Title: Optimisation of cache replacement policy using extreme learning machine

Authors: Swapnita Srivastava; P.K. Singh

Addresses: Department of Computer Science and Engineering, Madan Mohan Malaviya University, Gorakhpur, 273010, India ' Department of Computer Science and Engineering, Madan Mohan Malaviya University, Gorakhpur, 273010, India

Abstract: In multiprocessors, all the cores ordinarily share the last level cache (LLC). The memory systems of multi-core CPUs are often affected by irregular memory access patterns. The gap between the memory system and LLC initiates the search for an effective cache replacement policy (CRP). Current processors utilise a variant of the least recently used (LRU) policy to identify which should replace victims. However, there is a significant gap between the LRU policy and Belady's MIN policy, which is the ideal CRP in all the scenarios. Since Belady's algorithm needs future knowledge, it is optimal but not practically possible. This paper shows how CRP can be trained from recent cache accesses to guide future replacement decisions. Recent research on anticipating the reuse of cache blocks has enabled substantial improvement in cache speed and efficiency. This paper presents the ELM-SSO policy that uses salp swarm optimisation (SSO) to optimise the weights coefficients of extreme learning machine (ELM) to perform cache replacement classification. Furthermore, the use of SSO in optimising the ELM is examined to increase system accuracy and overcome the drawback of traditional ELM. The findings demonstrate that the proposed ELM-SSO policy outperforms the traditional cache replacement policy in terms of improvement rate, cache hit rate and cache miss rate. The proposed ELM-SSO policy improves the system performance by 36.66%, 6.25%, 11.71%, 11.35%, 10.32% and 10.99% over optimal (OPT), LRU, least frequently used (LFU), logistic regression (LR), K-nearest neighbour (K-NN) and neural network (NN) respectively.

Keywords: computer architecture; cache memory; latency; eviction set; hit ratio; Belady replacement policy; least recently used; salp swarm optimisation; extreme learning machine.

DOI: 10.1504/IJNT.2023.134036

International Journal of Nanotechnology, 2023 Vol.20 No.5/6/7/8/9/10, pp.829 - 850

Received: 30 Dec 2021
Accepted: 29 Mar 2022
Published online: 10 Oct 2023 *

Full-text access for editors Full-text access for subscribers Purchase this article Comment on this article