Advanced Processor Technologies Home
APT Advanced Processor Technologies Research Group

Workload-Adaptation in Memory Controllers

Ghasempour, Mohsen


Advanced development in processor design, increasing the heterogeneity of computer system, by involving Graphics Processing Units (GPUs), Field-Programmable Gate Arrays (FPGAs) and custom accelerators, and increasing the number of cores and threads in such systems puts extra pressure on the main memory, demanding a higher performance. Current computing trends are putting ever more pressure on main memory. In modern computer systems, this is generally Dynamic Random Access Memory (DRAM) which consists of a multi-level access hierarchy (e.g. Rank, Bank, Row etc.). This heterogeneity of structure implies different access latencies (and power consumption), resulting in performance differences according to memory access patterns. DRAM controllers manage access and satisfy the timing constraints and now employ complex scheduling and prediction algorithms to mitigate the effect on performance. This complexity can limit the scalability of a controller with the size of memory, while maintaining performance. The focus of this PhD thesis is to improve performance, reliability and scalability (with respect to memory size) of DRAM controllers. To this end, it covers three significant contributors to the performance and reliability of a memory controller: 'Address Mapping', 'Page Closure Policies' and 'Reliability Monitoring'. A detailed DRAM simulator is used as an evaluation platform throughout this work. The following contributions are presented in this thesis.

Hybrid Address-based Page PolicY (HAPPY): Memory controllers have used static page-closure policies to decide whether a row should be left open (open-page policy) or closed immediately (close-page policy) after use. The appropriate choice can reduce the average memory latency. Since access patterns are dynamic, static page policies cannot guarantee to deliver optimum execution time. Hybrid page policies can cover dynamic scenarios and are now implemented in state-of-the-art processors. These switch between open-page and close-page policies by monitoring the accesspattern of row hits/conflicts and predicting future behaviour. Unfortunately, as the size of DRAM memory increases, fine-grain tracking and analysis of accesses does not remain practical. HAPPY proposes a compact, memory address-based encoding technique which can maintain or improve page closure predictor performance while reducing the hardware overhead. As a case study, HAPPY is integrated, with a state-of-the-art monitor -- the Intel-adaptive open-page policy predictor employed by the Intel Xeon X5650 -- and a traditional Hybrid page policy. The experimental results show that using the HAPPY encoding applied to the Intel-adaptive page closure policy can reduce the hardware overhead by 5x for the evaluated 64 GB memory (up to 40x for a 512 GB memory) while maintaining the prediction accuracy.

Dynamic Re-arrangement of Address Mapping (DReAM): The initial location of data in DRAMs is determined and controlled by the 'address-mapping' and even modern memory controllers use a fixed and runtime-agnostic address-mapping. On the other hand, the memory access pattern seen at the memory interface level will be dynamically changed at run-time. This dynamic nature of memory access pattern and the fixed behaviour of address mapping process in DRAM controllers, implied by using a fixed address-mapping scheme, means that DRAM performance cannot be exploited efficiently. DReAM is a novel hardware technique that can detect a workload-specific address mapping at run-time based on the application access pattern. The experimental results show that DReAM outperforms the best evaluated baseline address mapping by 5%, on average, and up to 28% across all the workloads.

A Run-time Memory hot-row detectOR (ARMOR): DRAM needs refreshing to avoid data loss. Data can also be corrupted within a refresh interval by crosstalk caused by repeated accesses to neighbouring rows; this is the row hammer effect and is perceived as a potentially serious reliability and security threat. ARMOR is a novel technique which improves memory reliability by detecting which rows are potentially being "hammered" within the memory controller, which can then insert extra refresh operations. It can detect (and thus prevent) row hammer errors with minimal execution time overhead and hardware requirements. Alternatively by adding buffers inside the memory controller to cache such hammered rows, execution times are reduced with small hardware costs. The ARMOR technique is now the basis of a patent applicant and under process for commercial exploitation.

As a final step of this PhD thesis, an adaptive memory controller was developed integrating HAPPY, DReAM and ARMOR into a standard memory controller. The performance and the implementation cost of such an adaptive memory controller were compared against a state-of-the-art memory controller, as a baseline. The experimental results show that the adaptive memory controller outperforms the baseline, on average by 18%, and up to 35% for some workloads, while requiring around 6 KB-900 KB more storage than the baseline to support a wide range of memory sizes (from 4 GB up to 512 GB).