Banca de QUALIFICAÇÃO: JOÃO BATISTA FERNANDES

Uma banca de QUALIFICAÇÃO de DOUTORADO foi cadastrada pelo programa.
STUDENT : JOÃO BATISTA FERNANDES
DATE: 17/12/2021
TIME: 15:00
LOCAL: meet.google.com/mwm-ikgs-ezw
TITLE:

Task-homogeneous load balancing in heterogeneous distributed systems with adaptive asynchronous work-stealing

KEY WORDS:

Work-Stealing, scheduling, load balancing, A2WS, MPI.

PAGES: 30
BIG AREA: Ciências Exatas e da Terra
AREA: Ciência da Computação
SUMMARY:

The use of clusters and data centers is essential to some industry and science fields, including those with a higher amount of data to process, like meteorology studies and the oil and gas industry. Over the years, the demands for high-speed processing have increased, requiring updates of clusters to include new technologies. Many of these clusters already include different hardware for specific purposes, such as FPGA and GPU. However, the clusters can also have nodes with previous hardware versions, including CPUs of different companies with different technologies. Then, the clusters are becoming heterogeneous. One crucial challenge of heterogeneous clusters comes together with parallel programming, load balancing. For this, there are many scheduling methods for specific application situations. Among them, dynamic schedules are more flexible. Without using previous information, they are more practical to heterogeneous clusters where the performance of tasks is not previously known. In our purpose, we will work with heterogeneous environments using a dynamic scheduler method called work-stealing. This method makes the scheduling on runtime when faster processes steal tasks from slower ones. Our objective is to improve the work-stealing performance on heterogeneous clusters. For this, there are two main challenges to be considered: A smart distribution of the tasks, considering the speed of each node, and the immediate scheduling of executions, avoiding synchronous methods. To this end, we will introduce the Adaptive Asynchronous Work-Stealing (A2WS). This method takes advantage of information measured in application runtime to estimate the number of tasks each process needs to balance the execution. In our work, we already have some results that show the gain of our method related to the other similar schedules, such as the Cyclic Token Work-Stealing, Leader-Workers, and a static distribution. The results show, simultaneously, 11.31%, 13.94%, and 16.69% of runtime gain. We have some purposes for the future, including the tests in diverse real heterogeneous environments. We also intend to include tests using applications that demand more network usage to find other bottlenecks on our purpose.

BANKING MEMBERS:
Presidente - 1673543 - SAMUEL XAVIER DE SOUZA
Externa ao Programa - 1434796 - IDALMIS MILIAN SARDINA MARTINS
Externo à Instituição - ITALO AUGUSTO SOUZA DE ASSIS - UFERSA