RALB‐HC: A resource‐aware load balancer for heterogeneous cluster

Abstract

Summary In the heterogeneous computing environment, programmers map the applications either on CPUs or GPUs. However, this default mapping process does not produce improved results, particularly on the heterogeneous clusters. If one resource of the cluster is more compute capable, then most of the scheduling schemes favor that powerful device. In this scenario, the scheduling schemes overload the powerful resources while making all other compute resources remain under utilized. This load imbalance problem results in higher energy consumption and increased execution time. In this research, a novel Resource‐Aware Load Balancer for the Heterogeneous Cluster (RALB‐HC) is proposed that distributes workload based on resources computing capabilities and applications computing needs. The RALB‐HC uses supervised machine learning approach to classify applications using the static code‐features. The RALB‐HC framework comprises of two phases: (1) job mapping based on the availability of the resources and (2) the resource‐aware load balancing to achieve the higher resource utilization ratio. The experimental results on a large set of real‐world and synthetic workloads show that the RALB‐HC reduces execution time by 31.61%, increased resource utilization ratio by 67.8% and improved throughout 147.35% as compared to baseline scheduling schemes.

A load balance multi-scheduling model for OpenCL kernel tasks in an integrated cluster

“…The highly correlated data will result in lower accuracy because of lower predictive power, and hence, it should be evaluated empirically. Figure 4 shows that the features 0,6,15,12,16,8,22 and 20 have a negative correlation (Ahmed et al 2019a). The tree-based feature selection also validated the observation by ranking the same features on top and mentioned in Fig.…”

“…At the same time, there are some scientific applications (i.e., dot product or bread first search) that are inadequately performed on GPUs. The same applications often attain varying performance for different input data sizes (Khalid et al 2019;Ahmed et al 2019a). The applications that attain less gain should take a different strategy based on its input dimensions and type of operations.…”

“…Most of the research uses default classification settings. Default settings mostly result in under performance of the machine learning classification model (Ahmed et al 2019a;Ahmed et al 2020, Amrollahi et al 2020: the feature extraction, selection, model selection, hyper-parameter tuning and evaluation need to be optimized. Most research avenues do not consider multi-node application splitting or merger technique and the load balancing issue in the cluster of a heterogeneous environment.…”

“…TPOT The TPOT (Tchernykh et al 2016;Ahmed et al 2019a) uses genetic programming to construct features, selecting the machine learning model, and tune the selected model parameter. We provide the labeled data for each application to classification class to determine the device suitability.…”

An Evolutionary Model to Mine High Expected Utility Patterns From Uncertain Databases

“…Based on the MOEA/D framework, it can produce a set of equally disseminated solutions and has a great convergence as compared to the other MOEA algorithms such as NSGA-II [16] and SPEA-II [17]. In the field of data mining, the multiobjective evolutionary algorithms (MOEAs) are commonly utilized to solve the classification [13], and clustering and feature selection [13], [18] problem since the multi-criteria are considered into those problems that have to be optimized [13]. For example, the interesting pattern in the database depends upon the multiple measures, i.e., interestingness, support, confidence comprehensibility, and lift.…”