.8%.

With the rapid development of highperformance computing, computational fluid dynamics (CFD) has become an important part of hydrodynamics and aerodynamics. Mesh quality is the key factor that affects the accuracy and efficiency of CFD numerical calculation. However, the current the process of mesh quality discrimination is very time-consuming. The manpower time needed for this process takes up a large proportion in the whole numerical calculation process. A large number of artificial intelligence algorithms have been put forward to replace the human to efficiently complete all kinds of tedious tasks. In this paper, we propose a convolutional neural network (CNN) based mesh quality discrimination method, MeshNet. MeshNet uses residual neural network structure to learn mesh features and automatically judge the mesh quality. The experimental results show that the proposed network can greatly save labor time cost and achieve an accuracy of 94.41% for mesh quality discrimination.

With the increasing size of high performance computing systems, the expensive communication overhead between processors has become a key factor leading to the performance bottleneck. However, default process-to-processor mapping strategies do not take into account the topology of the interconnection network, and thus the distance spanned by communication messages may be particularly far. In order to enhance the communication locality, we propose a new topology-aware mapping method called TAMM. By generating an accurate description of the communication pattern and network topology, TAMM employs a two-step optimization strategy to obtain an efficient mapping solution for various parallel applications. This strategy first extracts an appropriate subset of all idle computing resources on the underlying system and then constructs an optimized one-to-one mapping with a refined iterative algorithm. Experimental results demonstrate that TAMM can effectively improve the communication performance on the Tianhe-2A supercomputer.

With the development of Internet applications, the data storage of many applications has some new characters. There always exists a huge amount of pictures in distributed file systems, and the size of these pictures is usually no more than 1M bytes. To meet the demands in small file storage system which contains so many pictures, we provide a kind of distributed file system to store large amounts of small files. In our system, small files (actual data files) are merged into large files (default 64M) which we call block, and we use two-indexed way to reduce the pressure of nameserver, which maintains the information of correspondence between blocks and dataservers. The result shows that the memory usage of block information in nameserver is less than 2G when the system capacity is 1PB. It shows good scalability. Meanwhile, to reduce the cost of storage, we introduce the technique of erasure code which is an alternative offers the same data protection but reduces significantly the storage consumption. After grouping, the cost of storage dropped by 25% compared with 2 replications.