Keynote: Eric Sedlar, VP and Technical Director of Oracle Labs
Explosive growth in dataset sizes and demand for ubiquitous access to information by both human and machine actors is fueling the use of warehouse-scale datacenters. To accommodate massive datasets and achieve their performance and resilience objectives, such datacenters rely on distributed memory systems. Distributed memory comes with a number of challenges, including consistency, scalability, availability and performance predictability. These challenges are amplified by an increasingly diverse datacenter workload portfolio, with established scale-out workloads sharing infrastructure with HPC, machine learning, and traditional server applications migrating to the cloud.
The goal of the workshop is to explore both established and emerging challenges in warehouse-scale memory systems, along with potential angles for attacking these challenges.
Topics of interest include, but are not limited to:
- Design, verification and evaluation of distributed memory systems
- Consistency and persistency models at scale
- Software and hardware support for distributed memory (e.g., memory disaggregation, shared memory semantics)
- Emerging use-cases and applications driving the evolution of distributed memory
- System-level support for distributed memory (e.g., scheduling, data placement, quality of service guarantees)
- Rack- and datacenter-scale interconnects
The workshop will include a mix of accepted presentations and invited talks.
Speaker: Eric Sedlar, VP and Technical Director of Oracle Labs.
Title: The Economics of Moving Compute to Storage and Other Interesting Distributed System Design Ideas.
Abstract: Large scale computing designers will periodically produce proposals to “blow up the boxes” and move around compute, memory and network elements in large-scale distributed system for various sorts of efficiency gains. However, the dominant criterion on how to evaluate such proposals is typically performance for the dollar, rather than looking at considerations like power that are easier for researchers to measure. Even worse, large-scale distributed systems typically are expected to run any arbitrary software including many popular open source packages, for reasons of developer productivity (even harder to measure). Oracle Labs built “Project RAPID”, a large-scale distributed analytic system involving custom silicon with a high-speed InfiniBand interconnect and a custom database engine, which provides some relevant experience to discuss the tradeoffs.
Bio: As VP & Technical Director of Oracle Labs, Eric Sedlar manages a team of close to 200 researchers and engineers worldwide. In his tenure in the Labs, Eric has started a number of long-term system research projects that have led to technology transfer into products, including the GraalVM programming language runtime, PGX Parallel Graph Analytics, and the Parfait tool for Program Analysis. His personal research interests have been in the field of data processing and the intersection with compiler technologies. Eric was the co-author of the SIGMOD Best Paper in 2009 and has been an inventor on 85 granted patents.
8.30 opening remarks
8.45 - 9.45 Keynote: Eric Sedlar, VP and Technical Director of Oracle Labs. The Economics of Moving Compute to Storage and Other Interesting Distributed System Design Ideas.
9.45 - 10.00 Irina Calciu, Aasheesh Kolli, Jayneel Gandhi, Stanko Novakovic, Marcos Aguilera, Rajesh Venkatasubramanian and Pratap Subrahmanyam (VMWare). Resource disaggregation for the 99%. Slides
10.00 - 10.15 Yizhou Shan and Yiying Zhang (Purdue University). Disaggregating Memory with Software-Managed Virtual Cache. Slides
10.15 - 10.45 BREAK
10.45 - 11.00 Aasheesh Kolli (VMware and Penn State), Jayneel Gandhi, Irina Calciu and Stanko Novakovic (VMWare). Remote Memory Persistency. Slides
11.00 - 11.15 Yilun Chen and Yiying Zhang (Purdue University). Split Container: Running Containers beyond Physical Machine Boundaries. Slides
11.15 - 11.30 Maciej Besta, Erik Henriksson and Torsten Hoefler (ETH Zurich). Lowering Diameter Enables Cost-Effective and High-Performance Networks. Slides
11.30 - 11.45 Magnus Norgren (Uppsala University), Andra Hugo (DDN Storage), Stefanos Kaxiras (Uppsala University) and Konstantinos Sagonas (Uppsala University). Unleashing Dynamic Task Scheduling at Rack-Scale. Slides
11.45 - 12.00 Shin-Yeh Tsai and Yiying Zhang (Purdue University). Mitsume: an Object-Based Remote Memory System. Slides
Boris Grot, University of Edinburgh
Vijay Nagarajan, University of Edinburgh
Mark Silberstein, Technion
Please contact Boris Grot at email@example.com