|
|
 |
|
|
|
|
|
Graphitti: A Framework for Self-adapting High Performance Massively Parallel Graph Processing
|
|
|
The summary of my proposal:
Fast and thorough data analysis is thought to be the key to the next paradigm of scientific discovery.
Graph analytics, the class of data analysis that deals with data collected from networked environments,
is emerging as a huge consumer of computational resources due to its complex data-hungry algorithms. Social networking,
personal medicine, text and graphics content analysis, and search engines, are only a
few examples where Terra-, Peta- or even Exa-scale graph processing is required.
Currently, graph processing environments and modern architectures are diverging: the
performance of the hardware is rapidly increasing by means of multi-layer heterogeneous
parallelism, while traditional graph processing workloads use a coarse-grain single-layer
parallelism model. This poor match leads to very inefficient use of modern high-
performance computing (HPC) architectures (only 10-15% of their peak performance),
which reflects in the low application performance and/or high infrastructure cost.
In this project, I will investigate solutions to efficiently use modern HPC architectures,
massively-parallel and heterogeneous, for large-scale graph processing. To do so, I will
propose (1) a novel methodology for modeling generic HPC architectures (massively
parallel and heterogeneous), (2) a hardware-agnostic methodology for modeling large-
scale graph processing workloads, (3) a method for evaluating the hardware-software fit
at the model level, and (4) validation against existing graph-processing benchmarks.
Further, I will design and implement the Graphitti framework, a system for optimally
matching the hardware and software models together, within given performance
requirements and resource constraints. Essentially, Graphitti is a hardware-software co-
design framework able to generate self-adapted optimal implementations for large-scale
graph processing workloads.
My Graphitti framework will have a strong impact for owners of graph data collections
(Google, Twitter, Facebook, graph analytics start-up companies), software and hardware
vendors (IBM, NVIDIA, AMD, ARM), and infrastructure-renting companies (cloud
computing providers like Microsoft, Amazon, or IBM).
|
|
|
|
|
| Letters of support
|
|
|
| Graphitti, the framework I am proposing, will have signifcant impact
on various branches of industry and academia. For a brief overview of this impact, take a look here .
Further, there are three industrial partners that have expressed their interest in this proposal. Their letters of support are listed bellow.
|
|
|
|
|
|
Support from AMD , USA - Dr. Lee Howes:
here .
|
|
|
Support from Vector Fabrics, The Netherlands - Dr. Paul Stravers:
here .
|
|
|
Support from Stream Computing , The Netherlands - CEO Vincent Hindriksen, MSc :
here .
|
| The VENI Proposal |
|
|
|
|
|
During the evaluation process, the proposal remains confidential. However, if you would like to see its technical content or discuss
it with me, please do not hesitate to contact me: A.L.Varbanescu@tudelft.nl.
|
|
|
| Curriculum Vitae |
|
|
|
|
|
NWO requires a detailed CV of the application. Here is a copy of my CV in the NWO format.
|
|
|
|
|
|
|
|
|
|
VENI Proposal 2012 
