"Big Data" is everywhere -- both the IT industry and the scientific computing community are routinely handling terabytes to petabytes of data. This preponderance of data has fueled the development of data-intensive scalable computing (DISC) systems that manage, process and store massive data-sets in a distributed manner. For example, Google and Yahoo have built their respective Internet services stack to distribute processing (MapReduce and Hadoop), to program computation (Sawzall and Pig) and to store the structured output data (Bigtable and HBase). Both these stacks are layered on their respective distributed file systems, GoogleFS and Hadoop distributed FS, that are designed "from scratch" to deliver high performance primarily for their anticipated DISC workloads.
However, cluster file systems have been used by the high-performance computing (HPC) community at even larger scales for more than a decade. These cluster file systems, including IBM GPFS, Panasas PanFS, PVFS, and Lustre, are required to meet the scalability demands of highly parallel I/O access patterns generated by scientific applications that execute simultaneously on tens to hundreds of thousands of nodes. Thus, given the importance of scalable storage to both the DISC and the HPC world, the authors take a step back and ask we are at a point where we can distill the key commonalities of these scalable file systems.
In Search of an API for Scalable File Systems: Under the Table or Above It? by Swapnil Patil, Garth A. Gibson, Gregory R. Ganger, Julio Lopez, Milo Polte, Wittawat Tantisiroj, and Lin Xiao, is not a paper about engineering yet another "right" file system or database, but rather about how do we evolve the most dominant data storage API -- the file system interface -- to provide the right abstraction for both DISC and HPC applications. What structures should be added to the file system to enable highly scalable and highly concurrent storage? Our goal is not to define the API calls per se, but to identify the file system abstractions that should be exposed to programmers to make their applications more powerful and portable. This paper highlights two such abstractions. First, the authors show how commodity large-scale file systems can support distributed data processing enabled by the Hadoop/MapReduce style of parallel programming frameworks. And second, the authors argue for an abstraction that supports indexing and searching based on extensible attributes, by interpreting BigTable as a file system with a filtered directory scan interface.