Cornell Systems Lunch

Spanner is a globally distributed transactional data management system that backs hundreds of mission-critical services at Google and operates on a massive scale: 10s of millions operations per second, 100s of petabytes, 1000s of databases. This talk will provide an overview of the system (NSDI 2012, SIGMOD 2017), zoom in on several current projects, and describe selected open problems.

Alex Khesin is a Principal Engineer at Google. During his 15 years at Google he has worked on databases, distributed file systems, and search engines. His core area of expertise is large distributed systems; his current focus is Spanner where he is leading a team developing Full Text Search subsystem.

He graduated from Cornell College of Engineering in 1994 with BS in Computer Science and received MS from New York University in 1998.

Isolation is an effective approach to building reliable systems from unreliable components. It is hard, if possible, to entirely eliminate bugs from software components; a more practical way is to assume the existence of bugs/vulnerabilities and live with them. This talk will present different ways to do system isolation, including software solutions like hypervisor, nested virtualization and same-level isolation, as well as hardware solutions like Intel SGX, AMD SEV and ARM TrustZone. It will also discuss the tradeoff between isolation and performance, and propose software/hardware co-designed systems to optimize the interaction between isolated domains.

Yubin Xia is an associate professor of IPADS group, Shanghai Jiao Tong University. His research areas include operating system, hypervisor, TEE (trusted execution environment), and computer architecture. He leads his team to implement T6, a TEE OS running on ARM TrustZone platform, which has been deployed on hundreds of millions of mobile phones over the world. His current main research interests lie in practical design and implementation for better system isolation as well as performance optimization.

Abstract: Transparency logs allow users to audit a potentially malicious service, paving the way towards a more accountable Internet. For example, in HTTPS, transparency logs are used to map each website to the history of its public keys, allowing websites to check if they have been impersonated. Yet, to achieve their full potential, transparency logs must be bandwidth-efficient when queried by users. Specifically, everyone should be able to (1) efficiently look up log entries by their key and (2) efficiently verify that the log remains append-only. Unfortunately, without additional trust assumptions, current transparency logs cannot provide both small-sized lookup proofs and small-sized append-only proofs. In fact, one of the proofs always requires bandwidth linear in the size of the log, making it expensive for everyone to query the log. In this paper, we address this gap with a new primitive called an append-only authenticated dictionary (AAD). We present efficient AADs from bilinear and RSA accumulators. Our construction is the first to achieve logarithmic sizes for both proof types.

Bio: Alin is a PhD candidate at MIT focusing on public-key distribution for HTTPS and secure messaging. His interests lie at the intersection of theory and practice: he enjoys applied cryptography and building systems. Recently, Alin worked on Catena, a system that leverages the Bitcoin blockchain to prevent equivocation attacks on public-key directories. In the past, Alin has worked on privacy-preserving file systems, private social networking and secure email. Alin's other interests include cryptocurrencies, anonymous networks and distributed systems in general.

Abstract: As data-driven applications continue to push the limits of data center infrastructure, achieving consistently low communication latency has emerged as a long-term challenge. Data-driven computations rely heavily on synchronous remote procedure calls (RPCs) to conduct computations close to data. As data center networks approach microsecond forwarding delays, existing hardware and software techniques become inadequate to hide these latencies. Traditionally, hardware has been used to hide nanoseconds, while operating systems can hide milliseconds, with no solution adequate for microsecond-scale latencies. Solutions to the communication latency challenge will require a combination of hardware, software, and networking techniques, laying out a research direction for years to come.

This talk focuses on software solutions to the communication latency problem. I will describe how communication latency became so significant, why existing hardware and software approaches cannot fully solve the latency problem, and why a future solution is unlikely to come from hardware or software alone. I then survey the progress to date in software solutions to the problem. If time permits, I will give an outlook on proposed hardware techniques and how they might combine with software solutions.

Simon is currently co-designing operating system networking and storage stacks with new hardware technologies to push server I/O efficiency an order of magnitude beyond today's capabilities. He is also working on networking issues that arise when pushing performance far beyond current capabilities. His group open-sources all built research systems.