There is a trend towards using very large deep neural networks (DNN) to improve the accuracy of complex machine learning tasks. However, the size of DNN model and its training speed are both limited by the use of a single GPU device. We are working on two systems, Tofu and DGL, that allow programmers train models across many GPU devices with the same ease as training on a single GPU.
Another challenge facing DNN deployment today is how to serve complex models in real time. Each generation of new GPU comes with increased hardware parallelism, which necessitates large batch sizes but results in high inference latency. We developed the BatchMaker inference system to enable low latency batched inference.
Minjie Wang, Chien-Chin Huang, Lingfan Yu, Jinyang Li
Many large-scale web applications today run across multiple data centers to improve fault tolerance and reduce network delays to users. These web applications need a geo-distributed storage backend to store and share data. Unfortunately, unlike storage systems operating within a single data center, a geo-distributed storage system faces the unpleasant tradeoff of consistency vs. performance because of large inter-data-center communication delay.
We study this fundmental consistency vs. performance tradeoff in terms of new consistency and programming models for geo-replicated storage.
Shuai Mu, Tiger Zhaoguo Wang, Jinyang Li
How can a machine specify a computation to another one and then, without executing the computation, check that the other machine carried it out correctly? This question is motivated by the cloud and other third-party computing models.
The Pepper project is tackling this question by reducing to practice powerful tools from complexity theory and cryptography: probabilistically checkable proofs (PCPs), efficient arguments, and related notions. Recently, this research area has seen a number of exciting results (ours and others); together, they suggest that in the future, PCP-related machinery could be a real tool for building actual systems.
Srinath Setty, Riad Wahby, Michael Walfish
Tomas Wies, Dennis Shasha, and graduate student Siddharth Krishna are developing a verifiable template framework for concurrent algorithms on search structures (data structures such as hash, B-tree, lists, binary search trees and hybrids of these that support key-value stores). Stratos Idreos, Michael Kester, and Dennis Shasha are using this template data structure to develop high performance adaptable data structures.
Juliana Freire, Fernando Chirigati and Dennis work on systems to support computational reproducibility using provenance and operating system support.
Dennis Shasha works on applied projects such as a database and tools to support cross-linguistic analysis, algorithms for making magnetic resonance imagery faster, and a database to support research in millimeter wave wireless, and a tool for automatic database tuning.