Authors:
William Maas
1
;
Fábio Rossi
2
;
Marcelo Luizelli
3
;
Philippe Navaux
1
and
Arthur Lorenzon
1
Affiliations:
1
Institute of Informatics, Federal University of Rio Grande do Sul, Brazil
;
2
Campus Alegrete, Federal Institute Farroupilha, Brazil
;
3
Campus Alegrete, Federal University of Pampa, Brazil
Keyword(s):
Parallel Computing, Cloud Computing, Cost, Performance.
Abstract:
The growing popularity of data-intensive applications in cloud computing necessitates a cost-effective approach to harnessing distributed processing capabilities. However, the wide variety of instance types and configurations available can lead to substantial costs if not selected based on the parallel workload requirements, such as CPU and memory usage and thread scalability. This situation underscores the need for scalable and economical infrastructure that effectively balances parallel workloads’ performance and expenses. To tackle this issue, this paper comprehensively analyzes performance, costs, and trade-offs across 18 parallel workloads utilizing 52 high-performance computing (HPC) optimized instances from three leading cloud providers. Our findings reveal that no single instance type can simultaneously offer the best performance and the lowest costs across all workloads. Instances that excel in performance do not always provide the best cost efficiency, while the most afford
able options often struggle to deliver adequate performance. Moreover, we demonstrate that by customizing instance selection to meet the specific needs of each workload, users can achieve up to 81.2% higher performance and reduce costs by 95.5% compared to using a single instance type for every workload.
(More)