Performance Analysis and Acceleration of Learned Hash-Indexes on Supercomputing Architectures

Identifier:  TFG:9490

Handle: https://hdl.handle.net/20.500.11797/TFG9490

Authors:  Palazón Balmaseda, Montserrat

Abstract:
The exponential growth of genomic data challenges the performance of genome analysis tools, especially during long-read mapping. Hash tables, widely used in the seeding phase, suffer from irregular memory access and poor cache locality on modern hardware. Learned hash indexes, using models like RMI to predict key locations, offer a promising alternative. This thesis analyzes their performance and shows that, with software optimizations such as batching, prefetching, and vectorization, learned indexes can outperform traditional ones. Our results show that they achieve a 2.90× speedup, a 3.87× reduction in MPKI, and reduce memory-bound cycles from 66.80% to 13.65%, unlocking their potential on modern processors.