Scaling Distributed Hierarchical File Systems Using NewSQL Databases

Sammanfattning: For many years, researchers have investigated the use of database technology to manage file system metadata, with the goal of providing extensible typed metadata and support for fast, rich metadata search. However, earlier attempts failed mainly due to the reduced performance introduced by adding database operations to the file system’s critical path. Recent improvements in the performance of distributed in-memory online transaction processing databases (NewSQL databases) led us to re-investigate the possibility of using a database to manage file system metadata, but this time for a distributed, hierarchical file system, the Hadoop Distributed File System (HDFS). The single-host metadata service of HDFS is a well-known bottleneck for both the size of the HDFS clusters and their throughput.In this thesis, we detail the algorithms, techniques, and optimizations used to develop HopsFS, an open-source, next-generation distribution of the HDFS that replaces the main scalability bottleneck in HDFS, single node in-memory metadata service, with a no-shared state distributed system built on a NewSQL database. In particular, we discuss how we exploit recent high-performance features from NewSQL databases, such as application-defined partitioning, partition pruned index scans, and distribution aware transactions, as well as more traditional techniques such as batching and write-ahead caches, to enable a revolution in distributed hierarchical file system performance.HDFS’ design is optimized for the storage of large files, that is, files ranging from megabytes to terabytes in size. However, in many production deployments of the HDFS, it has been observed that almost 20% of the files in the system are less than 4 KB in size and as much as 42% of all the file system operations are performed on files less than 16 KB in size. HopsFS introduces a tiered storage solution to store files of different sizes more efficiently. The tiers range from the highest tier where an in-memory NewSQL database stores very small files (<1 KB), to the next tier where small files (<64 KB) are stored in solid-state-drives (SSDs), also using a NewSQL database, to the largest tier, the existing Hadoop block storage layer for very large files. Our approach is based on extending HopsFS with an inode stuffing technique, where we embed the contents of small files with the metadata and use database transactions and database replication guarantees to ensure the availability, integrity, and consistency of the small files. HopsFS enables significantly larger cluster sizes, more than an order of magnitude higher throughput, and significantly lower client latencies for large clusters.Lastly, coordination is an integral part of the distributed file system operations protocols. We present a novel leader election protocol for partially synchronous systems that uses NewSQL databases as shared memory. Our work enables HopsFS, that uses a NewSQL database to save the operational overhead of managing an additional third-party service for leader election and deliver performance comparable to a leader election implementation using a state-of-the-art distributed coordination service, ZooKeeper.