The International Conference for High Performance Computing, Networking, Storage and Analysis
Big Data Analytics on Object Stores: A Performance Study.
Authors: Lukas Rupprecht (Imperial College London), Rui Zhang (IBM Corporation), Dean Hildebrand (IBM Corporation)
Abstract: Object Stores provide a cheap solution for storing, sharing, and accessing globally distributed, scientific datasets. Their scalability and high reliability makes them well suited as archive storage for data produced by HPC clusters. To make this archive active, i.e. to process the archived data, a separate analytics cluster is needed which adds overhead in terms of cost, maintenance, and performance as data needs to be copied between two systems. Running the analytics directly on the Object Store can greatly simplify the archive usage. We study the problems that arise when running an analytics system (Hadoop MapReduce) on top of an Object Store (OpenStack Swift). We conduct a set of detailed micro- and macro benchmarks and find that the static, consistent hashing based, object-node mapping adds significant overhead when object writes are not local. Additionally, repeated authentication calls slow down jobs which has a high performance impact, especially on interactive workloads.