Congratulations Dejan Kostić!
For winning a Best Paper Award
We met up with Dejan Kostić who has won a Best Paper Award at the ACM Symposium on Operating Systems Principles (SOSP) 2021 and talked about what the award will mean to him and his team.
What does this Best Paper Award mean to you?
"SOSP is the top conference in computer systems, and this award means a lot to me. I have never been happier with my research team and collaborators. It is very pleasant to look back on this year and see that my doctoral student Alireza Farshin (co-advised with Prof. Maguire) won the Google PhD Fellowship Award for 2021. We also recently received a Vetenskaprådet (VR, Swedish Research Council in English) grant with Magnus Boman (KTH), Marco Chiesa (KTH), and Sabine Koch (KI)."
"This research should have a positive impact on sustainable development, the economy, and the environment."
Could you tell us a bit about the paper?
"The system we built for this paper is called LineFS, and it is a culmination of multi-year efforts in a large collaboration. LineFS provides further performance improvements on Assise, which was featured at last year's USENIX OSDI conference (another top systems venue). My student Waleed Reda was the lead author and the paper presenter. So, I will tell you first a bit about Assise (the full title of that paper is "Performance and Availability via Client-local NVM in a Distributed File System").
Distributed file systems (DFSes) are a key building block of any modern computing system. The recent adoption of low-latency persistent memory modules (PMMs, sometimes called NVRAM) upends the long-established model of distributed file systems. The client machines make requests across the network for data hosted by remote servers on classic hard disks or Solid-state Drives (SSDs). Instead, by co-locating computation with PMM storage, we can provide applications with much higher IO performance and fast application fail-over. To demonstrate this, we built the Assise distributed file system and compared it to the state-of-the-art distributed file systems on a cluster with newly-announced Intel Optane DC PMMs. While executing widely-used cloud applications, we show that Assise improves write latency up to 22x, throughput up to 56x, fail-over time up to 103x, and scales up to 6x better than its counterparts while providing stronger consistency semantics.
However, it turns out that, in the systems hosting multiple so-called tenants, the CPU overhead created by this new class of distributed file systems is increasingly a burden to application performance. CPU and memory interference degrade application and storage performance, particularly for latency (the time it takes to read from or write to the file system). DFS offload to SmartNICs (a new programmable Network Interface Cards) is a promising solution to these problems. Still, it is challenging to fit the complex demands of a DFS onto simple, less-performing SmartNIC processors located across the PCIe bus. We present LineFS, a SmartNIC-offloaded, high-performance DFS with support for client-local persistent memory. To fully leverage the SmartNIC architecture, we decompose DFS operations into execution stages that can be offloaded to a parallel data-path execution pipeline on the SmartNIC. We implement LineFS on the Mellanox BlueField-1 SmartNIC and compare it to Assise while running widely-used application benchmarks. LineFS delivers about 80% performance improvement across the board. Besides providing quantitative improvements, LineFS demonstrates a qualitatively new functionality by allowing the data to be compressed before it is sent across the network, resulting in up to 72% data reduction and another 11% performance improvement.
This research should have a positive impact on sustainable development, the economy, and the environment. Both Assise and LineFS expose a traditional POSIX API, supporting a wide range of applications and services. Our file systems provide high IO performance and availability, which directly contributes to increased productivity. Moreover, they are also designed for efficiency, which translates to lower energy usage and operating expenses."
What are you working on at the moment?
"We are continuing our work on ultra-low-latency networked systems, primarily within my 2018-2024 ERC Consolidator Project called ULTRA. In this single-PI, 2-Million EUR project, we want to change how Internet Services are constructed dramatically. European Research Council (ERC) gives a lot of freedom to the researchers to pursue new research directions, and we certainly did that with the work on Assise and LineFS.
Our upcoming USENIX NSDI 2022 RedN paper will also show that Remote Direct Memory Access (RDMA) Network Interface Cards (NICs) are Turing Complete (capable of performing arbitrary computation)! This effectively allows us to use RDMA NICs as general-purpose processors. Leveraging this newfound capability will be critical for further improving the performance and energy efficiency of Internet services. As a matter of fact, in this paper, we demonstrate a 35x reduction in tail latency of busy servers. We are very excited about the potential impact of this work, and Waleed will present it in April 2022. Along with Assise, RedN will also be prominently featured in his upcoming PhD thesis."