Funding entity: MINECO (TIN2015-65277-R)
Principal Investigators: Manuel Prieto Matías / Luis Piñuel Moreno
Participating entities: Complutense University of Madrid
Starting date: 01/01/2016 Ending date: 31/12/2019
Project amount: 307,000 €
SUMMARY
Over the last few years, heterogeneity has become one of the main concepts explored by computer architects to improve performance while keeping power, energy and thermal efficiency in the face of continued technology miniaturization (as predicted by Moore’s Law) and slowed supply voltage reduction (i.e. the end of Dennard scaling). Heterogeneity comes in many flavours ranging from heterogeneous configurations in datacenters to processors with on-chip hardware accelerators, hybrid CPU/GPU platforms, single-ISA asymmetric multicores and multi-ISA heterogeneous multicores.
While the performance and power/energy opportunities that those processors and platforms offer have been extensively explored and analysed, important challenges are yet to be studied. In this project we address some of these issues regarding:
-
- Enhancing the Hardware/software interface on novel multicore architectures, including asymmetric multicore processors with different core types such as the ARM big.LITTLE architecture and homogeneous multicore architectures that include advanced policies for asymmetric resource allocation.
- Code generation and optimization for asymmetric and heterogeneous architectures, including methodologies and tools (runtimes) for designing and implementing applications and (auto) tuned libraries.
- Performance, energy and temperature monitoring, characterization and modelling, with a focus on (1) novel HPC architectures and accelerators and (2) datacenter-class systems.
- Thermal management in heterogeneous datacenter configurations with a focus on advanced cooling techniques.
- Workload allocation for the heterogeneous distributed platforms behind the mobile cloud computing paradigm.
On the other hand, there is also a broad consensus among architects and developers that the effective use of those complex platforms requires the development of holistic approaches and new methodologies for system design. Based on this hypothesis our goals also include the design and implementation of the following two case studies that address complementary challenges for our society:
-
- The development of non-invasive ambulatory monitorization mechanisms for biometric and bioelectric variables, with a focus on (1) performance and energy-related issues and (2) robust modeling and prediction in the context of neurological disorders.
- The development of a set of auto-tuned libraries to serve as a reference for developers who want to accelerate media encoders using the new HVEC standard. The main contributions in this area will come from (1) heterogeneous-aware implementations (2) the design of new arithmetic optimizations.