Shared Virtual Memory for Heterogeneous Embedded Systems on Chips

The most difficult obstacles for exploiting the full potential of modern heterogeneous embedded systems on chip (HESoCs) are posed by the partitioned memory models between host processor and accelerators. Sharing data requires the programmer to manually orchestrate data copies between virtually and physically addressed main memory sections, including the translation of virtual address pointers. This limits performance and is completely prohibitive for the heterogeneous implementation of applications operating on complex, pointer-rich data structures.
This thesis investigates the design of transparent, zero-copy shared virtual memory (SVM) frameworks for HESoCs that al- low to simply pass pointers between host and accelerators, and thereby improve both programmability and performance. Compared to full-fledged hardware solutions for SVM found in highperformance computing systems, the proposed mixed hardware-software designs are i) better suited for area- and power-constrained embedded systems, ii) less intrusive to the hardware architecture of both host and accelerators, and iii) allow for greater flexibility, while offering competitive or even superior performance.