EECS 280 Becomes Third Largest Course at U-M
University of Michigan CSE, April 2017
Autonomous vehicle security: Autonomous vehicle software, such as Apollo, is implemented as a distributed system, with multiple modules communicating using message passing. I am creating tools that can automatically identify vulnerabilities in the ways these modules interact, which developers can use to secure these critical systems.
GPU multi-kernel execution: When multiple kernels are run at the same time on the same GPU, they can often acheive higher throughput than when they are run consecutively, because there are times they have complementary resource requirements. During other times, they can interfere. My work finds resource paritions that limit the impact of interference.
GPU register file design: GPUs need to have hundreds of kilobytes of register file, because so many threads are executing simultaneously. However, not many of these registers are accessed in any given period of time. RegLess (published MICRO 2017) is a technique to save energy using a much smaller register structure that stores only active registers.
GPU memory coalescing: Nearby threads on a GPU tend to access nearby locations in memory, allowing requests to the same cache lines to be merged to increase memory throughput. WarpPool (published MICRO 2015) used a new type of memory locality between loads made by different thread groups to merge more requests.
Scratch That (But Cache This): A Hybrid Register Cache / Scratchpad for GPUs
Jonathan Bailey, John Kloosterman, Scott Mahlke
RegLess: Just-in-Time Operand Staging for GPUs
John Kloosterman, Jonathan Beaumont, D. Anoushe Jamshidi, Jonathan Bailey, Trevor Mudge, Scott Mahlke
WarpPool: Sharing Requests with Inter-Warp Coalescing for Throughput Processors
John Kloosterman, Jonathan Beaumont, Mick Wollman, Ankit Sethia, Ron Dreslinski, Trevor Mudge, and Scott Mahlke
local_malloc: malloc() for OpenCL local memory (poster)
John Kloosterman, Joel Adams
ACM SRC Poster, SC13