METERG: Measurement-Based End-to-End Performance Estimation Technique in QoS-Capable Multiprocessors [abstract] (IEEE Xplore, PDF)
Jae W. Lee and Krste Asanovic
Proceedings of the 12th IEEE Real-Time and Embedded Technology and Applications Symposium (RTAS), April 2006.
Multiprocessor systems present serious challenges in the design of
real-time systems due to the wider variation of execution time of an
instruction sequence compared to a uniprocessor system. Even if
non-determinism is tightly controlled by adding conventional QoS
support, it is generally difficult to find the minimal hardware
resource request settings (e.g., memory bandwidth) for a given
user-level performance goal (e.g., transactions per second). In this
paper, we introduce the METERG (Measurement-Time Enforcement and
Run-Time Guarantee) QoS system that provides an easy method of
obtaining a tight estimate of the lower bound on end-to-end
performance for a given configuration of resource reservations. Every
QoS-capable block in the METERG system supports two operation modes
for each task requiring QoS: enforcement mode for estimating the lower
bound on a task's execution time and deployment mode for maximizing
its performance. We evaluate the effectiveness of our approach with an
execution-driven multiprocessor simulator implementing the METERG QoS
memory subsystem. We show that the performance lower bound is easy to
obtain by simply running an application in enforcement mode, and that
this estimated lower bound is tight.