Alapfogalmak
Efficiently accelerate learned index structures using memoization-based incremental training and FPGA acceleration.
Kivonat
Learned indexes use machine learning models to map keys to positions in key-value indexes. Existing systems face performance bottlenecks during retraining, especially for string keys. SIA introduces an algorithm-hardware co-designed solution to reduce retraining complexity and accelerate training using FPGA. SIA offers higher throughput compared to state-of-the-art systems like ALEX, LIPP, and SIndex on real-world benchmarks.
Statisztikák
SIA-accelerated learned indexes offer 2.6× and 3.4× higher throughput on YCSB and Twitter cache trace benchmarks.
Increased retraining times negatively impact inference throughput.
SIA provides a substantial performance boost compared to software-only solutions.
Idézetek
"We develop a memoization-based incremental training scheme."
"SIA combines algorithmic and hardware innovations for high query throughput."
"Compared to baseline learned indexes, SIA offers significant speedups."