About Me

I am a Ph.D. candidate in the Computer Science Department (DCC) at Federal University of Minas Gerais (UFMG), Brazil. Before that I received my M.Sc. at the same institution. And before that I received my B.Sc. in Computer Science from Federal University of Uberlândia (UFU), Brazil.

Broadly, I am interested in research problems concerning systems’ performance, specially large-scale distributed systems. Recently, I have been working on performance diagnosis and reconfiguration of data parallel programming models (a.k.a. Big Data) and in particular, the following challenges have been catching my attention (and stealing some of my sleep :P):

• Task Model in Data-Parallel Frameworks: Data-itensive frameworks (e.g., Spark and Flink) are based on specific task models that describe how resources (memory, CPU, disk, etc.) in parallel computations are consumed.

  • Is there a more deterministic model that would make easy to identify which resources are more important throughout tasks’ execution?
  • Would this model facilitate resource overlapping among concurrent tasks?
  • Which would be the gains (if any) in doing that, considering modern operating systems?

• Scalable (Single-)Graph Pattern Mining: Graph pattern mining is every task that regards discovering, matching, searching and summarizing “interesting” subgraph structures from a single (large) input graph. Instead of worrying about definitions of what is “interesting”, I am personally interested in the development of execution models for supporting those kind of computations. Last but not least, given the combinatorial nature, I would like to find out how far we can abstract models without serious performance degradation, and how expressive those models can be in order to facilitate the implementation of graph mining algorithms.

• Multi-Resource Characterization in Parallel Programs: Parallel processing frameworks perform computation by statically requesting multi-dimensional resource containers from cluster schedulers (e.g., YARN and Mesos). Community experiences have shown that tuning these dimensions appropriately taking into account each application individually can have great impact in performance.

  • What would be the right set of application profiles that captures most of behaviors regarding multi-resource tuning in parallel applications?
  • How to use that knowledge to automate the process of resource requesting? Ideally I would not want these aspects to be transparent to the user.

• Autonomic Computing: As computational demands can greatly vary in a parallel execution due to faults, (over,under)subscription of resources, heterogeneous environments, and so on, modern distributed systems would leverage autonomic computing characteristics and principles.

  • How far are we from actually being capable of incorporating such concepts in large-scale distributed systems?
  • How feasible is to design and implement systems with autonomic computation as a requirement, rather than a feature?

Selected Publications

• Diagnosing Performance Bottlenecks in Massive Data Parallel Programs

  Vinicius Dias, Rubens Moreira, Wagner Meira Jr., Dorgival Guedes

  CCGrid 2016

  pdf slides code doi

• Dynamic Reconfiguration of Data Parallel Programs

  Vinicius Dias, Wagner Meira Jr., Dorgival Guedes

  SBAC-PAD 2016

  pdf slides code doi

More

My CV My CV on Lattes Platform

• (in portuguese) Processamento de Dados Massivos (Big-Data) com Spark

  Vinicius Dias, Dorgival Guedes

  EVCOMP-2016 Escola de Verão em Computação do PPGCC/DCC

  slides video event

• viniciusvdias@dcc.ufmg.br
• Speed Lab, Universidade Federal de Minas Gerais, Belo Horizonte, BRA.