Reproduce “PIM-tree: A Skew-resistant Index for Processing-in-Memory” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Hongbo Kang, Yiwei Zhao, Guy E. Blelloch, Laxman Dhulipala, Yan Gu, Charles McGuffey, Phillip B. Gibbons
PIM-tree introduces a skew-resistant ordered index for Processing-in-Memory (PIM) systems, which dynamically assigns tasks between the host CPU and PIM modules to maintain high throughput across varying data and query distributions. The method achieves up to 69.7x higher throughput than previous PIM indexing solutions and offers competitive performance with traditional non-PIM indexes, even under extreme data skew.
Key insights
- Dynamic load balancing between a powerful host CPU and distributed PIM modules is crucial for achieving robust skew-resistance in PIM systems, effectively utilizing the strengths of both architectures.
- Strategic, partial replication of search subtrees ('shadow subtrees') within PIM modules can significantly reduce communication costs without incurring prohibitive space overheads.
- Optimizations like 'chunking' to improve memory locality and reduce data structure height are highly effective in enhancing the practical performance of PIM-based data structures.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.