Abstract
Following an invariant growth in the required computational performance
of processors, the multicore revolution started around 20 years ago. This
revolution was mainly an answer to power dissipation constraints restrict-
ing the increase of clock frequency in single-core processors. The multicore
revolution not only brought in the challenge of parallel programming, i.e.
being able to develop software exploiting the entire capabilities of many-
core architectures, but also the challenge of programming heterogeneous
platforms. The question of “on which processing element to map a specific
computational unit?”, is well known in the embedded community. With the
introduction of general-purpose graphics processing units (GPGPUs), dig-
ital signal processors (DSPs) along with many-core processors on different
system-on-chip platforms, heterogeneous parallel platforms are nowadays
widespread over several domains, from consumer devices to media process-
ing platforms for telecom operators. Finding mapping together with a suit-
able hardware architecture is a process called design-space exploration. This
process is very challenging in heterogeneous many-core architectures, which
promise to offer benefits in terms of energy efficiency. The main problem is
the exponential explosion of space exploration. With the recent trend of in-
creasing levels of heterogeneity in the chip, selecting the parameters to take
into account when mapping software to hardware is still an open research
topic in the embedded area. For example, the current Linux scheduler has
poor performance when mapping tasks to computing elements available in
hardware. The only metric considered is CPU workload, which as was shown
in recent work does not match true performance demands from the applica-
tions. Doing so may produce an incorrect allocation of resources, resulting
in a waste of energy. The origin of this research work comes from the ob-
servation that these approaches do not provide full support for the dynamic
behavior of stream processing applications, especially if these behaviors are
established only at runtime. This research will contribute to the general goal
of developing energy-efficient solutions to design streaming applications on
heterogeneous and parallel hardware platforms. Streaming applications are
nowadays widely spread in the software domain. Their distinctive characteristic is the retrieving of multiple streams of data and the need to process
them in real time. The proposed work will develop new approaches to ad-
dress the challenging problem of efficient runtime coordination of dynamic
applications, focusing on energy and performance management.
of processors, the multicore revolution started around 20 years ago. This
revolution was mainly an answer to power dissipation constraints restrict-
ing the increase of clock frequency in single-core processors. The multicore
revolution not only brought in the challenge of parallel programming, i.e.
being able to develop software exploiting the entire capabilities of many-
core architectures, but also the challenge of programming heterogeneous
platforms. The question of “on which processing element to map a specific
computational unit?”, is well known in the embedded community. With the
introduction of general-purpose graphics processing units (GPGPUs), dig-
ital signal processors (DSPs) along with many-core processors on different
system-on-chip platforms, heterogeneous parallel platforms are nowadays
widespread over several domains, from consumer devices to media process-
ing platforms for telecom operators. Finding mapping together with a suit-
able hardware architecture is a process called design-space exploration. This
process is very challenging in heterogeneous many-core architectures, which
promise to offer benefits in terms of energy efficiency. The main problem is
the exponential explosion of space exploration. With the recent trend of in-
creasing levels of heterogeneity in the chip, selecting the parameters to take
into account when mapping software to hardware is still an open research
topic in the embedded area. For example, the current Linux scheduler has
poor performance when mapping tasks to computing elements available in
hardware. The only metric considered is CPU workload, which as was shown
in recent work does not match true performance demands from the applica-
tions. Doing so may produce an incorrect allocation of resources, resulting
in a waste of energy. The origin of this research work comes from the ob-
servation that these approaches do not provide full support for the dynamic
behavior of stream processing applications, especially if these behaviors are
established only at runtime. This research will contribute to the general goal
of developing energy-efficient solutions to design streaming applications on
heterogeneous and parallel hardware platforms. Streaming applications are
nowadays widely spread in the software domain. Their distinctive characteristic is the retrieving of multiple streams of data and the need to process
them in real time. The proposed work will develop new approaches to ad-
dress the challenging problem of efficient runtime coordination of dynamic
applications, focusing on energy and performance management.
| Original language | English |
|---|---|
| Awarding Institution |
|
| Supervisors/Advisors |
|
| Award date | 4 Oct 2023 |
| Place of Publication | Åbo |
| Publisher | |
| Print ISBNs | 978-952-12-4305-9 |
| Electronic ISBNs | 978-952-12-4306-6 |
| Publication status | Published - 2023 |
| MoE publication type | G5 Doctoral dissertation (article) |