Coursera Flash Sale
40% Off Coursera Plus for 3 Months!
Grab it
Explore the theoretical foundations and practical applications of Processing-in-Memory (PIM) technology in this comprehensive lecture by Phillip Gibbons from Carnegie Mellon University. Learn how PIM addresses the growing challenge of data movement costs in computing by enabling computation directly on compute resources embedded within memory modules. Discover the fundamental differences between programming and algorithm design for PIM systems compared to traditional parallel or distributed computing environments. Examine the key trade-offs and limitations inherent in PIM architectures, with particular focus on the tension between minimizing communication overhead and achieving optimal load balance. Delve into PIM-optimized data structures as replacements for traditional indexing methods including B-trees, radix trees, and kd-trees, understanding how these specialized indexes provide provable performance guarantees regardless of query patterns or data distribution skew. Review experimental results from UPMEM's 2,560-module PIM system demonstrating performance improvements of up to 59x over existing PIM indexing approaches. Gain insights into ongoing research in PIM-friendly OLTP database design and implementation, positioning yourself at the forefront of this emerging computational paradigm that promises to revolutionize how we approach memory-intensive computing tasks.
