Efficient Execution of Mixed Application Workloads in a Hard Real-Time Multicore System

1. Efficient Execution of Mixed Application Workloads in a Hard Real-Time Multicore System Marco Paolieri (BSC/UPC) Eduardo Quiñones (BSC) Francisco J. Cazorla (BSC) Mateo Valero (BSC/UPC) RePP Workshop Grenoble, 15th October

2. Future Real-Time Systems • Current real-time embedded systems require higher performance than provided by current processors • Increasing safety, comfort, number and quality of services

3. Architecture What about predictability?

4. Multicores in RTESs: disadvantages • It is harder to perform WCET analysis for multicore processors than for single-core because of Inter-thread Interferences • Inter-thread interferences accessing shared resources make the execution time vary • Execution time, and so the WCET of a HRT depend on the mixed application workload • Where: • WCETa WCET est without interferences • d deadline • ETa,b ET of a running with b • ETa,c ET of a running with c Is it possible to use multicores to execute mixed application workloads? WCETa d ETa,b ETa,c deadline miss

5. Our Goal • Execute efficiently mixed application workload providing • predictability to HRTs • maximizing performance of NHRTs with the resources not used by HRTs

6. Regarding HRTs • Our multicore architecture guarantees by design that the maximum delay a request accessing a shared resource may suffer due to inter-thread interferences has an Upper Bound Delay (UBD) • Inter-thread interferences < UBD • Round Robin provides UBD based on the number of requestors UBD = (NHRT – 1) * LBUS where LBUS is the latency of the bus and NHRT the total number of HRTs running at the same time • The impact of inter-thread interferences on WCET is up to 40% [Paolieri et al. Hardware Support for WCET Analysis of Multicore Hard Real-Time Systems, ISCA’09]

7. What about Non Real-Time Tasks? • HRTs are accessing shared resources as soon as they are available, what we call Average-Case Resource Managament (AC-RM) • NHRTs could be starving before accessing shared resources • To execute efficiently a mixed application workload • It is required to maximize the perfomance of NHRTs

8. Our Proposal • No advantages if executing the HRTs before their WCET • We evaluate a resource management policy, called Worst-Case Resource Management (WC-RM) • Every access to a shared resource from a HRT is stalled by UBD cycles • This forces HRTs to be executed closer to their WCET • It provides performance to NHRTs still guaranteeing that HRTs meet their deadlines • Reducing the variance between average case execution time and WCET performances of NHRTs can increase

9. How It Works 1.04 1.01 1.03 1.06 1.05 1.05 1.04 1.05 1.06 1.05 1.04

10. Experimental Setup • As HRT benchmarks we use: • A real HRT application provided by Honeywell • 3D collision avoidance algorithm • EEMBC Automotive • As NHRT benchmarks we use • MediaBench, MiBench, SPEC CPU 2006

11. Results 1 • Baseline: throughput of NHRTs using the AC-RM

12. Effect on HRTs 1 • Baseline: Performance of HRTs using AC-RM • WC-RM is between 1 and 2 % close to its WCET estimation

13. Conclusions • Our architecture is designed from a WCET point of view • It implements a hardware shared-resource management policy, called Worst-Case Resource Management (WC-RM) • It improves the performance level of NHRTs when running in a mixed application workload

14. Thanks for the attention!