Extending the Amalthea model to introduce hardware heterogeneity
Posted: Fri Jun 01, 2018
Title: Extending the Amalthea model to introduce hardware heterogeneity
Authors: Paolo Burgio and Marko Bertogna
Abstract:
In the last years, power-efficient embedded platforms based on heterogeneous multi- and many-core accelerators paved their way in the automotive domain, where size, weight and power (SWaP) are today primary concerns.
Unfortunately, the hardware complexity of such architectures makes "traditional" approaches to timing analysis and real-time scheduling almost unapplicable in the domain.
One of the main reasons is that, these methodologies and tools were originally designed for systems made of few cores, with a simple shared memory sistem, and they lose their effectiveness when applied, e.g., to the complex hierarchical memory system of modern many-cores.
To tackle this problem, in this work we extend a well-known system and execution model, Amalthea, to cope with hardware heterogeneity.
We choose Amalthea because it is widely adopted in the industrial automotive domain, and it is compliant with Autosar.
More specifically, we introduce in the Amalthea specifications the concept of a many-core co-processor/accelerators, which can be either reprogrammable such as a GPU or application specific, such as a NN accelerator, capturing key aspects of the system, such as timing behavior, memory footprint and data transfers between the host and the accelerator.
Our model extensions will support researchers and engineers in designing future real-time automotive systems based on power-efficient embedded architectures.
Authors: Paolo Burgio and Marko Bertogna
Abstract:
In the last years, power-efficient embedded platforms based on heterogeneous multi- and many-core accelerators paved their way in the automotive domain, where size, weight and power (SWaP) are today primary concerns.
Unfortunately, the hardware complexity of such architectures makes "traditional" approaches to timing analysis and real-time scheduling almost unapplicable in the domain.
One of the main reasons is that, these methodologies and tools were originally designed for systems made of few cores, with a simple shared memory sistem, and they lose their effectiveness when applied, e.g., to the complex hierarchical memory system of modern many-cores.
To tackle this problem, in this work we extend a well-known system and execution model, Amalthea, to cope with hardware heterogeneity.
We choose Amalthea because it is widely adopted in the industrial automotive domain, and it is compliant with Autosar.
More specifically, we introduce in the Amalthea specifications the concept of a many-core co-processor/accelerators, which can be either reprogrammable such as a GPU or application specific, such as a NN accelerator, capturing key aspects of the system, such as timing behavior, memory footprint and data transfers between the host and the accelerator.
Our model extensions will support researchers and engineers in designing future real-time automotive systems based on power-efficient embedded architectures.