Ecosyste.ms: Awesome

An open API service indexing awesome lists of open source software.

Awesome Lists | Featured Topics | Projects

https://github.com/erikzenker/diploma

Latex sources of my diploma thesis from december 2014 on the topic: "A flexible and portable approach for communication in distributed computing systems"
https://github.com/erikzenker/diploma

Last synced: 7 days ago
JSON representation

Latex sources of my diploma thesis from december 2014 on the topic: "A flexible and portable approach for communication in distributed computing systems"

Awesome Lists containing this project

README 

        
# A Flexible and Portable Approach for Communication in Distributed Computing Systems



* [Latex sources](thesis/diplom.tex) of my diploma thesis from december 2014

* [Precompiled pdf](thesis/diplom.pdf)



## Introduction

The domain of high performance computing is subject to constant develop-

ment.

The main focus of this development is the increase of computing

power [ref:performance_development]. More computing power enables high

performance applications to solve more complex problems or allows for an increase of

the problem size. Current state of the art computing systems reached the area of one

petaflop, that are 10 to the power of 15 floating point operations per second (petaflop/s), in the year

2008 [ref:ibm_roadrunner]. But only a small number of applications worldwide have

taken advantage of these so called petascale systems.

One application that has shown scalability on petascale systems, is PIConGPU

[ref:picongpu_scale]. It is a particle in cell physics simulation that was running

on the Titan supercomputer at the Oak Ridge National Laboratory in the United

States [ref:titan], reaching 7.1 petaflop/s peak performance by utilizing 18.000 Graphics

Processing Units (GPUs). However, applications capable of fully utilizing a petascale

system can often be expected to desire systems with even more computational power.

Therefore, current computing systems will no longer adequately cover the future needs

of complex, computation intensive scientific and industrial applications. To provide more

computing power to these applications, a rapid progress in the development of computers

and algorithms is necessary. The next milestone in the development of supercomputers

are exascale systems, capable of at least one exaflop/s, which represents a thousandfold

increase over the first petascale system.



On the one hand, not all questions about the construction of such a system are

solved. On the other hand, it is certain that exascale will increase the amount and

complexity of computers to a new level [ref:cresta], amplifying the challenge of reliability,

programmability and usability of such systems.



Considering an exascale system, it is not sufficient to simply port the applications

previously running on petascale systems. The sole, massive increase in size of the

computing system will lead to a decrease in the mean time to failure (MTTF) and

an increased importance of data locality. Furthermore, the emergence of accelerator

hardware leads to computing systems that are more heterogeneous and hierarchical, as

well as increasingly complex to program and use [ref:accel]. Therefore, applications

running on high performance computers have to be adapted to utilize the upcoming

generation of supercomputers. Current applications lack techniques that form the basis

for load balancing and fault tolerance. Thus, there is a need for a smart description

concept for high performance applications that models upcoming exascale computing

systems with respect to these techniques.



Many applications exploit the parallel power of clusters by interconnecting their parallel

entities via network. These applications can be enhanced by a smart communication

approach. This approach will enhance already existing communication libraries by a

1further layer, which is necessary for the upcoming supercomputers. This additional

layer is separated into three components: an abstract communication layer, a model

of the communication topology, and a mapping from the topology onto the abstract

communication layer. These mappings can be changed at any time during the application

execution. In connection with the exchangeable communication library, it forms a flexible

and portable communication approach for the upcoming supercomputer generation. A

prototype, that combines these three components in a single system, was implemented and

deployed in simulation applications. The message passing interface (MPI) was used within

this prototype as communication back-end underneath the communication abstraction

layer. It maps common communication processes onto equivalent MPI methods.

The following Chapter 2 first motivates the design of the system and subsequently

provides a discussion about of the system components. Selected implementation details

of the system is presented in Chapter 3. The developed prototype is evaluated and

compared in contrast to an MPI implementation in Chapter 4. Chapter 5 provides ideas

to enhance the system in future work. Finally, Chapter 6 summarizes the developed

system and Chapter 7 provides an outlook for future development.



#  Dependencies

* latex

* biber

* make



# Compile

```

cd thesis

make

```