An efficient design space exploration framework to optimize power-efficient heterogeneous many-core multi-threading embedded processor architectures

By the middle of this decade, uniprocessor architecture performance had hit a roadblock due to a combination of factors, such as excessive power dissipation due to high operating frequencies, growing memory access latencies, diminishing returns on deeper instruction pipelines, and a saturation of available instruction level parallelism in applications. An attractive and viable alternative embraced by all the processor vendors was multi-core architectures where throughput is improved by using micro-architectural features such as multiple processor cores, interconnects and low latency shared caches integrated on a single chip. The individual cores are often simpler than uniprocessor counterparts, use hardware multi-threading to exploit thread-level parallelism and latency hiding and typically achieve better performance-power figures. The overwhelming success of the multi-core microprocessors in both high performance and embedded computing platforms motivated chip architects to dramatically scale the multi-core processors to many-cores which will include hundreds of cores on-chip to further improve throughput. With such complex large scale architectures however, several key design issues need to be addressed. First, a wide range of micro-architectural parameters such as L1 caches, load/store queues, shared cache structures and interconnection topologies and non-linear interactions between them define a vast non-linear multi-variate micro-architectural design space of many-core processors; the traditional method of using extensive in-loop simulation to explore the design space is simply not practical. Second, to accurately evaluate the performance (measured in terms of cycles per instruction (CPI)) of a candidate design, the contention at the shared cache must be accounted in addition to cycle-by-cycle behavior of the large number of cores which superlinearly increases the number of simulation cycles per iteration of the design exploration. Third, single thread performance does not scale linearly with number of hardware threads per core and number of cores due to memory wall effect. This means that at every step of the design process designers must ensure that single thread performance is not unacceptably slowed down while increasing overall throughput. While all these factors affect design decisions in both high performance and embedded many-core processors, the design of embedded processors required for complex embedded applications such as networking, smart power grids, battlefield decision-making, consumer electronics and biomedical devices to name a few, is fundamentally different from its high performance counterpart because of the need to consider (i) low power and (ii) real-time operations. This implies the design objective for embedded many-core processors cannot be to simply maximize performance, but improve it in such a way that overall power dissipation is minimized and all real-time constraints are met. This necessitates additional power estimation models right at the design stage to accurately measure the cost and reliability of all the candidate designs during the exploration phase.;In this dissertation, a statistical machine learning (SML) based design exploration framework is presented which employs an execution-driven cycle-accurate simulator to accurately measure power and performance of embedded many-core processors. The embedded many-core processor domain is Network Processors (NePs) used to processed network IP packets. Future generation NePs required to operate at terabits per second network speeds captures all the aspects of a complex embedded application consisting of shared data structures, large volume of compute-intensive and data-intensive real-time bound tasks and a high level of task (packet) level parallelism. Statistical machine learning (SML) is used to efficiently model performance and power of candidate designs in terms of wide ranges of micro-architectural parameters. The method inherently minimizes number of in-loop simulations in the exploration framework and also efficiently captures the non-linear interactions between the micro-architectural design parameters. To ensure scalability, the design space is partitioned into (i) core-level micro-architectural parameters to optimize single core architectures subject to the real-time constraints and (ii) shared memory level micro-architectural parameters to explore the shared interconnection network and shared cache memory architectures and achieves overall optimality. The cost function of our exploration algorithm is the total power dissipation which is minimized, subject to the constraints of real-time throughput (as determined from the terabit optical network router line-speed) required in IP packet processing embedded application.