Sky Computing
Towards Utility Computing for the Cloud
News
October 23, 2023
Professors Alvin Cheung, Joseph Gonzalez, and Joseph Hellerstein win awards at VLDB Conference 2023
CS Associate Professor Alvin Cheung has won the 2023 Very Large Data Bases (VLDB) Early Career Research Contribution Award. The award, which includes a $2,000 prize, recognizes researchers who have made a significant impact through a specific contribution to the field since completing their Ph.D.
August 21, 2023
Come see us at SOSP ’23!
Accepted papers by Micah Murray, Matei Zaharia, Woosuk Kwon, Zhuohan Li, Siyuan Zhuang, Lianmin Zheng, Joseph Gonzalez, Hao Zhang, Ion Stoica, and Emma Dauterman at the 29th ACM Symposium on Operating Systems Principles.
August 18, 2023
‘Unlocking the new next frontier’: UC Berkeley researchers develop innovative AI ‘Gorilla’
Researchers from the Sky Computing lab and the Berkeley AI Research, or BAIR, recently released Gorilla, a large language model, or LLM, designed to revolutionize the way AI algorithms function. The researchers behind Gorilla include Shishir Patil, Tianjun Zhang, Prof. Joseph Gonzalez, and Xin Wang.
July 7, 2023
Come see us at OSDI ’23!
Ion Stoica will be giving the Keynote Address on Sky Computing at the 17th USENIX Symposium on Operating Systems Design and Implementation. Emma Dauterman, Siyuan Zhuang, Audrey Cheng, Romil Bhardwaj, and Zhuohan Li will be presenting at the technical sessions.
April 20, 2023
Ion Stoica and Joseph Gonzalez receive AWS AI Amazon Research Award on “A Unified Platform for Training and Serving Large Models”
Amazon Research Awards (ARA) provides unrestricted funds and AWS Promotional Credits to academic researchers investigating various research topics in multiple disciplines. Awardees, who represent 54 universities in 14 countries, have access to Amazon public datasets, along with AWS AI/ML services and tools.
Events
December 13, 2023
Database Seminar: Nikolaos (Nikos) Tziavelis (Northeastern University) – “Efficient Ranked Access over Joins”
Join queries over multiple tables can produce a huge output that is infeasible to compute. Even when it is feasible, it is often not efficient when users have particular preferences over the answers in the output and are interested in accessing only a small subset according to that ranking; either t…
December 8, 2023
Sky Systems Seminar: Zhihao Jia (CMU) – Building Systems for Fast, Efficient, and Affordable Large Language Models
The high computational and memory requirements of large language models (LLMs) make it challenging to train and serve them cheaply and efficiently. For example, serving a LLAMA-2-70B on NVIDIA A100 GPUs can only utilize 2% of the available compute resources. In this talk, I will present two systems …
December 6, 2023
Sky Security Seminar: Alin Tomescu (Aptos Labs) – UTT: Sensibly-Anonymous Decentralized Payments without zkSNARKs
We present UTT, a system for decentralized e-cash with accountable privacy….
December 6, 2023
Database Seminar: Mira Mezini (Technical University Darmstadt) – “Programming Abstractions for Safe and Secure Local-FirstSoftware”
Today’s computing infrastructure is massively distributed across back-end geo-replicated clouds and millions of increasingly powerful front-end devices. …
Publications
October 2023
Efficient Memory Management for Large Language Model Serving with PagedAttention.
October 2023
The Story of GraphLab – From Scaling Machine Learning to Shaping Graph Systems Research.
October 2023
Multiversion Hindsight Logging for Continuous Training.
August 2023
Energy-based Predictive Representations for Partially Observed Reinforcement Learning.
August 2023
Optimizing the cloud? Don’t train models. Build oracles!
August 2023
Efficient Data Sharing across Trust Domains.
August 2023
HOLMES: Efficient Distribution Testing for Secure Collaborative Learning.
August 2023
Test Accuracy vs. Generalization Gap: Model Selection in NLP without Accessing Training or Testing Data.
Ion Stoica – Featured Projects
SkyPilot
To comply with the increasing number of government regulations about data placement and processing, and to protect themselves against major cloud outages, many users want the ability to easily migrate their workloads between clouds. We propose doing so not by imposing uniform and comprehensive standards, but by creating a fine-grained two-sided market via intercloud brokers. SkyPilot is an intercloud broker that treats the cloud ecosystem not just as a collection of individual and largely incompatible clouds but as a more integrated Sky of Computing. SkyPilot enables users to run Machine Learning and Data Science batch jobs seamlessly on any cloud, reduce cloud costs substantially, tap into best-in-class hardware on different clouds, and enjoy higher resource availability.
Skyplane
Cloud applications are increasingly distributing data across multiple regions and cloud providers in response to privacy regulations, to take advantage of specialized hardware, and to prevent vendor lock-in. Unfortunately, wide-area bulk data transfers are often slow and expensive due to egress fees. This work aims to reduce both the latency and the cost of inter-cloud bulk transfer by using a variety of techniques, including overlay routing, multiple instances, multiple TCP connections, and taking advantage of different network tiers. Together, these techniques allow Skyplane to significantly improve object transfer throughput and lower the costs.
Natacha Crooks – Featured Project
Basil
Basil explores the design of SQL databases with high integrity and decentralized trust. How can traditional functionality like ACID transactions and SQL queries be efficiently implemented when trust is decentralized among n distinct parties, of which a subject can misbehave.
Joseph Gonzalez – Featured Project
Gorilla
The Gorilla project is designed to connect large language models (LLMs) with a wide range of services and applications exposed through APIs. Imagine if ChatGPT could interact with thousands of services, ranging from Instagram and Doordash to tools like Google Calendar and Stripe, to help you accomplish tasks. This may be how we interact with computers and even the web in the future. Gorilla is an LLM that we train using a concept we call retriever-aware training (RAT), which picks the right API to perform a task that a user can specify in natural language. Gorilla also introduces an Abstract Syntax Tree (AST) based sub-tree matching algorithm, which for the first time allows us to measure hallucination of LLMs!
Raluca Ada Popa – Featured Project
MC2
MC2 is a platform for running secure analytics and machine learning on encrypted data. With MC2, organizations can safely upload their confidential data to the cloud in encrypted form and securely compute analytics and machine learning without exposing the unencrypted data to the cloud provider. MC2 also enables secure collaboration among multiple organizations, where the data owners can use the platform to jointly analyze their collective data without revealing their individual data to each other.
Koushik Sen – Featured Project
FuzzFactory
FuzzFactory is domain-specific fuzz testing tool that generalizes coverage-guided fuzzing to domain-specific testing goals. FuzzFactory allows users to guide the fuzzer’s search process without having to modify the core search algorithm.
Sky Computing Story
Berkeley’s computer science division has an ongoing tradition of 5-year collaborative research labs. Recent labs included the AMPLab (ended in 2016) and the RISELab. These labs have had significant impact in both academia and industry. Past labs publish their research at top conferences in systems, databases, and machine learning. On the industrial side, AMPLab and RISELab fostered several successful startups (Databricks, Opaque, Ponder, Anyscale, to name a few). We are excited to announce the Berkeley Sky Computing Lab where we will strike to make cloud computing a true commodity.
Context
The Sky Computing Lab represents the next chapter of data-intensive systems research at Berkeley. Recent years have seen the explosion of cloud computing. Applications are moving their data and computation to the cloud; on-premise services are dying. In doing so, companies have to make difficult choices between the myriad of cloud providers, each with different services or hardware. Lock-in, whether through artificial migration costs, legal constraints or engineering baggage is real. In the Sky Computing Lab, we will leverage distributed systems, programming languages, security, and machine learning to decouple the services that a company wants to implement from the choice of a specific cloud. Much like the Internet today, cloud computing should be an undifferentiated commodity. Applications should run seamlessly on any or multiple clouds.
Mission
Our mission in the Sky Computing Lab is to transform the cloud into an undifferentiated commodity and ease application burden. As in previous labs, we’re all in — working on everything from basic research to software development, all in the Berkeley tradition of open publication and open source software. Our founding team consists of experts in distributed systems, machine learning, security and programming languages. We’ll use this space to lay out our ideas and progress as we go.
Commitment to Diversity
Sky Computing is guided by Berkeley’s Principles of Community and is committed to providing a safe and caring research environment for every member of our community. We believe that a diverse student body, faculty, and staff are essential to the open exchange of ideas that Sky Computing Lab is founded on.
Our head is in the cloud. We are heading for the SKY.
Sponsors
Affiliated Companies
