This talk presents ATLAS-MP, a combination of runtime and kernel
scheduler, to process arbitrary collection of real-time jobs.
ATLAS-MP inherits the accessible API and execution time prediction
from the ATLAS infrastructure and extends it with primitives to
express intra-task parallelism. The scheduler is able to take full
advantage of SMP systems and balances load automatically over all
Multi-Processor Look-Ahead Scheduling
Verteidigung der Master-Arbeit
I compare benchmark results with the theoretical utilization bounds of
ATLAS-MP and discuss current limitations as well as possible
strategies to refine the design of ATLAS-MP.
Intel's Running Average Power Limit (RAPL) processor feature provides the
possibility to gather energy statistics about the processor. Various researchers
have already shown that these statistics can be used to calculate the energy
consumption of individual programs. However, their approaches are limited to
restricted systems which do not support processor sharing or which can just
utilize one processor. So far, nobody used RAPL for energy measurements in a
shared multicore system. In this talk, I will present an approach which tries to
achieve exactly this task. I will explain which obstacles need to be overcome to
solve the problem and present some first measurement results of my implementation.
S-RAPL - Hardware Based Energy Accounting for Multicore Systems
Zwischenpräsentation der Diplomarbeit, Sondertermin: 12:30 Uhr, APB E001