Anshuman Tripathi

BACK

Partition Oriented Affinity-Aware ERFair Scheduler (POAES).

Anshuman Tripathi, Dr. Arnab Sarkar, Prof. P.P Chakrabarti

Embedded Systems Laboratory, Department of Computer Science and Engineering, IIT Kharagpur

GENERAL : In this project we developed a proportionally fair scheduler for hard real time tasks, in multiprocessor environments, which incurrs very low task migration overheads, for a wide range of system load. We conducted simulations to study the overall effect of this saving in terms of task migrations to the scheduling overhead. It turns out, that a partion oriented scheduler which is aware of the processor and task affinities, incurs less scheduling overhead is most cases of loosely and tightly coupled systems. We compared the algorithm proposed (POAES) to three other ERFair algorithms, one being naive ERFair, and other two were POES and StickyERFair (both of wich aim towards reduction in task migrations).

END RESULT : It turns out that POAES beats the performance of the other four algorithms studied in this project, by a very clear margin, in case of both loosley coupled systems and tightly coupled systems. The simulation results show promissing improvments over the other algorithms, and provides an overall picture of the scheduling time computation (with was missing in the older works).

STATUS : The work been submitted to IEEE Transactions on Computers as a short paper.

FUTUTRE WORK:
  • We studied the efficiency of the algorithms on only generated datasets, implementing the algorithm on a embedded device and testing its performance in practice can give much more insigthful data
  • Since the scheduling overhead is lower for the POAES algorithm, the saving in scheduling time, can be used as a source of slack generation for tasks, which might be useful in procrastination scheduling
RESOURCES: RESULTS:
  • The number of task migrations incurred by POAES are least (almost negligible) as compared to ERFair, Sticky-ERFair and POES
  • Based on simulation studies, POAES outperforms POES in most scenerios, performing slightly poor only in case of low number of processors and tightly coupled cache system
  • Based on a typical time quanta of 1 ms, scheduler overhead for POAES is only around 1%, compared to POES(6%), Sticky-ERFair(14%) and naive ERFair(31%)