Back to search

National Technical University of Athens (NTUA)

On application design for manycore processing systems in the domain of neuroscience

Abstract

dc:description

In recent years, the rapidly growing field of human neuromodelling has undergone significant changes. Neuroscientists have been taking impressive steps in unveiling the elaborate functionality of the human brain. In doing so, complex mathematical models have been the focus of efforts for describing detailed electrochemical processes that govern the human brain's behaviour. Such efforts require tremendous computational power in order to render, simulate and analyze in traditional computing systems. As such, the field of computational neuroscience presents an imposing challenge that the realm of high performance computing is tasked with meeting.The evolution of our understanding and mapping of the human brain has been accompanied by a steady increase in the processing power made available in manycore processors. Processors such as Intel's Xeon Phi line of products have grown to incorporate more advanced computing capability over the years. Due to their nature, they also provide traditional parallel coding tools, which can significantly impact the ease at which applications can be developed, tested and marketed. As a result, manycore CPUs are presented as an attractive alternative to other high-performance computing fabrics, such as GPUs and FPGAs.In this Doctoral thesis, we investigate the impact that manycore processors can have in the domain of computational neuroscience, specifically from the viewpoint of high-detail neuromodelling. By identifying a lack of research efforts in high-performance, large-scale, detailed neuronal simulations, the thesis presents the development of a simulator rich `in biophysical detail in manycore x86-based processors. Furthermore, the simulator acts as a means to study how manycore processors have evolved in architecture and behaviour, as well as highlight their strengths and drawbacks, in an effort to understand the role that they can play in the landscape of high-performance neuromodelling.This Doctoral thesis presents the complete development effort of the aforementioned simulator. The effort commences with an application specifically designed for the earliest, experimental manycore processors. We meticulously re-configure the simulator and its implementation design in order to take advantage of the continuously evolving architecture of manycore processors. Through this process, the simulator incorporates more modelling options, supports a wider range of simulation parameters and operates on a scalable, modern manycore system. The end product of this thesis is a simulator that constitutes an efficient solution for studying demanding neuronal models, in terms of both performance and energy. The thesis starts with a design that can simulate an average network of 50k neurons and 2 million synapses in 40 minutes for every second of simulated brain activity; the final design on a modern, small cluster of manycore processors vastly improves on this design by simulating 2 million neurons and 2 billion synapses in under 10 minutes for every second of simulated brain activity.Our point of focus and contributions lie in the analysis of manycore processor performance when tasked with demanding neuromodelling workloads. Through the proposed simulator, we highlight how the significant wealth of neuromodelling parameters affects simulation in different manycore processors. The thesis will provide a clear map on matching the correct amount, and type, of hardware to different network simulation configurations. As such, we take an important step towards defining proper utilization of high-performance hardware in order to match simulation challenges imposed by the domain of computational neuroscience.Furthermore, significant effort is expended in incorporating the simulator in a larger, collaborative framework aimed at serving as an online resource for high-performance neuromodelling simulations. The designed framework, named “BrainFrame”, leverages a heterogeneous ensemble of accelerators, namely manycore processors, FPGAs and GPUs, in order to provide efficient solutions for different modelling and network configurations. We provide a proof of value in the framework by identifying different use cases where a switch in the underlying accelerator hardware yields significant gains in performance, thus reinforcing the value of heterogeneity in high-performance neuromodelling. In particular, the framework further highlights the value of matching the accelerating hardware of choice with the computational workload at hand by indicating differences in performance by orders of magnitude between the different accelerators examined in different network configuration scenarios.

Degree

thesis:*
Grantor dc:publisher
National Technical University of Athens (NTUA)
Year dc:date
2019

Author and committee

dc:creator, dc:contributor.*
Authors dc:creator
  • Chatzikonstantis, Georgios
  • Χατζηκωνσταντής, Γεώργιος

Subjects

dc:subject × 22

Rights

Language dc:language
eng

Identifiers

dc:identifier.*
Identifier
10.12681/eadd/49353

Chain of custody

source
Harvested from
Greek National Archive of PhD Theses
Base URL
phdtheses.ekt.gr/eadd_oai/request
Last updated
2026-11-18
Source record
OAI-PMH GetRecord
citation

Chatzikonstantis, Georgios; Χατζηκωνσταντής, Γεώργιος. On application design for manycore processing systems in the domain of neuroscience. National Technical University of Athens (NTUA), 2019. http://hdl.handle.net/10442/hedi/49353