A researcher from our Computer Science department, along with international collaborators, has developed a new hybrid technique to enhance performance and energy efficiency in approximate real-time computing systems.
Researchers from Durham University and a number of international collaborators (University of Essex (UK), IISER Bhopal (India), NTNU (Norway)) have developed a new hybrid technique, PRECIOUS, which enhances the performance and energy efficiency of approximate real-time computing systems. This innovation addresses the conflicting demands of time-critical applications by optimising how tasks are scheduled and managed on heterogeneous computing platforms.
In approximate computing, an on-time, slightly less precise result is preferred over a late but perfectly accurate one. This approach is common in applications like mobile target tracking and multimedia processing. The new research, outlined in a paper titled "PRECIOUS: Approximate Real-Time Computing in MLC-MRAM based Heterogeneous CMPs" , tackles the challenge of maximising the Quality of Service (QoS) of these tasks while staying within power limits and meeting deadlines.
PRECIOUS employs a hybrid offline-online method to achieve its goals.
A key element of PRECIOUS is its use of MLC-MRAM, a type of non-volatile memory that offers double the storage density of traditional MRAM. This allows for a larger, more efficient last-level cache (LLC). The researchers developed a new block management technique, called as MLC_In_S_Out_HS, to mitigate the high write latency and energy consumption associated with hard bits in MLC-MRAM. This technique prioritises placing data from energy-efficient in-order (InO) cores in the faster "soft-bit" region of the cache, while also directing some requests from high-performance out-of-order (OoO) cores to this region. This helps maintain performance while saving energy.
This research represents the first work to maximise QoS for a dependent approximate time-critical task set on a heterogeneous multicore processor by using a novel scheduling strategy and a MLC-MRAM-based cache. The hybrid approach provides a robust solution for a wide range of real-time applications where power consumption and time criticality are contradictory constraints.
Further research is planned to extend PRECIOUS to support concurrent task execution within clusters and to incorporate dynamic voltage and frequency scaling (DVFS) and task migration strategies to further improve performance and energy efficiency.