Spring 2011 (GF)

worked on: Large Scale data processing with MapReduce on GPU/CPU hybrid systems

read two papers about Phoenix, a mapreduce implementation for multi-core processors.
spent some days to figure out how to use Mark framework and run some samples, but couldn't fully understand. These works has been done:
- Configured system (windows) to run Mars, including cuda and SDK installation as well as VS9 configuring.
- Corrected some typos in the code (library mismatching)
- Asking authors about problems, and got this answer: "I must apologize that mars_v2 is buggy and complex, and we don't maintain the code base any more, I strongly recommend you to try the latest version on linux"
- tried to install mars_v2 on Linux, but it is still buggy and complex. It seems this frame work could run only with certaing configuration, and with older versions of CUDA.
Explored Mars to find its algorithm, and found in co-processing mode (Hybrid) they partition input data into two parts, one for CPU processing, the other for GPU processing. After the map stage, they merge data on CPU side, then dispatch data again to CPU workers and GPU workers.
Looked at phonix, another System for MapReduce Programming from Stanford. It was the comparison base for Mars.
- Spent 2 days for writing resume and being prepared for YouTube interview.

Fall 2010 (TA)

Submissions:
- SmartAd: a smart autonomous system for effective advertising in video (ICME 11)

worked on:
- Estimating the click-through rate for new ads with semantic and feature based similarity

algorithms

worked on:
- Accelarting online auction using GPU
- Estimating the click-through rate for new ads with semantic and feature based similarity

algorithms

submitted
- Accelerating online auctions with Optimized Parallel GPU based algorithms: Accelerating Vickrey-Clarke-Groves (VCG) Mechanism (proposal for GPU Gem book)

Courses:
- CMPT-705: Algorithm
- CMPT-771: Internet Architecture and Protocols