Alireza Asl Nemati

This study compares three Federated Learning (FL) algorithms—FedAvg, FedProx, and MOON—by evaluating their performance in both IID and non-IID settings. We found that FedAvg performs best in IID scenarios, offering quick convergence and high accuracy. However, in non-IID settings, MOON stood out as the top performer, thanks to its contrastive learning method, providing better stability and accuracy across heterogeneous data. FedProx showed improvements over FedAvg in handling non-IID data but was less effective than MOON. Our findings suggest that for environments with IID data, FedAvg is ideal, while MOON is more suitable for non-IID cases. We also highlight the need for further research into personalized FL, regularization techniques, and the integration of multimodal data.

The Longest Palindromic Substring (LPS) problem is a classic and fundamental task in string processing with broad applications in bioinformatics, natural language processing, data security, and compression. Manacher's algorithm is the most efficient sequential solution, achieving linear time O(n); however, for very large inputs, its execution can still be time-consuming in modern high-performance computing (HPC) environments. In this work, we first implement the sequential Manacher algorithm and evaluate it on a supercomputing platform: the results confirm the expected linear behavior, yet also indicate that merely using multi-core hardware without effective parallelization is insufficient. We then survey major prior parallel approaches for palindrome-related problems and, based on this study, design and implement a parallel variant of Manacher's algorithm using the Message Passing Interface (MPI). Our approach distributes the input across processes and merges local results with minima