Back to search

National Technical University of Athens (NTUA)

Ανάπτυξη μεθοδολογίας για δυναμική ανάθεση μνήμης σε πολυπύρηνα επεξεργαστικά συστήματα

Abstract

dc:description

This dissertation presents a methodology for creating dynamic memory allocators on computing systems, ranging from embedded systems to high-workload servers. Our thesis is that modern implementations of dynamic memory allocators have become very generic and do not use sufficiently features and functions of the modern computing systems where they are being executed. We propose a framework, that we call dmmlib, through which application developers can create their own library for dynamic memory management. By selecting different memory organizations, policies and mechanisms, the produced library can be used directly by the applications with recompiling them being optional. The dmmlib framework offers also the possibility to re-configure the allocator at the runtime in order to improve further the application performance, as well as the ability to profile extensively the memory usage.In this dissertation, we show that dmmlib can be used efficiently in both embedded systems with limited software ecosystems and general-purposed, high-performance systems, while it could act as a study and development field of policies and mechanisms for dynamic memory management. We begin by evaluating existing approaches in dynamic memory management, including popular allocators that are used in real-world applications and operating systems, and we show the advantages and possible shortcomings in modern applications. Next, we present dmmlib, and we motivate the implementations of policies and the broader mechanisms. We also derive of a possible way of executing and profiling applications against dmmlib and how the firsts can re-configure the latter at their runtime.We present also two examples of dmmlib usage on top of three different platforms. The first two refer to embedded systems, while the last one to general-purposed systems. The very first uses P2012, a many-core accelerator for embedded systems. Dynamic memory management in this platform requires extra treatment, since a single processing unit is responsible for the resource management of a cluster with 16 processing elements. We port and evaluate dmmlib at this platform. Our results show that the allocators produced by dmmlib can be up to 2.6 times faster than the official allocator of the platform, without compromising additional memory space for the metadata. In the next embedded platform, we show how processor scalability affects the performance of dynamic memory allocators. Experimental results show that low-level memory managers may not cope with scalability if the processors exceed a certain number. Finally, our last target platform is more generic and our work aims on predicting the sizes of future requests and how the current allocator can use such information in order to re-configure and improve himself at runtime. Our approach introduces minimal overhead, while it manages to keep memory fragmentation low.

Degree

thesis:*
Grantor dc:publisher
National Technical University of Athens (NTUA)
Year dc:date
2016

Author and committee

dc:creator, dc:contributor.*
Authors dc:creator
  • Koutras, Ioannis
  • Κούτρας, Ιωάννης

Subjects

dc:subject × 16

Rights

dc:rights
Statement dc:rights
  • BY
Language dc:language
gre

Identifiers

dc:identifier.*
Identifier
10.12681/eadd/39120

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

Koutras, Ioannis; Κούτρας, Ιωάννης. Ανάπτυξη μεθοδολογίας για δυναμική ανάθεση μνήμης σε πολυπύρηνα επεξεργαστικά συστήματα. National Technical University of Athens (NTUA), 2016. http://hdl.handle.net/10442/hedi/39120