LABIOS Receives Karsten Schwan Best Paper Award at HPDC'19
The paper "LABIOS: A Distributed Label-Based I/O System" has received the Karsten Schwan Best Paper Award at the ACM 28th International Symposium on High-Performance Parallel and Distributed Computing (HPDC'19). The award, named in honor of the late Professor Karsten Schwan for his contributions to high-performance systems, recognizes the most outstanding paper presented at the symposium.
The LABIOS architecture
LABIOS supports conflicting I/O workloads from scientific and Big Data applications under a single storage system. Its central idea is the label, a new data representation. A label carries a unique identifier, a pointer to the data, and the operation to perform on it, so the instruction of an I/O request travels separately from its data.
In LABIOS, a Label Manager builds labels from each I/O request and publishes them to a distributed label queue. A Content Manager holds the request data in a distributed in-memory warehouse, and a Catalog Manager maintains user and system metadata. A Label Dispatcher sends labels to workers using configurable scheduling policies, and a Worker Manager balances load across the worker pool and commissions or decommissions workers as needed. Because the application only enqueues labels, its I/O becomes asynchronous.
The paper reports that LABIOS boosts I/O performance by up to 17x and reduces overall execution time by 40 to 60 percent in the workloads it evaluates. It also supports heterogeneous storage resources, offers storage elasticity, and enables in-situ analytics through data provisioning.
The team
The work was led by Anthony Kougkas as part of his doctoral research, with Hariharan Devarajan and Jay Lofstead (Sandia National Laboratories), under the supervision of Dr. Xian-He Sun. The label-based I/O method was later granted US Patent 11,630,834, and LABIOS continues as an active project under NSF award OAC-2313154 (opens in a new tab).