Throughput processors such as GPUs continue to provide higher peak arithmetic capability. Designing a high throughput memory system to keep the computational units busy is very challenging. Future throughput processors must continue to exploit data locality and utilize the on-chip and off-chip resources in the memory system more effectively to further improve the memory system throughput. This dissertation advocates orchestrating the thread scheduler with the cache management algorithms to alleviate GPU cache thrashing and pollution, avoid bandwidth saturation and maximize GPU memory system throughput. Based on this principle, this thesis work proposes three mechanisms to improve the cache efficiency and the memory throughput. This thesis work enhances the thread throttling mechanism with the Priority-based Cache Allocation mechanism (PCAL). By estimating the cache miss ratio with a variable number of cache-feeding threads and monitoring the usage of key memory system resources, PCAL determines the number of threads to share the cache and the minimum number of threads bypassing the cache that saturate memory system resources. This approach reduces the cache thrashing problem and effectively employs chip resources that would otherwise go unused by a pure thread throttling approach. We observe 67% improvement over the original as-is benchmarks and a 18% improvement over a better-tuned warp-throttling baseline. This work proposes the AgeLRU and Dynamic-AgeLRU mechanisms to address the inter-thread cache thrashing problem. AgeLRU prioritizes cache blocks based on the scheduling priority of their fetching warp at replacement. Dynamic-AgeLRU selects the AgeLRU algorithm and the LRU algorithm adaptively to avoid degrading the performance of non-thrashing applications. There are three variants of the AgeLRU algorithm: (1) replacement-only, (2) bypassing, and (3) bypassing with traffic optimization. Compared to the LRU algorithm, the above mentioned three variants of the AgeLRU algorithm enable increases in performance of 4%, 8% and 28% respectively across a set of cache-sensitive benchmarks. This thesis work develops the Reuse-Prediction-based cache Replacement scheme (RPR) for the GPU L1 data cache to address the intra-thread cache pollution problem. By combining the GPU thread scheduling priority together with the fetching Program Counter (PC) to generate a signature as the index of the prediction table, RPR identifies and prioritizes the near-reuse blocks and high-reuse blocks to maximize the cache efficiency. Compared to the AgeLRU algorithm, the experimental results show that the RPR algorithm results in a throughput improvement of 5% on average for regular applications, and a speedup of 3.2% on average across a set of cache-sensitive benchmarks. The techniques proposed in this dissertation are able to alleviate the cache thrashing, cache pollution and resource saturation problems effectively. We believe when these techniques are combined, they will synergistically further improve GPU cache efficiency and the overall memory system throughput. / text

The continued growth of the computational capability of throughput processors has made throughput processors the platform of choice for a wide variety of high performance computing applications. Graphics Processing Units (GPUs) are a prime example of throughput processors that can deliver high performance for applications ranging from typical graphics applications to general-purpose data parallel (GPGPU) applications. However, this success has been accompa- nied by new performance bottlenecks throughout the memory hierarchy of GPU-based systems. This dissertation identifies and eliminates performance bottlenecks caused by major sources of interference throughout the memory hierarchy. Specifically, we provide an in-depth analysis of inter- and intra-application as well as inter- address-space interference that significantly degrade the performance and efficiency of GPU-based systems. To minimize such interference, we introduce changes to the memory hierarchy for systems with GPUs that allow the memory hierarchy to be aware of both CPU and GPU applications’ charac- teristics. We introduce mechanisms to dynamically analyze different applications’ characteristics and propose four major changes throughout the memory hierarchy. First, we introduce Memory Divergence Correction (MeDiC), a cache management mecha- nism that mitigates intra-application interference in GPGPU applications by allowing the shared L2 cache and the memory controller to be aware of the GPU’s warp-level memory divergence characteristics. MeDiC uses this warp-level memory divergence information to give more cache space and more memory bandwidth to warps that benefit most from utilizing such resources. Our evaluations show that MeDiC significantly outperforms multiple state-of-the-art caching policies proposed for GPUs. Second, we introduce the Staged Memory Scheduler (SMS), an application-aware CPU-GPU memory request scheduler that mitigates inter-application interference in heterogeneous CPU-GPU systems. SMS creates a fundamentally new approach to memory controller design that decouples the memory controller into three significantly simpler structures, each of which has a separate task, These structures operate together to greatly improve both system performance and fairness. Our three-stage memory controller first groups requests based on row-buffer locality. This grouping allows the second stage to focus on inter-application scheduling decisions. These two stages en- force high-level policies regarding performance and fairness. As a result, the last stage is simple logic that deals only with the low-level DRAM commands and timing. SMS is also configurable: it allows the system software to trade off between the quality of service provided to the CPU versus GPU applications. Our evaluations show that SMS not only reduces inter-application interference caused by the GPU, thereby improving heterogeneous system performance, but also provides better scalability and power efficiency compared to multiple state-of-the-art memory schedulers. Third, we redesign the GPU memory management unit to efficiently handle new problems caused by the massive address translation parallelism present in GPU computation units in multi- GPU-application environments. Running multiple GPGPU applications concurrently induces significant inter-core thrashing on the shared address translation/protection units; e.g., the shared Translation Lookaside Buffer (TLB), a new phenomenon that we call inter-address-space interference. To reduce this interference, we introduce Multi Address Space Concurrent Kernels (MASK). MASK introduces TLB-awareness throughout the GPU memory hierarchy and introduces TLBand cache-bypassing techniques to increase the effectiveness of a shared TLB. Finally, we introduce Mosaic, a hardware-software cooperative technique that further increases the effectiveness of TLB by modifying the memory allocation policy in the system software. Mosaic introduces a high-throughput method to support large pages in multi-GPU-application environments. The key idea is to ensure memory allocation preserve address space contiguity to allow pages to be coalesced without any data movements. Our evaluations show that the MASK-Mosaic combination provides a simple mechanism that eliminates the performance overhead of address translation in GPUs without significant changes to GPU hardware, thereby greatly improving GPU system performance. The key conclusion of this dissertation is that a combination of GPU-aware cache and memory management techniques can effectively mitigate the memory interference on current and future GPU-based systems as well as other types of throughput processors.

Recent graphics processing units (GPUs) have emerged as a promising platform for general purpose computing and have been shown to be very efficient in executing parallel applications with regular control and memory access behavior. Current GPU architectures primarily adopt the single-instruction multiple-thread (SIMT) programming model that balances programmability and hardware efficiency. With SIMT, the programmer writes application code to be executed by scalar threads and each thread is supported with conditional branch and fine-grained load/store instruction for ease of programming. At the same time, the hardware and software collaboratively enable the grouping of scalar threads to be executed in a vectorized single-instruction multiple-data (SIMD) in-order pipeline, simplifying hardware design. As GPUs gain momentum in being utilized in various application domains, these throughput processors will increasingly demand more efficient execution of irregular applications. Current GPUs, however, suffer from reduced thread-level parallelism, underutilization of compute resources, inefficient on-chip caching, and waste in off-chip memory bandwidth utilization for highly irregular programs with divergent control and memory accesses. In this dissertation, I develop techniques that enable simple, robust, and highly effective performance optimizations for SIMT-based throughput processor architectures such that they can better manage irregularity. I first identify that previously suggested optimizations to the divergent control flow problem suffers from the following limitations: 1) serialized execution of diverging paths, 2) lack of robustness across regular/irregular codes, and 3) limited applicability. Based on such observations, I propose and evaluate three novel mechanisms that resolve the aforementioned issues, providing significant performance improvements while minimizing implementation overhead. In the second half of the dissertation, I observe that conventional coarse-grained memory hierarchy designs do not take into account the massively multi-threaded nature of GPUs, which leads to substantial waste in off-chip memory bandwidth utilization. I design and evaluate a locality-aware memory hierarchy for throughput processors, which retains the advantages of coarse-grained accesses for spatially and temporally local programs while permitting selective fine-grained access to memory. By adaptively adjusting the access granularity, memory bandwidth and energy consumption are reduced for data with low spatial/temporal locality without wasting control overheads or prefetching potential for data with high spatial locality. / text