VIRTUAL MEMORY ON A MANY-CORE NOC

Many-core devices are likely to become increasingly common in real-time and embedded systems as computational demands grow and as expectations for higher performance can generally only be met by by increasing core numbers rather than relying on higher clock speeds. Network-on-chip devices, where multiple cores share a single slice of silicon and employ packetised communications, are a widely-deployed many-core option for system designers. As NoCs are expected to run larger and more complex programs, the small amount of fast, on-chip memory available to each core is unlikely to be sufficient for all but the simplest of tasks, and it is necessary to find an efficient, effective, and time-bounded, means of accessing resources stored in off-chip memory, such as DRAM or Flash storage. The abstraction of paged virtual memory is a familiar technique to manage similar tasks in general computing but has often been shunned by real-time developers because of concern about time predictability. We show it can be a poor choice for a many-core NoC system as, unmodified, it typically uses page sizes optimised for interaction with spinning disks and not solid state media, and transports significant volumes of subsequently unused data across already congested links. In this work we outline and simulate an efficient partial paging algorithm where only those memory resources that are locally accessed are transported between global and local storage. We further show that smaller page sizes add to efficiency. We examine the factors that lead to timing delays in such systems, and show we can predict worst case execution times at even safety-critical thresholds by using statistical methods from extreme value theory. We also show these results are applicable to systems with a variety of connections to memory.