Quantifying and Mitigating Computational Inefficiency of Genomics Data Analysis

摘要

In this paper, we performed a comprehensive study of quantifying and mitigating computational inefficiency of current genomic analysis approaches. First, we found current parallelization approaches that have limited scalability due to either unexploited parallelism or low utilization of system resource. Thus, we proposed Spark-Gene, which is on the basis of Spark in-memory programming model. To test the performance of our Spark-Gene, we used WGS in the GATK as the test case. We show that Spark-Gene reduces the execution time of WGS analysis from 19 hours to 30 minutes with a speedup in excess of 37-fold at 256 CPU cores. Furthermore, we identified that garbage collection is the major scalable bottleneck of better parallel efficiency for native in-memory computing model. Second, we quantified microarchitectural inefficiency for typical genomic applications and uncovered opportunities for microarchitectural optimizations for the design of genomic domain-specific accelerator, especially on specializing concurrency, computation and memory hierarchy. This paper is to leverage state-of-art big-data technologies to improve parallelization for genomics analysis and motivate the integration of accelerators into the genomic analysis computing system.