By David Murphy
A collaborative multi-institutional research team featuring KAUST Associate Professor of Computer Science Marco Canini won a best paper award at the 28th ACM Symposium on Operating Systems Principles (SOSP 2021). The biennial event held from October 26-29, 2021, is the world's premier forum for researchers, developers, programmers and teachers of computer systems technology to advance the theory and practice of modern computer systems software.
The paper, titled "LineFS: Efficient SmartNIC Offload of a Distributed File System with Pipeline Parallelism," is the result of an ongoing collaborative effort between Canini and his colleagues at the Korea Advanced Institute of Science and Technology (KAIST), South Korea; the University of Texas at Austin (UT Austin), U.S.; the Royal Institute of Technology (KTH), Sweden; and the Université catholique de Louvain (UCLouvain), Belgium.
Led by KAIST researchers Jongyul Kim and Professor Youngjin Kwon, the SOSP Best Paper win follows on from a previous successful partnership that also resulted in a publication at the prestigious 14th USENIX Symposium on Operating Systems Design and Implementation (OSDI 2020).
Professor Simon Peter at UT Austin and the team had been investigating the implications of persistent memory on the design of file systems. The close collaborative links with both UT Austin and KAIST came about when Canini and one of his former students, Waleed Reda (currently pursuing a dual Ph.D. degree at UCLouvain and KTH), came together to create a distributed file system named Assise.
“Assise questioned the long-established model of remote storage for distributed file systems and encouraged collocating computation with persistent memory storage. By doing so, Assise provides applications with much higher performance, sub-second application failover and stronger consistency.”
Supporting a modern computing evolution
SmartNICs are a modern technological evolution of network interface cards (NICs) that traditionally enable computers to send and receive data. SmartNICs—a programmable network adapter card with accelerators and Ethernet connectivity that can expedite infrastructure applications—offer increased computational and offload capabilities to process data directly on its path to/from a network.
“Our paper outlines how LineFS realizes a high-performance distributed filesystem with persistent memory by offloading certain crucial operations on SmartNICs. The SmartNICs free up precious CPU cycles and enable improvements in application latency and throughput. LineFS pushes the boundaries further by integrating SmartNICs as a first-class component of the distributed filesystem pioneered with Assise,” Canini emphasized.
“An ongoing project that my SANDS Lab is working on is how to leverage SmartNICs to accelerate distributed deep learning. This year, the group introduced in-network aggregation to efficiently support collective communication for dense and sparse inputs using programmable network switches. Our next focus is to leverage the functions of SmartNICs to increase the data-gathering performance of GPUs,” he concluded.