A differentiated proposal of three dimension i/o performance characterization model focusing on storage environments

Abstract

The I/O bottleneck remains a central issue in high-performance environments. Cloud computing, high-performance computing (HPC) and big data environments share many underneath difficulties to deliver data at a desirable time rate requested by high-performance applications. This increases the possibility of creating bottlenecks throughout the application feeding process by bottom hardware devices located in the storage system layer. In the last years, many researchers have been proposed solutions to improve the I/O architecture considering different approaches. Some of them take advantage of hardware devices while others focus on a sophisticated software approach. However, due to the complexity of dealing with high-performance environments, creating solutions to improve I/O performance in both software and hardware is challenging and gives researchers many opportunities. Classifying these improvements in different dimensions allows researchers to understand how these improvements have been built over the years and how it progresses. In addition, it also allows future efforts to be directed to research topics that have developed at a lower rate, balancing the general development process. This research present a three-dimension characterization model for classifying research works on I/O performance improvements for large scale storage computing facilities. This classification model can also be used as a guideline framework to summarize researches providing an overview of the actual scenario. We also used the proposed model to perform a systematic literature mapping that covered ten years of research on I/O performance improvements in storage environments. This study classified hundreds of distinct researches identifying which were the hardware, software, and storage systems that received more attention over the years, which were the most researches proposals elements and where these elements were evaluated. In order to justify the importance of this model and the development of solutions that targets I/O performance improvements, we evaluated a subset of these improvements using a a real and complete experimentation environment, the Grid5000. Analysis over different scenarios using a synthetic I/O benchmark demonstrates how the throughput and latency parameters behaves when performing different I/O operations using distinct storage technologies and approaches.