Processing-in-Memory: Theory and Practice
Tuesday, November 18, 2025 - 1:00pm to 2:00pm
Location:
32-G575
Speaker:
Hongbo Kang
Biography:
Hongbo Kang is a PhD candidate in the Department of Computer Science at Tsinghua University. His research focuses on theoretically and practically efficient parallel algorithms, particularly for novel hardware. His previous research centered on processing-in-memory (PIM), as part of a project led by Professor Phil Gibbons of Carnegie Mellon University. In this area, his contributions include a theoretical model and the design and implementation of efficient parallel algorithms. His work demonstrates that collaboration between traditional CPUs and in-memory processors can significantly reduce data movement, improving both asymptotic and practical worst-case performance. His PIM-tree, a PIM-optimized ordered index, received the Best Paper Runner-Up Award at VLDB 2023. His research interests also include non-volatile memory systems and learned indexes.
Seminar group:
Data movement is fast becoming the dominant cost in computing. Processing-in-Memory (a.k.a., near-data-processing), an idea dating back to 1970, is now re-emerging as a key technique for reducing costly data movement, by enabling computation to be executed on compute resources embedded in memory modules.
While there has been considerable recent work on the architecture/technology side of PIM, there has been very little work addressing the theory/programming side. This leaves open critical questions: How should programming and algorithm design for PIM differ from traditional parallel or distributed settings? What are its fundamental limitations and trade-offs?
This talk presents our recent results addressing these questions. As a driving application kernel, we focus on designing PIM-friendly indexes, i.e., what are PIM-optimized replacements for B-trees, radix trees, and kd-trees? Our indexes address head-on the inherent tension between minimizing communication and achieving load balance in PIM systems, providing provable guarantees regardless of query or data skew. Experimental results on UPMEM's 2048-module PIM system demonstrate that our indexes outperform prior PIM indexes by up to 59x. Finally, we will outline several future research directions we are exploring.