Linpeng Tang will present his pre FPO, " Efficient processing and delivery of multimedia data" on Friday, October 20, 2017 at 2pm in CS 402. The members of his committee are as follows: Readers: Prof. Mike Freedman, Prof. Wyatt Lloyd Non-readers: Prof. Jen Rexford, Prof. Kyle Jamieson Advisor: Prof. Kai Li Everyone is invited to attend his talk. The talk title and abstract follow below. Title: Efficient processing and delivery of multimedia data Abstract: With the rise of smart mobile devices and faster networks, multimedia data, such as photos and videos, are becoming a central part of people's digital lives in recent years. Meanwhile, new challenges have also emerged on how to process and delivery such data efficiently. This thesis I will investigate three challenging problems related to this area. The first part is about how to cache photo content in the large-scale content delivery network. We have designed and implemented RIPQ, an advanced photo caching framework on SSD. The Flash Translation Layer on SSD devices doesn't cope well with the random writes generated by advanced caching algorithms, causing lower througput and reduced device lifespan. To solve this issue, RIPQ aggregates small random writes, co-locates similarly prioritized content, and lazily moves updated content to reduce device overhead. Our evaluation on Facebook’s photo trace shows that two advanced caching algorithms running on RIPQ increase hit ratios up to ~20% over the current FIFO system, incur low overhead, and achieve high throughput. It has also since been deployed on the Facebook photo serving stack. The second part is on how to improve the streaming quality of social media video content. I will present CHESS, a scalable and accurate video popularity prediction algorithm. By accuratly predicting the popular videos, CHESS allows us to focus intensive processing on a small amount of videos, while improving user experience for most viewers. CHESS only requires constant space per video while delivering superior predictions using a combination of historical access patterns with social signals, in a unified online learning framework. In evaluation we find that re-encoding popular videos predicted by CHESS enables 80% Facebook user watch time to benefit from a computing intensive encoding, with only one-third computing overhead compared to a recent production heuristic. Finally, deep learning emerges as the most effective machine learning method for analyzing many types of content. Computing is a bottleneck for performing such analytics at scale, and I will present Graphi, a generic and high-performance execution engine for deep learning computation on many-core CPUs. We find that the key is finding the right balance between intra/inter operation parallelization, while minimizing resource interference across the weak cores. We therefore have designed and implemented Graphi using intelligent scheduling algorithms with feedback from a profiler, and a thread management system. Our experiments on show that the various training times with Graphi are 2.1× to 9.5× faster than those with TensorFlow on a 68-core Intel Xeon Phi processor.