Efficient Processing and Delivery of Multimedia Data

Abstract

The explosion of multimedia data on the Internet in recent years has greatly enriched people's online experience. However, it also poses a great challenge to analyze and process, and then deliver such content to the worldwide audience. This dissertation presents novel approaches to improve the overall efficiency of the stack by tailoring software design to hardware properties, as well as optimize systems by exploiting workload characteristics using learning-based approaches.

First, to improve the caching performance of the flash-memory caches for content delivery network, this thesis proposes RIPQ, a framework for efficient and advanced caching with flash memory. Traditional implementations of these algorithms generate random writes that perform poorly on flash devices, decreasing the device's performance and lifespan. RIPQ overcomes this issue by aggregating small writes, colocating items with similar priorities, and perform lazy updates to achieve low over- head. By providing a priority queue interface, it allows a variety of caching algorithms to be easily implemented.

Second, this thesis proposes Chess, which uses popularity prediction for higher quality video streaming. Although better encodings improve video streaming, they are also compute-intensive, and it is infeasible to encode all videos uploaded to Face- book with the highest quality codec. However, because the accesses to videos are highly skewed, we may obtain most of the benefit by only running the compute- intensive encoding on a small portion of popular videos, and the challenge lies in how to accurately and scalably run popularity prediction to detect those videos before- hand. Chess meets this demand by designing an approximate but fast base predictor with the access history information, and using an online learning method to combine multiple such predictors as well as the social signals to boost accuracy.

Lastly, this thesis investigates how to accelerate deep learning models on many-core CPUs. Deep learning is now widely used for analyzing multimedia data, but it is compute-intensive, which constitutes its major bottleneck. The manycore CPU, combining both high FLOPS and a flexible computing model, is a promising solution to this problem. However, existing frameworks are still mainly optimized for GPU, and do not run efficiently on this architecture. To overcome this issue, this thesis proposes Graphi, the first attempt to accelerate the execution of computation graphs for deep learning models on this architecture. Graphi determines the optimal parallel settings with a profiling step, runs concurrent operations with low contention, and further reduces execution makespan with critical-path first scheduling. This thesis demonstrated that these techniques can achieve significant speedups over TensorFlow on manycore CPUs.