IoT edge platform has become popular in various distributed environments. The edge devices need to communicate BigData among them or with the cloud servers by collaborating with AI technologies for finding events from the applications. Those devices exchange data streams from such as distributed sensors and remote image/video devices. We focus on an acceleration technique for the communication performance using a stream-based lossless data compression technology. This paper proposes a parallelization technique for the compression process in a software environment running on a multicore processor. The technique invokes concurrent compression processes assigned to multiple threads with splitting a data stream to chunks. The paper exposes three scheduling methods for assigning the chunks to the threads: in-order, hybrid and out-of-order. As an original data order of chunks must be obtained in decompression side, the proposed technique introduces packeting mechanisms in each chunk by adding headers to support the scheduling methods. Through experimental performance evaluations, we discuss the packeting overhead focusing on compression ratio and speedup by the parallelization with three scheduling methods.

This poster presentation proposes a novel lossless data compression algorithm that completely supports stream data called Adaptive Stream-based Entropy coding. It compresses continuous stream data using a look-up table without stalling/buffering by assigning the fewest bits according to instant entropy. The mechanism is suitable for hardware implementation.

Dictionary-based lossless data compression algorithms mainly replace a frequent data pattern in the inputted data to a compressed symbol, and to decompress vice versa. The mechanism potentially has an overhead problem regarding the number of symbol matchings in the table. This work focuses on a technique to reduce the number of searches in the dictionary using a bank separation technique. This poster presentation shows design and implementation of the technique applied to LCT-DLT.

Capturing human movement has become available in detail due to the advancement of motion sensor technology integrated by micro-machine and also due to the one of optical recording by high speed and high resolution image sensors. Therefore, we can easily record the human activity as the body movement BigData and analyze it to quest skill to become an expert of a target body movement. Especially, in the sports activity, the quest for becoming an expert athlete has been tried by using a mathematical model of an ideal body movement experienced from the biomechanics approach. The skill is discussed by comparing the differences from the predicted coordinates of body parts captured during the target performance. However, the approach potentially includes difficulties such as modeling the body control from the dynamics system for all human movements. And also the approach needs for adjusting jitters of the individual characteristics. Therefore, when applying the conventional approach, we must discuss a huge number of combinations of mathematical models and then we would find a model for the ideal body movement. To overcome the difficulty, this paper proposes an approach to visualize skill differences among experts and beginners from the BigData called the skill grouping method. It exploits the skill groups clustered by machine learning approach based on a kernel method. This paper shows applications of the skill grouping method from sports activities. Those show validities for finding the skill differences comparing to the BigData of skillful athletes, and also the one for managing skill transition of an athlete in a timeline.

When we consider any applications that use large data continuously produced, it is necessary for the system developer to apply some fast method that migrates the data stream to the processors. Even if we consider the internal communications of a BigData processing system, applications that treat dataflow such as from a sensor system with tens of channels to a peripheral bus for interconnections among processing modules are currently facing a critical frequency problem to exchange data in the busses because the data size has become very large. One of the best solutions to improve the situation is to compress the exchanged data stream during the transfer in the interconnection among processing modules. However, the conventional compression mechanisms used by software solutions such as ZIP and LZW need to aggregate the compressed data and a table that includes the information for recovering the compressed data to the original one. This paper shows a novel compression mechanism based on the symbol pair matching that uses a coherent and static lookup table with a limited number of entries of the symbol pairs. Building a compression pipeline with multiple tables we can implement an effective data path of the stream-based compression with a reconfigurable and flexible compression ratio applying trained tables from the original data characteristics. This paper shows the algorithm design and an implementation example on an FPGA using the content addressable memory and reports the performance of the hardware.