Master Degree / Yüksek Lisans Tezleri

Permanent URI for this collectionhttps://hdl.handle.net/11147/3008

Browse

Filters

2022 - 20234

Settings

Author: search.filters.author.Öz, Işıl

Search Results

Now showing 1 - 4 of 4
  • Master Thesis
    Community Detection on Gpu: a Comprehensive Analysis, Unified Memory Enhancement, and Memory Access Optimization
    (2023) Dinçer, Emre; Öz, Işıl
    Recent years have experienced a slowdown in the development of traditional systems that use only the Central Processing Unit (CPU). However, significant progress has been made in the development of heterogeneous systems utilizing not only the CPU but also the Graphics Processing Unit (GPU). NVIDIA, one of the GPU manufacturers, through its CUDA platform, has increased the interest of many researchers in heterogeneous systems by providing a means to program GPUs more easily. The ease of application development provided by the CUDA platform and the performance gains offered by these heterogeneous systems have encouraged many researchers to develop algorithms and applications that operate on these systems. One such algorithm that is frequently used in data analysis is the community detection algorithm. Although there are applications that implement this algorithm to run on GPUs, and while these applications work efficiently for many datasets, they either fail to work or experience significant performance loss for large datasets that exceed the GPU's memory capacity. In this thesis, we analyzed Rundemanen, which is one of the community detection applications running on GPU. We also made enhancements that enable Rundemanen to process datasets larger than the GPU's memory capacity by utilizing CUDA's Unified Memory. Lastly, we tested various optimization methods to use Unified Memory more efficiently. By using our memory-access advises, in comparison to the naive version, we obtained up to 62x and 8x performance gain with artificial oversubscription scenarios and for datasets that already do not fit into GPU memory, respectively.
  • Master Thesis
    Performance-Reliability Tradeoff Analysis for Safety-Critical Systems With Gpus
    (2023) Sezgin, Yağızcan; Öz, Işıl
    GPUs were mostly used for image processing purposes when they were first introduced. These applications can be considered non-critical, and they were not given sufficient importance for reliability. Due to the evolving nature of GPUs, they offer highly parallelized architecture and provide extremely powerful computation, they become one of the most crucial parts of the systems that have complex applications in safety-critical domains such as automotive and space to fulfill the high computational demand. In this thesis, we evaluate the performance and reliability tradeoff in the safety-critical domain. We propose software-based redundancy schemes with different spheres of replications on the GPU4S benchmark in the safety-critical domain. Our proposal includes profiling the baseline application without any redundancy, applying fault injection using NVBitFI and changing implementation manually according to proposed redundancy schemes, measuring performance metrics such as execution time, memory copy operations, and power consumption on the real hardware that is widely used on target domain instead of using well-known GPU simulators to see actual performance. We reveal that our proposed redundancy schemes are managed to eliminate all the soft errors in the cases if we apply full redundancy for single-kernel benchmarks, for the reliability evaluation with the cost of performance degradation, depending on the application. We show that most soft errors can be eliminated using partial redundancy for complex applications, with a small performance impact.
  • Master Thesis
    Evaluating Impacts of Micro-Architectural Metrics on Error Resilience and Performance of General Purpose Gpu Applications
    (01. Izmir Institute of Technology, 2023) Topçu, Burak; Öz, Işıl
    Rapidly growing data processing tasks require powerful and energy-efficient heterogeneous computing systems, and GPUs take on a significant mission for those systems in accelerating heavy workloads by executing multiple parallel tasks concurrently. Increasing architectural complexity and widening employment of GPUs bring error resiliency concerns for safety-critical applications. Furthermore, approaches that enhance performance and reduce energy dissipation handle error resiliency on GPUs through approximate computing solutions. Evaluating error resiliency in terms of either identifying error proneness of a system or investigating approximations without much disturbing the output necessities robust knowledge about the execution of a program on a device. In this thesis, we develop a runtime performance and power monitoring tool visualizing the execution with detailed micro-architectural metrics. By utilizing the tool, we acquire several fundamental understandings about runtime performance bottlenecks and how perturbations affect output quality. Afterward, we propose a framework predicting fault vulnerability for error-resilient GPU applications. The framework can accurately estimate error tolerance and saves from analyzing the fault occurrence probability requiring significant effort. Depending on the performance bottlenecks observed with the tool and the error propagation gained during prediction experiments, we introduce a hardware-based approximation computing approach targeting to improve the performance and power of GPU programs, especially memory-bound ones. The approximation method, which resolves memory utilization bottlenecks at runtime, enhances performance by 1.49× (up to 2.1×) and diminishes energy consumption by 28.4% (up to 52.6%) while maintaining the accuracy on the output above 98%.
  • Master Thesis
    Compiler-Managed Fault Tolerance Techniques for General Purpose Graphics Processing Units
    (Izmir Institute of Technology, 2022) Kaya, Ercüment; Öz, Işıl; Öz, Işıl
    As the use of graphics processing units evolves for general-purpose computations besides inherently-fault tolerant graphics programs, soft error reliability becomes a first-class citizen in program design. In this thesis, we aim to increase the reliability of general-purpose graphics processing units. We propose compiler-based redundancy schemes for graphics processing units. Our framework replicates the annotated kernel function by a programmer at compile time. Our selective redundancy approach enables us to provide full redundancy with no error and partial redundancy with an acceptable error rate with higher performance. We develop different schemes to satisfy the performance and memory requirements of the general-purpose graphics processing unit applications. We build our framework on top of the LLVM compiler framework to increase the reliability of applications that exploit the CUDA programming model and evaluate our schemes for the applications from the PolyBench benchmark suite. We reveal that our partial redundancy approach improves the reliability with a small performance overhead and our full redundancy schemes provide complete fault coverage with varying performance differences based on the application's characteristics.