Real-time Query Performance Optimization Of Time-series Database Based On Heterogeneous Architecture

With the rapid development of Internet technology and the rise of the Internet of Things,huge Internet service monitoring services,industrial sensors,mobile terminals and so on have generated massive time series data,the scale of which often exceeds 10 billion.As for managing these data,current systems have begun to show an insufficiency in performance.In order to effectively store data,some systems use disks to organize data,which have enough space for building efficient indexes,but disk I/O limits the query performance.Others have designed solutions based on new architectures,hoping to improve system performance through high-performance hardware devices such as SSD and Memory.However,the index retrieval efficiency and massive data decompression ultimately limit the query efficiency.In order to solve these problems,we designed and implemented PCQO(Parallel Computing for Time Series Query Optimization with GPU),a query optimization solution by using parallel computing with GPU(Graphics Processing Unit).The optimization idea of PCQO is to accelerate the intersection operation of the inverted index and the data decompression operation executed in real-time time series query through GPU parallel computing to reduce the query latency.At the same time,according to the analysis of production data,an interesting observation is that real-time query requests for time series data only involve data within the latest few hours,so PCQO chooses to store recent data and its indexes in GPU memory,this approach can greatly reduce the time overhead caused by data transmission from main memory to GPU global memory.The experimental results show that PCQO can effectively improve the real-time performance of the request response.The intersection speed of inverted index is improved by over 2 times,and the processing speed of decompressing data is increased by over 5 to 8 times,and overall,the query performance has improved by over 1.6 times.As for these evaluations,the larger the data scale is,the more improvement they are.