Time-sharing Multithreading on Stream-based Lossless Data Compression
Time-sharing Multithreading on Stream-based Lossless Data Compression
复制标题
基于流的无损数据压缩的分时多线程
DOI：
10.1109/candar.2017.42
复制
发表时间：
2017
期刊：
影响因子：
--
通讯作者：
Koichi Marumo and Shinichi Yamagiwa
中科院分区：
文献类型：
--
作者：
Maruta Ei;Wang Jingwen;Kotani Tomomi;Tsuda Hiroyuki;Nakano Tomoko;Imai Kenji;Sumigama Seiji;Niwa Yoshimitsu;Mitsui Takashi;Yoshida Shigeru;Yamashita Mamoru;Nawa Akihiro;Tamakoshi Koji;Kajiyama Hiroaki;Kikkawa Fumitaka;Koichi Marumo and Shinichi Yamagiwa
为了解决大数据迁移中通信数据路径的性能饱和问题，无损数据压缩技术是一种提高路径带宽的解决方案。然而，目前的数据压缩机制需要以非常低的延迟来处理数据流（例如传感器数据），以避免路径中的开销。的快速流无损数据压缩，使用基于字典的符号查找机制。当将其应用于非常快的路径时，由于字典查找操作成为硬件中最长延迟路径的瓶颈，因此硬件延迟大大增加并且时钟速度降低。本文提出了一种将多线程技术应用于字典查找操作的性能改进技术。该技术使得LCA-DLT的单个模块能够通过在压缩/解压缩流水线的混叠阶段中划分压缩定时来接受多个数据流。根据具有两个线程的硬件实现的性能评估，虽然数据压缩带宽在逻辑上变为原始单线程LCA-DLT的一半，但分时多线程技术减少了所需的硬件资源并提高了时钟频率。
To address the performance saturation in communication data path at migrating BigData, the lossless data compression technique is a solution to enhance the bandwidth of the path. However, the recent data compression mechanism needs to treat data streams such as sensor data with very low latency to avoid overhead in the path. This paper focuses on a new loss less data compression mechanism called LCA-DLT that implements a hardware-based fast stream lossless data compression using dictionary-based symbol lookup mechanism. When applying it to a very fast path, the hardware latency increases largely and the clock speed degrades because the dictionary lookup operation becomes bottleneck of the longest delay path in the hardware. This paper proposes a performance improvement technique applying multithreading technique in the dictionary lookup operation. The technique enables a single module of the LCA-DLT to accept multiple data streams by dividing the compression timing in babble stage of the compression/decompression pipeline. According to performance evaluation by a hardware implementation with two threads, although the data compression bandwidth logically becomes half of the original single thread LCA-DLT, the time-sharing multithreading technique reduces required hardware resources and improves the clock frequency.