Reproduce “Hydro: Adaptive Query Processing of ML Queries” 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
Gaurav Tarlok Kakkar, Jiashen Cao, Aubhro Sengupta, Joy Arulraj, Hyesoon Kim
Hydro, a system from Georgia Institute of Technology, introduces an Adaptive Query Processing (AQP) framework specifically designed for machine learning (ML) queries within database management systems. It dynamically optimizes predicate evaluation order and efficiently manages hardware resources for ML User-Defined Functions (UDFs), achieving multi-fold speedups (e.g., up to 11.52x) on diverse ML workloads compared to static baselines.
Key insights
- For concurrent ML UDFs, prioritizing faster-executing predicates (cost-driven routing) can be more effective than selectivity-driven or score-driven approaches in minimizing overall query time.
- Dynamically adjusting predicate priority based on cache hit rates (reuse-aware routing) is crucial for maintaining performance when UDF costs fluctuate mid-query due to caching.
- Batch-agnostic parallelization and data-aware load balancing significantly improve GPU utilization and throughput for ML UDFs, especially when UDF execution costs vary with input data characteristics.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.