knor: A NUMA-Optimized In-Memory, Distributed and Semi-External-Memory k-means Library

TL;DR

The paper introduces knor, a NUMA-optimized, scalable k-means library with in-memory, distributed, and semi-external memory modules, significantly improving performance and scalability for large datasets on modern hardware.

Contribution

It presents a novel NUMA-aware parallelization scheme and a versatile library that outperforms existing solutions for large-scale k-means clustering.

Findings

knor in-memory module boosts performance by over 10x on single machines

semi-external memory module scales to billions of points using SSDs

outperforms commercial distributed solutions by more than an order of magnitude

Abstract

k-means is one of the most influential and utilized machine learning algorithms. Its computation limits the performance and scalability of many statistical analysis and machine learning tasks. We rethink and optimize k-means in terms of modern NUMA architectures to develop a novel parallelization scheme that delays and minimizes synchronization barriers. The \textit{k-means NUMA Optimized Routine} (\textsf{knor}) library has (i) in-memory (\textsf{knori}), (ii) distributed memory (\textsf{knord}), and (iii) semi-external memory (\textsf{knors}) modules that radically improve the performance of k-means for varying memory and hardware budgets. \textsf{knori} boosts performance for single machine datasets by an order of magnitude or more. \textsf{knors} improves the scalability of k-means on a memory budget using SSDs. \textsf{knors} scales to billions of points on a single machine, using a fraction of the resources that distributed in-memory systems require. \textsf{knord} retains \textsf{knori}'s performance characteristics, while scaling in-memory through distributed computation in the cloud. \textsf{knor} modifies Elkan's triangle inequality pruning algorithm such that we utilize it on billion-point datasets without the significant memory overhead of the original algorithm. We demonstrate \textsf{knor} outperforms distributed commercial products like H$_2$O, Turi (formerly Dato, GraphLab) and Spark's MLlib by more than an order of magnitude for datasets of $10^7$ to $10^9$ points.