TY - CHAP A1 - Schmidtke, Robert A1 - Schintke, Florian A1 - Schütt, Thorsten T1 - From Application to Disk: Tracing I/O Through the Big Data Stack T2 - High Performance Computing ISC High Performance 2018 International Workshops, Frankfurt/Main, Germany, June 24 - 28, 2018, Revised Selected Papers, Workshop on Performance and Scalability of Storage Systems (WOPSSS) N2 - Typical applications in data science consume, process and produce large amounts of data, making disk I/O one of the dominating — and thus worthwhile optimizing — factors of their overall performance. Distributed processing frameworks, such as Hadoop, Flink and Spark, hide a lot of complexity from the programmer when they parallelize these applications across a compute cluster. This exacerbates reasoning about I/O of both the application and the framework, through the distributed file system, such as HDFS, down to the local file systems. We present SFS (Statistics File System), a modular framework to trace each I/O request issued by the application and any JVM-based big data framework involved, mapping these requests to actual disk I/O. This allows detection of inefficient I/O patterns, both by the applications and the underlying frameworks, and builds the basis for improving I/O scheduling in the big data software stack. Y1 - 2018 U6 - https://doi.org/10.1007/978-3-030-02465-9_6 SP - 89 EP - 102 ER - TY - JOUR A1 - Seibert, Felix A1 - Peters, Mathias A1 - Schintke, Florian T1 - Improving I/O Performance Through Colocating Interrelated Input Data and Near-Optimal Load Balancing JF - Proceedings of the IPDPSW; Fourth IEEE International Workshop on High-Performance Big Data, Deep Learning, and Cloud Computing (HPBDC) N2 - Most distributed file systems assign new files to storage servers randomly. While working well in some situations, this does not help to optimize the input performance for most MapReduce computations’ data access patterns. In this work, we consider an access pattern where input files are partitioned into groups of heterogeneous size. Each group is accessed by exactly one process. We design and implement a data placement strategy that places these file groups together on the same storage server. This colocation approach is combined with near-optimal storage load balancing. To do so, we use a classical scheduling approximation algorithm to solve the NP hard group assignment problem. We argue that local processing is not only beneficial because of reduced network traffic, but especially because it imposes an even resource schedule. Our experiments, based on the parallel processing of remote sensing images, reveal an enormous reduction of network traffic and up to 39 % faster input read times. Further, simulations show that our approximate assignments limit storage server imbalances to less than 5 % above the theoretical minimum, in contrast to more than 85 % with random assignment. Y1 - 2018 U6 - https://doi.org/10.1109/IPDPSW.2018.00081 N1 - Best Paper Award VL - 2018 SP - 448 EP - 457 ER -