1 Guanhua Wang AMPLab, Soda Hall [email protected] EECS, UC Berkeley http://www.cs.berkeley.edu/~guanhua/ EDUCATION University of California, Berkeley 08/2015-06/2020 Ph.D. Student in Computer Science, AMPLab Advisor: Ion Stoica Hong Kong University of Science and Technology 09/2012-06/2015 Master of Philosophy in Computer Science and Engineering GPA: 4.0/4.3 (A+ range) Advisor: Lionel M. Ni Southeast University, China 08/2008-06/2012 Bachelor of Engineering in Computer Science and Technology GPA: 3.5/4, Ranking: 1/140 (overall ranking) First Prize of the Outstanding Graduation Project in Jiangsu Province (T op 0.1%) WORK EXPERIENCE Research Intern @ Microsoft Research (MSR), Redmond, WA 05/2017 – 08/2017 Speeding up communication among GPUs within DGX-1 machine, and across multiple machines. Mentors: Amar Phanishayee (Researcher), Chuanxiong Guo (Principle Researcher) Research Intern @ Microsoft Research (MSR), Redmond, WA 05/2016 – 08/2016 Improving image query accuracy (Microsoft’s Bing Image Search Engine) using CNTK Deep Learning Toolkits. Mentors: Sanjeev Mehrotra (Principle Software Architect), Jin Li (Partner Researcher Manager) RESEARCH PROJECTS Big Data Ø Gemini: Boosting Spark Performance with GPU accelerators Apache Spark is a MapReduce-based platform for large-scale data analysis. In Gemini project, we try to use GPU replace CPU for data processing in Spark system. The high-level idea is to leverage GPU’s higher parallelization in order to speed-up system performance. Ø Pipeline Shuffle: Reducing Shuffle I/O Latency in Spark The problem of Spark system is that, we need to first write the Mappers’ output data to the disk, and then send the data to the Reducers. The I/O latency on disk writing is huge. Pipeline Shuffle is trying to send the Mappers’ output data to Reducers once it is produced. At the same time, we write the shuffle data (i.e. Mappers’ output) to disk for fault tolerance. Ø Benefault: Scheduling via Task Preemption in Heterogeneous Environments Multiple jobs may content for scarce resources in a cluster (e.g. nodes with GPU). If long batch job occupies scarce resources, new short-interactive queries needs to wait for the long batch job to finish, which increase latency significantly. Existing solution is to kill/preempt the long task, and let short task run first. The victim/killed task is a waste of resources. Instead of killing long task for preemption, Benefault checkpoints and resumes victim tasks efficiently under the hood. Computer Networks Ø WiHear: We Can Hear You with Wi-Fi! The key insight of WiHear is to sense and recognize the radiometric impact of mouth movements on Wi-Fi signals. WiHear achieves hearing people talking just like lip-reading.