Senior Autonomy Engineer @ Uber Advanced Technologies Group
Graduate Research Assistant @ Carnegie Mellon University
Research Intern @ HP Labs
Education:
Bachelor of Engineering (BE) @
Tsinghua University
About:
My research topic is software systems for big data analytics. For more details, see my homepage http://users.ece.cmu.edu/~hengganc/
RESEARCH PROJECTS:
LazyTable:
A parameter server framework for Distributed Machine Learning. It manages the global shared model parameters for data parallel ML applications, where distributed application workers iterate on partitioned input data in parallel and concurrently make adjustment to the
My research topic is software systems for big data analytics. For more details, see my homepage http://users.ece.cmu.edu/~hengganc/
RESEARCH PROJECTS:
LazyTable:
A parameter server framework for Distributed Machine Learning. It manages the global shared model parameters for data parallel ML applications, where distributed application workers iterate on partitioned input data in parallel and concurrently make adjustment to the shared model parameters based on the input and current parameter values. We build several representative ML applications on top of LazyTable, including PageRank, Topic Modeling, and Collaborative Filtering.
Stale Synchronous Parallel (SSP):
A model for synchronizing the progress of parallel workers. SSP is a middle ground between BSP (where the workers wait at each barrier) and the fully asynchronous parallel (where the workers never wait). SSP allows each worker to be a bounded number of iterations ahead of the slowest one, thus making data staleness a tunable parameter and allowing users to explicitly trade data freshness for speed.
Papers published in NIPS'13 and ATC'14. SSP is supported in LazyTable.
Parameter server for iterative ML:
Many iterative ML algorithms (PageRank, SGD, and Gibbs Sampling) result in the same (or nearly the same) sequence of accesses to parameters repeating each iteration. We propose a method to collect this repeating access pattern from the application as well as specializations that exploit this information to speed up parameter servers. Experiments show that these specializations reduce per-iteration time by 33% to 98% for our applications.
Paper published in SoCC'14. These ideas are implemented in a new version of LazyTable.
PhD Candidate @ From September 2012 to Present (3 years 2 months) Research Intern @ From May 2014 to August 2014 (4 months) Palo Alto, CA
Doctor of Philosophy (PhD), Electrical and Computer Engineering @ Carnegie Mellon University From 2012 to 2018 Bachelor of Engineering (BE), Electronic Engineering @ Tsinghua University From 2008 to 2012
Websites:
http://users.ece.cmu.edu/~hengganc/
Looking for a different
Henggang Cui?
Get an email address for anyone on LinkedIn with the ContactOut Chrome extension