中央研究院 資訊科學研究所

With hardware accelerators (e.g, TPUs and NPUs) shrinking the execution time in computation kernels, the bottleneck in machine learning applications has shifted. Reading inputs from the storage device and shuffling datasets become the most critical stage in the pipeline of train machine learning models. To address this emerging bottleneck and improve application performance, we need to take a holistic design approach, instead of the conventional single-point approach.

In talk, Hung-Wei will share his research experiences in revisiting the interaction between the storage device and the host computer to accelerate machine learning applications. To reduce the latency of reading input, Hung-Wei’s research team redefined the interface of an SSD to change data resolutions. To address the overhead in data shuffling, Hung-Wei’s research team also revamped the NVMe storage driver and eliminate unnecessary data movement between the processor and the system main memory.

In addition to design better systems for machine learning, we can also use machine learning results to design a better system. In this talk, Hung-Wei will also talk about his recent research in applying machine learning models to improve the lifetime of SSDs by 19%.

Hung-Wei is currently an assistant professor in the Department of Computer Science and Department of Electrical and Computer Engineering, NC State University where he is now leading the Extreme Storage & Computer Architecture Laboratory. He is recognized by facebook faculty research award for his research in accelerating data-intensive applications through revisiting the storage system design. Prior to joining NCSU, Hung-Wei was a postdoctoral scholar of the Non-volatile Systems Laboratory with Professor Steven Swanson and a lecturer of the Department of Computer Science and Engineering at University of California, San Diego. His thesis work with Professor Dean Tullsen, data-triggered threads, was selected by IEEE Micro "Top Picks from Computer Architecture" in 2012.