A Task-Aware Fine-Grained Storage Selection Mechanism for In-Memory Big Data Computing Frameworks

摘要

In-memory big data computing, widely used in hot areas such as deep learning and artificial intelligence, can meet the demands of ultra-low latency service and realtime data analysis. However, existing in-memory computing frameworks usually use memory in an aggressive way. Memory space is quickly exhausted and leads to great performance degradation or even task failure. On the other hand, the increasing volumes of raw data and intermediate data introduce huge memory demands, which further deteriorate the short of memory. To release the pressure on memory, those in-memory frameworks provide various storage schemes options for caching data, which determines where and how data is cached. But their storage scheme selection mechanisms are simple and insufficient, always manually set by users. Besides, those coarse-grained data storage mechanisms cannot satisfy memory access patterns of each computing unit which works on only part of the data. In this paper, we proposed a novel task-aware fine-grained storage scheme auto-selection mechanism. It automatically determines the storage scheme for caching each data block, which is the smallest unit during computing. The caching decision is made by considering the future tasks, real-time resource utilization, and storage costs, including block creation costs, I/O costs, and serialization costs under each storage scenario. The experiments show that our proposed mechanism, compared with the default storage setting, can offer great performance improvement, especially in memory-constrained circumstances it can be as much as 78%.