Avanced Systenm Lab - 2022

This repo contains the code for the Advanced System Lab course held @ ETHZ

Goal

Optimize the Non-Negative Matrix Factorization with multiplicative updates algorithm proposed by Lee and Seung. The optimization is done based on the given configuration and uses AVX2 instructions.

Team

• Daniele Coppola dcoppola@student.ethz.ch
• Viktor Gsteiger vgsteiger@student.ethz.ch (18-054-700)
• Masa Nesic mnesic@student.ethz.ch
• Matteo Oldani moldani@student.ethz.ch

Content

[[TOC]]

Configuration:

The code in this repository has been developed for the following system configuration. We suggest using the same or a similar configuration to achieve comparable results.

• CPU: Intel(R) Core(TM) i5-6600K CPU @ 3.50GHz
• L1d cache: 128 KiB
• L1i cache: 128 KiB
• L2 cache: 1 MiB
• L3 cache: 6 MiB
• DRAM: 16 GB DDR4 @ 2133 MT/s (2 x 8 GB)

NOTE: Turbo boost has been disabled.

Installation:

The codebase requires the prior installation of either OpenBLAS or Intel MKL ( can be installed with sudo apt install intel-mkl on Linux).

Other than that the codebase does not require any other installations.

Operating Instructions:

The operating of the codebase is relatively straightforward thanks to the makefile included in the /code folder. The different parts of the code can be compiled using different commands in the makefile. Following we explain the commands that can be used:

• basline: Compile and link the baseline implementations. The resulting binary can be found in build/baseline. The binary can be run either with the parameters <baseline number> <m> <n> <r> or <baseline number> <file-path> and it executes a test run of the baselines and returns the error of the factorization.
• performance: Compile and link the performance calculation. The resulting binary can be found in build/performance. The binary can be run with the parameters <program number [1,2,3,4,5]> <min size matrix> <max size matrix> <number of test> <output_file>[?] and either returns the performance results or saves them to the provided output file.
• roofline: Compile and link the roofline calculation. The resulting binary can be found in build/roofline. The binary can be run with the parameters <program number [1,2,3,4,5]> <input_size> <number of test> and should be used together with a roofline analyser tool such as Intel Advisor or similar
• tests: Compile and link the testing and performance evaluation utility. The resulting binary can be found in build/tests. The binary can be run with the following parameter <path to testfiles>[?] and it executes the testing and performance evaluation. The results will be printed to the standard output where for each registered function the test result and a cycle count will be returned.
• optimisations: Compile and link the created optimisations. The resulting binary can be found in build/optimisations. This does not include a main file and can therefore not be run.
• least_square: Compile and link the least squares NNMF implementation. The resulting binary can be found in build/least_square. The binary can be run either with the following parameters <m> <n> <r> or -1. The first parameters will run a performance evaluation with the provided matrix sizes while the second will run a predefined increasing matrix size evaluation for plotting.
• clean: Clean the repository from any binary files