Predictable and composable system-on-chip memory controllers

Contemporary System-on-Chip (SoC) become more and more complex, as increasing integration results in a larger number of concurrently executing applications. These applications consist of tasks that are mapped on heterogeneous multi-processor platforms with distributed memory hierarchies, where SRAMs and SDRAMs are shared by a variety of arbiters. Some applications have real-time requirements, meaning that they must perform a particular computation before a deadline to guarantee functional correctness, or to prevent quality degradation. Mapping the applications on the platform such that all real-time requirements are satisfied is very challenging. The number of possible mappings of tasks to processing elements and data structures to memories may be large, and appropriate configuration settings must be determined once the mapping is chosen. Verifying that a particular mapping satisfies all application requirements is typically done by system-level simulation. However, resource sharing causes interference between applications, making their temporal behaviors inter-dependent. All concurrently executing applications must hence be verified together, causing the verification complexity of the system to increase exponentially with the number of applications. Together these factors contribute to making the integration and verification process a dominant part of SoC development, both in terms of time and money. Predictable and composable systems are proposed to manage the increasing verification complexity. Predictable systems provide lower bounds on application performance, while applications in composable systems are completely isolated and cannot affect each other’s temporal behavior by even a single clock cycle. Predictable systems enable formal verification that covers all possible interactions with the platform. However, this assumes that the behavior of an application is captured in a performance model, which is not the case for many applications. Composability offers a complementary verification approach by letting these applications be verified independently by simulation with linear verification complexity. A limitation of current predictable and composable systems is that there are no memory controllers supporting the concepts in a general way. Current SRAM controllers can be shared in a predictable way with a variety of arbiters, but are only composable if statically scheduled or shared using time-division multiplexing. Existing SDRAM controllers are not composable, and are either unpredictable or limited to applications that are statically scheduled. This thesis addresses the limitations of current predictable and composable systems by proposing a general predictable and composable memory controller, thereby addressing the mapping and verification problem in embedded systems. The proposed memory controller is divided into a front-end and a back-end. The back-end is specific for DDR2/DDR3 SDRAM and makes the memory behave in a predictable manner using precomputed memory patterns that are dynamically combined at run time. The front-end contains buffering and an arbiter in the class of Latency-Rate (LR) servers, which is a class with many well-known predictable arbiters. We extend this class with a Credit-Controlled Static-Priority (CCSP) arbiter that is developed specifically for shared resources with latency-critical requestors and high loads, such as memories. Three key features of CCSP are: 1) It accommodates latency-critical requestors with low bandwidth requirements without wasting bandwidth. 2) Over-allocated bandwidth can be made negligible at an increased area cost, without affecting latency. 3) It has a small implementation that runs fast enough to keep up with most DDR2/DDR3 memories. The proposed front-end is general and can be used with other predictable resources, such as SRAM controllers. The proposed memory controller hence supports multiple arbiter and memory types, thus addressing the diversity in modern SoCs. The combination of front-end and predictable memory behaves like a LR server, which is the shared resource abstraction used in this work. In essence, a LR server guarantees a requestor a minimum bandwidth and a maximum latency, enabling formal verification of real-time requirements. The LR server model is compatible with several commonly used formal analysis frameworks, such as network calculus and data-flow analysis. Our memory controller hence allows any combination of predictable memory and LR arbiter to be used transparently for formal verification of applications with any of these frameworks. Sharing a predictable memory at run-time results in interference between requestors, making the memory controller non-composable. This is addressed by adding a Delay Block to the front-end that delays all signals sent from the front-end to a requestor to always emulate worst-case interference. This makes requestors unable to affect each other’s temporal behavior, which is sufficient to guarantee composability on the level of applications. Our predictable memory controller hence offers composable service with a variety of memory and arbiter types, which widely extends the scope of composable platforms. Another benefit of this approach is that it enables composable service to be dynamically enabled and disabled, enabling requestors that do not require composable service to use slack bandwidth to improve performance. The predictable and composable memory controller is supported by a configuration flow that automatically computes memory patterns and arbiter settings to satisfy given bandwidth and latency requirements. The flow uses abstraction to separate the configuration of the memory and the arbiter, enabling settings to be computed in a streamlined fashion for all supported memories and arbiters