Search results for: parallelization

Authors: Malihe Aliasgari, Yousef Nejatbakhsh

Abstract:

The era of Big Data and the immensity of real-life datasets compels computation tasks to be performed in a distributed fashion, where the data is dispersed among many servers that operate in parallel. However, massive parallelization leads to computational bottlenecks due to faulty servers and stragglers. Stragglers refer to a few slow or delay-prone processors that can bottleneck the entire computation because one has to wait for all the parallel nodes to finish. The problem of straggling processors, has been well studied in the context of distributed computing. Recently, it has been pointed out that, for the important case of linear functions, it is possible to improve over repetition strategies in terms of the tradeoff between performance and latency by carrying out linear precoding of the data prior to processing. The key idea is that, by employing suitable linear codes operating over fractions of the original data, a function may be completed as soon as enough number of processors, depending on the minimum distance of the code, have completed their operations. The problem of matrix-matrix multiplication in the presence of practically big sized of data sets faced with computational and memory related difficulties, which makes such operations are carried out using distributed computing platforms. In this work, we study the problem of distributed matrix-matrix multiplication W = XY under storage constraints, i.e., when each server is allowed to store a fixed fraction of each of the matrices X and Y, which is a fundamental building of many science and engineering fields such as machine learning, image and signal processing, wireless communication, optimization. Non-secure and secure matrix multiplication are studied. We want to study the setup, in which the identity of the matrix of interest should be kept private from the workers and then obtain the recovery threshold of the colluding model, that is, the number of workers that need to complete their task before the master server can recover the product W. The problem of secure and private distributed matrix multiplication W = XY which the matrix X is confidential, while matrix Y is selected in a private manner from a library of public matrices. We present the best currently known trade-off between communication load and recovery threshold. On the other words, we design an achievable PSGPD scheme for any arbitrary privacy level by trivially concatenating a robust PIR scheme for arbitrary colluding workers and private databases and the proposed SGPD code that provides a smaller computational complexity at the workers.

Keywords: coded distributed computation, private information retrieval, secret sharing, stragglers

Procedia PDF Downloads 91

Authors: Francesca Schiavello

Abstract:

This work aims to create a hybrid algorithm, combining Quantum Approximate Optimization Algorithm (QAOA) with an Evolutionary Algorithm (EA) in the place of traditional gradient based optimization processes. QAOA’s were first introduced in 2014, where, at the time, their algorithm performed better than the traditional best known classical algorithm for Max-cut graphs. Whilst classical algorithms have improved since then and have returned to being faster and more efficient, this was a huge milestone for quantum computing, and their work is often used as a benchmarking tool and a foundational tool to explore variants of QAOA’s. This, alongside with other famous algorithms like Grover’s or Shor’s, highlights to the world the potential that quantum computing holds. It also presents the reality of a real quantum advantage where, if the hardware continues to improve, this could constitute a revolutionary era. Given that the hardware is not there yet, many scientists are working on the software side of things in the hopes of future progress. Some of the major limitations holding back quantum computing are the quality of qubits and the noisy interference they generate in creating solutions, the barren plateaus that effectively hinder the optimization search in the latent space, and the availability of number of qubits limiting the scale of the problem that can be solved. These three issues are intertwined and are part of the motivation for using EAs in this work. Firstly, EAs are not based on gradient or linear optimization methods for the search in the latent space, and because of their freedom from gradients, they should suffer less from barren plateaus. Secondly, given that this algorithm performs a search in the solution space through a population of solutions, it can also be parallelized to speed up the search and optimization problem. The evaluation of the cost function, like in many other algorithms, is notoriously slow, and the ability to parallelize it can drastically improve the competitiveness of QAOA’s with respect to purely classical algorithms. Thirdly, because of the nature and structure of EA’s, solutions can be carried forward in time, making them more robust to noise and uncertainty. Preliminary results show that the EA algorithm attached to QAOA can perform on par with the traditional QAOA with a Cobyla optimizer, which is a linear based method, and in some instances, it can even create a better Max-Cut. Whilst the final objective of the work is to create an algorithm that can consistently beat the original QAOA, or its variants, due to either speedups or quality of the solution, this initial result is promising and show the potential of EAs in this field. Further tests need to be performed on an array of different graphs with the parallelization aspect of the work commencing in October 2023 and tests on real hardware scheduled for early 2024.

Keywords: evolutionary algorithm, max cut, parallel simulation, quantum optimization

Procedia PDF Downloads 31