Keynote 2: Making Cloud Systems Reliable and Dependable: Challenges and Opportunities

Speaker: Lidong Zhou

Abstract:As our society increasingly depends on cloud services, the reliability of the underlying cloud systems has become a significant challenge. The well-established foundation for system reliability, such as consensus protocols, proves insufficient to immune cloud systems from catastrophic outages. In this talk, we share our experiences with real failures observed in production cloud systems, reveal that a new category of subtle failures, referred to as gray failures, is the source of major availability breakdowns and performance anomalies we see in cloud systems, and discuss how a new foundation can be laid out to address the challenges of gray failures.

Bio: Lidong Zhou is an Assistant Managing Director of Microsoft Research Asia, responsible for research in the System and Networking area. Previously, he was a Principal Researcher managing the Systems Research Group at Microsoft Research Redmond (2014-2017), a Principal Researcher managing the Systems Research Group at Microsoft Research Asia (2008-2014), and a Researcher at Microsoft Research Silicon Valley (2002-2008). His research has been advancing the state of art in both the theory and practice of scalable and reliable distributed systems powering on-line cloud services, while making key technical contributions to production large-scale platforms and services in Microsoft. Lidong is on the editorial board of ACM Transactions on Storage and served on the program committees for top system conferences such as SOSP and OSDI. He was the general co-chair of SOSP 2017 in Shanghai, after years of effort of bringing the top system conference to the Asia-Pacific region. Lidong received his Ph.D. and M.S. in Computer Science from Cornell University and B.S. in Computer Science from Fudan University.

Keynote3: Deconstructing RDMA-enabled Distributed Transactions: Hybrid is Better!

Speaker: Haibo Chen

Abstract:In this paper, I will present a systematic comparison between different RDMA primitives with a combination of various optimizations using representative OLTP workloads. More specifically, we first implement and compare different RDMA primitives with existing and our new optimizations upon a single well-tuned execution framework. This gives us insights into the performance characteristics of different RDMA primitives. Then we investigate the implementation of optimistic concurrency control (OCC) by comparing different RDMA primitives using a phase-by-phase approach with various transactions from TPC-C, SmallBank, and TPC-E. Our results show that no single primitive (one-sided or two-sided) wins over the other on all phases. We further conduct an end-to-end comparison of prior designs on the same codebase and find none of them is optimal. Based on the above studies, we build DrTM+H, a new hybrid distributed transaction system that always embraces the optimal RDMA primitives at each phase of transactional execution. Evaluations using popular OLTP workloads including TPC-C and SmallBank show that DrTM+H achieves over 7.3 and 90.4 million transactions per second on a 16-node RDMA-capable cluster (ConnectX-4) respectively, without locality assumption. This number outperforms the pure one-sided and two- sided systems by up to 1.89X and 2.96X for TPC-C with over 49% and 65% latency reduction.

Bio: Haibo Chen is a Professor at the School of Software, Shanghai Jiao Tong University, where he co-founds and leads the Institute of Parallel and Distributed Systems (IPADS) (http://ipads.se.sjtu.edu.cn). He currently also serves as the Chief Scientist for OS and directs the OS Kernel Lab. Haibo’s main research interests are building scalable and dependable systems software, by leveraging cross-layering approaches spanning computer hardware system virtualization and operating systems. He is currently the steering committee co-chair of ACM APSys and Chair of ACM SIGOPS ChinaSys, serves on program committees of ASPLOS 2019, EuroSys 2019, Usenix ATC 2019 and SOSP 2019 and the editorial board of ACM Transactions on Storage.

Keynote 4: RDMA in Data Centers: from Cloud Computing to Machine Learning

Speaker: Chuanxiong Guo

Abstract:Data centers are being built around the world to meet the exponentially increasing demands for cloud computing. The same increasing demands drive the networking speed increase from 10Gb/s to 100Gb/s or higher and the e2e latency reduction from milliseconds to microseconds. The traditional software-based TCP/IP, however, cannot keep up with the increasing demand. Consequently, RDMA (Remote Direct Memory Access), once introduced in HPC, now is experiencing a renaissance in Ethernet-based data centers, at a much larger scale. In this talk, we will discuss the safety and performance challenges that we have addressed in deploying RDMA at scale. We will also discuss the new opportunities brought by RDMA and explore the new role that RDMA will play in a more heterogeneous system and networking infrastructure for machine learning. Specifically, we will study how RDMA helps accelerate various DNN models in distributed DNN training.

Bio: Chuanxiong Guo is a director of the AI Lab of Bytedance Inc. Before that he was a Principal Researcher at Microsoft Research. He is currently working on data center networking and machine learning systems. Several of his envisions including DCN virtualization, DCN monitoring, and ServerSwitch generated both academic and industrial impacts. Several of the systems that he designed and implemented, including Pingmesh and RDMA/RoCEv2 were widely adopted by the industry.