Matching Items (18)
Filtering by
- All Subjects: Computer Science
- Member of: ASU Electronic Theses and Dissertations
Description
We are expecting hundreds of cores per chip in the near future. However, scaling the memory architecture in manycore architectures becomes a major challenge. Cache coherence provides a single image of memory at any time in execution to all the cores, yet coherent cache architectures are believed will not scale to hundreds and thousands of cores. In addition, caches and coherence logic already take 20-50% of the total power consumption of the processor and 30-60% of die area. Therefore, a more scalable architecture is needed for manycore architectures. Software Managed Manycore (SMM) architectures emerge as a solution. They have scalable memory design in which each core has direct access to only its local scratchpad memory, and any data transfers to/from other memories must be done explicitly in the application using Direct Memory Access (DMA) commands. Lack of automatic memory management in the hardware makes such architectures extremely power-efficient, but they also become difficult to program. If the code/data of the task mapped onto a core cannot fit in the local scratchpad memory, then DMA calls must be added to bring in the code/data before it is required, and it may need to be evicted after its use. However, doing this adds a lot of complexity to the programmer's job. Now programmers must worry about data management, on top of worrying about the functional correctness of the program - which is already quite complex. This dissertation presents a comprehensive compiler and runtime integration to automatically manage the code and data of each task in the limited local memory of the core. We firstly developed a Complete Circular Stack Management. It manages stack frames between the local memory and the main memory, and addresses the stack pointer problem as well. Though it works, we found we could further optimize the management for most cases. Thus a Smart Stack Data Management (SSDM) is provided. In this work, we formulate the stack data management problem and propose a greedy algorithm for the same. Later on, we propose a general cost estimation algorithm, based on which CMSM heuristic for code mapping problem is developed. Finally, heap data is dynamic in nature and therefore it is hard to manage it. We provide two schemes to manage unlimited amount of heap data in constant sized region in the local memory. In addition to those separate schemes for different kinds of data, we also provide a memory partition methodology.
ContributorsBai, Ke (Author) / Shrivastava, Aviral (Thesis advisor) / Chatha, Karamvir (Committee member) / Xue, Guoliang (Committee member) / Chakrabarti, Chaitali (Committee member) / Arizona State University (Publisher)
Created2014
Description
Audio signals, such as speech and ambient sounds convey rich information pertaining to a user’s activity, mood or intent. Enabling machines to understand this contextual information is necessary to bridge the gap in human-machine interaction. This is challenging due to its subjective nature, hence, requiring sophisticated techniques. This dissertation presents a set of computational methods, that generalize well across different conditions, for speech-based applications involving emotion recognition and keyword detection, and ambient sounds-based applications such as lifelogging.
The expression and perception of emotions varies across speakers and cultures, thus, determining features and classification methods that generalize well to different conditions is strongly desired. A latent topic models-based method is proposed to learn supra-segmental features from low-level acoustic descriptors. The derived features outperform state-of-the-art approaches over multiple databases. Cross-corpus studies are conducted to determine the ability of these features to generalize well across different databases. The proposed method is also applied to derive features from facial expressions; a multi-modal fusion overcomes the deficiencies of a speech only approach and further improves the recognition performance.
Besides affecting the acoustic properties of speech, emotions have a strong influence over speech articulation kinematics. A learning approach, which constrains a classifier trained over acoustic descriptors, to also model articulatory data is proposed here. This method requires articulatory information only during the training stage, thus overcoming the challenges inherent to large-scale data collection, while simultaneously exploiting the correlations between articulation kinematics and acoustic descriptors to improve the accuracy of emotion recognition systems.
Identifying context from ambient sounds in a lifelogging scenario requires feature extraction, segmentation and annotation techniques capable of efficiently handling long duration audio recordings; a complete framework for such applications is presented. The performance is evaluated on real world data and accompanied by a prototypical Android-based user interface.
The proposed methods are also assessed in terms of computation and implementation complexity. Software and field programmable gate array based implementations are considered for emotion recognition, while virtual platforms are used to model the complexities of lifelogging. The derived metrics are used to determine the feasibility of these methods for applications requiring real-time capabilities and low power consumption.
The expression and perception of emotions varies across speakers and cultures, thus, determining features and classification methods that generalize well to different conditions is strongly desired. A latent topic models-based method is proposed to learn supra-segmental features from low-level acoustic descriptors. The derived features outperform state-of-the-art approaches over multiple databases. Cross-corpus studies are conducted to determine the ability of these features to generalize well across different databases. The proposed method is also applied to derive features from facial expressions; a multi-modal fusion overcomes the deficiencies of a speech only approach and further improves the recognition performance.
Besides affecting the acoustic properties of speech, emotions have a strong influence over speech articulation kinematics. A learning approach, which constrains a classifier trained over acoustic descriptors, to also model articulatory data is proposed here. This method requires articulatory information only during the training stage, thus overcoming the challenges inherent to large-scale data collection, while simultaneously exploiting the correlations between articulation kinematics and acoustic descriptors to improve the accuracy of emotion recognition systems.
Identifying context from ambient sounds in a lifelogging scenario requires feature extraction, segmentation and annotation techniques capable of efficiently handling long duration audio recordings; a complete framework for such applications is presented. The performance is evaluated on real world data and accompanied by a prototypical Android-based user interface.
The proposed methods are also assessed in terms of computation and implementation complexity. Software and field programmable gate array based implementations are considered for emotion recognition, while virtual platforms are used to model the complexities of lifelogging. The derived metrics are used to determine the feasibility of these methods for applications requiring real-time capabilities and low power consumption.
ContributorsShah, Mohit (Author) / Spanias, Andreas (Thesis advisor) / Chakrabarti, Chaitali (Thesis advisor) / Berisha, Visar (Committee member) / Turaga, Pavan (Committee member) / Arizona State University (Publisher)
Created2015
Description
Stream processing has emerged as an important model of computation especially in the context of multimedia and communication sub-systems of embedded System-on-Chip (SoC) architectures. The dataflow nature of streaming applications allows them to be most naturally expressed as a set of kernels iteratively operating on continuous streams of data. The kernels are computationally intensive and are mainly characterized by real-time constraints that demand high throughput and data bandwidth with limited global data reuse. Conventional architectures fail to meet these demands due to their poorly matched execution models and the overheads associated with instruction and data movements.
This work presents StreamWorks, a multi-core embedded architecture for energy-efficient stream computing. The basic processing element in the StreamWorks architecture is the StreamEngine (SE) which is responsible for iteratively executing a stream kernel. SE introduces an instruction locking mechanism that exploits the iterative nature of the kernels and enables fine-grain instruction reuse. Each instruction in a SE is locked to a Reservation Station (RS) and revitalizes itself after execution; thus never retiring from the RS. The entire kernel is hosted in RS Banks (RSBs) close to functional units for energy-efficient instruction delivery. The dataflow semantics of stream kernels are captured by a context-aware dataflow execution mode that efficiently exploits the Instruction Level Parallelism (ILP) and Data-level parallelism (DLP) within stream kernels.
Multiple SEs are grouped together to form a StreamCluster (SC) that communicate via a local interconnect. A novel software FIFO virtualization technique with split-join functionality is proposed for efficient and scalable stream communication across SEs. The proposed communication mechanism exploits the Task-level parallelism (TLP) of the stream application. The performance and scalability of the communication mechanism is evaluated against the existing data movement schemes for scratchpad based multi-core architectures. Further, overlay schemes and architectural support are proposed that allow hosting any number of kernels on the StreamWorks architecture. The proposed oevrlay schemes for code management supports kernel(context) switching for the most common use cases and can be adapted for any multi-core architecture that use software managed local memories.
The performance and energy-efficiency of the StreamWorks architecture is evaluated for stream kernel and application benchmarks by implementing the architecture in 45nm TSMC and comparison with a low power RISC core and a contemporary accelerator.
This work presents StreamWorks, a multi-core embedded architecture for energy-efficient stream computing. The basic processing element in the StreamWorks architecture is the StreamEngine (SE) which is responsible for iteratively executing a stream kernel. SE introduces an instruction locking mechanism that exploits the iterative nature of the kernels and enables fine-grain instruction reuse. Each instruction in a SE is locked to a Reservation Station (RS) and revitalizes itself after execution; thus never retiring from the RS. The entire kernel is hosted in RS Banks (RSBs) close to functional units for energy-efficient instruction delivery. The dataflow semantics of stream kernels are captured by a context-aware dataflow execution mode that efficiently exploits the Instruction Level Parallelism (ILP) and Data-level parallelism (DLP) within stream kernels.
Multiple SEs are grouped together to form a StreamCluster (SC) that communicate via a local interconnect. A novel software FIFO virtualization technique with split-join functionality is proposed for efficient and scalable stream communication across SEs. The proposed communication mechanism exploits the Task-level parallelism (TLP) of the stream application. The performance and scalability of the communication mechanism is evaluated against the existing data movement schemes for scratchpad based multi-core architectures. Further, overlay schemes and architectural support are proposed that allow hosting any number of kernels on the StreamWorks architecture. The proposed oevrlay schemes for code management supports kernel(context) switching for the most common use cases and can be adapted for any multi-core architecture that use software managed local memories.
The performance and energy-efficiency of the StreamWorks architecture is evaluated for stream kernel and application benchmarks by implementing the architecture in 45nm TSMC and comparison with a low power RISC core and a contemporary accelerator.
ContributorsPanda, Amrit (Author) / Chatha, Karam S. (Thesis advisor) / Wu, Carole-Jean (Thesis advisor) / Chakrabarti, Chaitali (Committee member) / Shrivastava, Aviral (Committee member) / Arizona State University (Publisher)
Created2014
Description
Energy consumption of the data centers worldwide is rapidly growing fueled by ever-increasing demand for Cloud computing applications ranging from social networking to e-commerce. Understandably, ensuring energy-efficiency and sustainability of Cloud data centers without compromising performance is important for both economic and environmental reasons. This dissertation develops a cyber-physical multi-tier server and workload management architecture which operates at the local and the global (geo-distributed) data center level. We devise optimization frameworks for each tier to optimize energy consumption, energy cost and carbon footprint of the data centers. The proposed solutions are aware of various energy management tradeoffs that manifest due to the cyber-physical interactions in data centers, while providing provable guarantee on the solutions' computation efficiency and energy/cost efficiency. The local data center level energy management takes into account the impact of server consolidation on the cooling energy, avoids cooling-computing power tradeoff, and optimizes the total energy (computing and cooling energy) considering the data centers' technology trends (servers' power proportionality and cooling system power efficiency). The global data center level cost management explores the diversity of the data centers to minimize the utility cost while satisfying the carbon cap requirement of the Cloud and while dealing with the adversity of the prediction error on the data center parameters. Finally, the synergy of the local and the global data center energy and cost optimization is shown to help towards achieving carbon neutrality (net-zero) in a cost efficient manner.
ContributorsAbbasi, Zahra (Author) / Gupta, Sandeep K. S. (Thesis advisor) / Chakrabarti, Chaitali (Committee member) / Shrivastava, Aviral (Committee member) / Wu, Carole-Jean (Committee member) / Arizona State University (Publisher)
Created2014
Description
In recent years, we have observed the prevalence of stream applications in many embedded domains. Stream programs distinguish themselves from traditional sequential programming languages through well defined independent actors, explicit data communication, and stable code/data access patterns. In order to achieve high performance and low power, scratch pad memory (SPM) has been introduced in today's embedded multicore processors. Current design frameworks for developing stream applications on SPM enhanced embedded architectures typically do not include a compiler that can perform automatic partitioning, mapping and scheduling under limited on-chip SPM capacities and memory access delays. Consequently, many designs are implemented manually, which leads to lengthy tasks and inferior designs. In this work, optimization techniques that automatically compile stream programs onto embedded multi-core architectures are proposed. As an initial case study, we implemented an automatic target recognition (ATR) algorithm on the IBM Cell Broadband Engine (BE). Then integer linear programming (ILP) and heuristic approaches were proposed to schedule stream programs on a single core embedded processor that has an SPM with code overlay. Later, ILP and heuristic approaches for Compiling Stream programs on SPM enhanced Multicore Processors (CSMP) were studied. The proposed CSMP ILP and heuristic approaches do not optimize for cycles in stream applications. Further, the number of software pipeline stages in the implementation is dependent on actor to processing engine (PE) mapping and is uncontrollable. We next presented a Retiming technique for Throughput optimization on Embedded Multi-core processors (RTEM). RTEM approach inherently handles cycles and can accept an upper bound on the number of software pipeline stages to be generated. We further enhanced RTEM by incorporating unrolling (URSTEM) that preserves all the beneficial properties of RTEM heuristic and also scales with the number of PEs through unrolling.
ContributorsChe, Weijia (Author) / Chatha, Karam Singh (Thesis advisor) / Vrudhula, Sarma (Committee member) / Chakrabarti, Chaitali (Committee member) / Shrivastava, Aviral (Committee member) / Arizona State University (Publisher)
Created2012
Description
Nearly 60% of the world population uses a mobile phone, which is typically powered by a system-on-chip (SoC). While the mobile platform capabilities range widely, responsiveness, long battery life and reliability are common design concerns that are crucial to remain competitive. Consequently, state-of-the-art mobile platforms have become highly heterogeneous by combining a powerful SoC with numerous other resources, including display, memory, power management IC, battery and wireless modems. Furthermore, the SoC itself is a heterogeneous resource that integrates many processing elements, such as CPU cores, GPU, video, image, and audio processors. Therefore, CPU cores do not dominate the platform power consumption under many application scenarios.
Competitive performance requires higher operating frequency, and leads to larger power consumption. In turn, power consumption increases the junction and skin temperatures, which have adverse effects on the device reliability and user experience. As a result, allocating the power budget among the major platform resources and temperature control have become fundamental consideration for mobile platforms. Dynamic thermal and power management algorithms address this problem by putting a subset of the processing elements or shared resources to sleep states, or throttling their frequencies. However, an adhoc approach could easily cripple the performance, if it slows down the performance-critical processing element. Furthermore, mobile platforms run a wide range of applications with time varying workload characteristics, unlike early generations, which supported only limited functionality. As a result, there is a need for adaptive power and performance management approaches that consider the platform as a whole, rather than focusing on a subset. Towards this need, our specific contributions include (a) a framework to dynamically select the Pareto-optimal frequency and active cores for the heterogeneous CPUs, such as ARM big.Little architecture, (b) a dynamic power budgeting approach for allocating optimal power consumption to the CPU and GPU using performance sensitivity models for each PE, (c) an adaptive GPU frame time sensitivity prediction model to aid power management algorithms, and (d) an online learning algorithm that constructs adaptive run-time models for non-stationary workloads.
Competitive performance requires higher operating frequency, and leads to larger power consumption. In turn, power consumption increases the junction and skin temperatures, which have adverse effects on the device reliability and user experience. As a result, allocating the power budget among the major platform resources and temperature control have become fundamental consideration for mobile platforms. Dynamic thermal and power management algorithms address this problem by putting a subset of the processing elements or shared resources to sleep states, or throttling their frequencies. However, an adhoc approach could easily cripple the performance, if it slows down the performance-critical processing element. Furthermore, mobile platforms run a wide range of applications with time varying workload characteristics, unlike early generations, which supported only limited functionality. As a result, there is a need for adaptive power and performance management approaches that consider the platform as a whole, rather than focusing on a subset. Towards this need, our specific contributions include (a) a framework to dynamically select the Pareto-optimal frequency and active cores for the heterogeneous CPUs, such as ARM big.Little architecture, (b) a dynamic power budgeting approach for allocating optimal power consumption to the CPU and GPU using performance sensitivity models for each PE, (c) an adaptive GPU frame time sensitivity prediction model to aid power management algorithms, and (d) an online learning algorithm that constructs adaptive run-time models for non-stationary workloads.
ContributorsGupta, Ujjwala (Author) / Ogras, Umit Y. (Thesis advisor) / Chakrabarti, Chaitali (Committee member) / Kishinevsky, Michael (Committee member) / Dutt, Nikil (Committee member) / Arizona State University (Publisher)
Created2018
Description
Deep neural networks (DNN) have shown tremendous success in various cognitive tasks, such as image classification, speech recognition, etc. However, their usage on resource-constrained edge devices has been limited due to high computation and large memory requirement.
To overcome these challenges, recent works have extensively investigated model compression techniques such as element-wise sparsity, structured sparsity and quantization. While most of these works have applied these compression techniques in isolation, there have been very few studies on application of quantization and structured sparsity together on a DNN model.
This thesis co-optimizes structured sparsity and quantization constraints on DNN models during training. Specifically, it obtains optimal setting of 2-bit weight and 2-bit activation coupled with 4X structured compression by performing combined exploration of quantization and structured compression settings. The optimal DNN model achieves 50X weight memory reduction compared to floating-point uncompressed DNN. This memory saving is significant since applying only structured sparsity constraints achieves 2X memory savings and only quantization constraints achieves 16X memory savings. The algorithm has been validated on both high and low capacity DNNs and on wide-sparse and deep-sparse DNN models. Experiments demonstrated that deep-sparse DNN outperforms shallow-dense DNN with varying level of memory savings depending on DNN precision and sparsity levels. This work further proposed a Pareto-optimal approach to systematically extract optimal DNN models from a huge set of sparse and dense DNN models. The resulting 11 optimal designs were further evaluated by considering overall DNN memory which includes activation memory and weight memory. It was found that there is only a small change in the memory footprint of the optimal designs corresponding to the low sparsity DNNs. However, activation memory cannot be ignored for high sparsity DNNs.
To overcome these challenges, recent works have extensively investigated model compression techniques such as element-wise sparsity, structured sparsity and quantization. While most of these works have applied these compression techniques in isolation, there have been very few studies on application of quantization and structured sparsity together on a DNN model.
This thesis co-optimizes structured sparsity and quantization constraints on DNN models during training. Specifically, it obtains optimal setting of 2-bit weight and 2-bit activation coupled with 4X structured compression by performing combined exploration of quantization and structured compression settings. The optimal DNN model achieves 50X weight memory reduction compared to floating-point uncompressed DNN. This memory saving is significant since applying only structured sparsity constraints achieves 2X memory savings and only quantization constraints achieves 16X memory savings. The algorithm has been validated on both high and low capacity DNNs and on wide-sparse and deep-sparse DNN models. Experiments demonstrated that deep-sparse DNN outperforms shallow-dense DNN with varying level of memory savings depending on DNN precision and sparsity levels. This work further proposed a Pareto-optimal approach to systematically extract optimal DNN models from a huge set of sparse and dense DNN models. The resulting 11 optimal designs were further evaluated by considering overall DNN memory which includes activation memory and weight memory. It was found that there is only a small change in the memory footprint of the optimal designs corresponding to the low sparsity DNNs. However, activation memory cannot be ignored for high sparsity DNNs.
ContributorsSrivastava, Gaurav (Author) / Seo, Jae-Sun (Thesis advisor) / Chakrabarti, Chaitali (Committee member) / Berisha, Visar (Committee member) / Arizona State University (Publisher)
Created2018
Description
Vision processing on traditional architectures is inefficient due to energy-expensive off-chip data movements. Many researchers advocate pushing processing close to the sensor to substantially reduce data movements. However, continuous near-sensor processing raises the sensor temperature, impairing the fidelity of imaging/vision tasks.
The work characterizes the thermal implications of using 3D stacked image sensors with near-sensor vision processing units. The characterization reveals that near-sensor processing reduces system power but degrades image quality. For reasonable image fidelity, the sensor temperature needs to stay below a threshold, situationally determined by application needs. Fortunately, the characterization also identifies opportunities -- unique to the needs of near-sensor processing -- to regulate temperature based on dynamic visual task requirements and rapidly increase capture quality on demand.
Based on the characterization, the work proposes and investigate two thermal management strategies -- stop-capture-go and seasonal migration -- for imaging-aware thermal management. The work present parameters that govern the policy decisions and explore the trade-offs between system power and policy overhead. The work's evaluation shows that the novel dynamic thermal management strategies can unlock the energy-efficiency potential of near-sensor processing with minimal performance impact, without compromising image fidelity.
The work characterizes the thermal implications of using 3D stacked image sensors with near-sensor vision processing units. The characterization reveals that near-sensor processing reduces system power but degrades image quality. For reasonable image fidelity, the sensor temperature needs to stay below a threshold, situationally determined by application needs. Fortunately, the characterization also identifies opportunities -- unique to the needs of near-sensor processing -- to regulate temperature based on dynamic visual task requirements and rapidly increase capture quality on demand.
Based on the characterization, the work proposes and investigate two thermal management strategies -- stop-capture-go and seasonal migration -- for imaging-aware thermal management. The work present parameters that govern the policy decisions and explore the trade-offs between system power and policy overhead. The work's evaluation shows that the novel dynamic thermal management strategies can unlock the energy-efficiency potential of near-sensor processing with minimal performance impact, without compromising image fidelity.
ContributorsKodukula, Venkatesh (Author) / LiKamWa, Robert (Thesis advisor) / Chakrabarti, Chaitali (Committee member) / Brunhaver, John (Committee member) / Arizona State University (Publisher)
Created2019
Description
The detection and segmentation of objects appearing in a natural scene, often referred to as Object Detection, has gained a lot of interest in the computer vision field. Although most existing object detectors aim to detect all the objects in a given scene, it is important to evaluate whether these methods are capable of detecting the salient objects in the scene when constraining the number of proposals that can be generated due to constraints on timing or computations during execution. Salient objects are objects that tend to be more fixated by human subjects. The detection of salient objects is important in applications such as image collection browsing, image display on small devices, and perceptual compression.
This thesis proposes a novel evaluation framework that analyses the performance of popular existing object proposal generators in detecting the most salient objects. This work also shows that, by incorporating saliency constraints, the number of generated object proposals and thus the computational cost can be decreased significantly for a target true positive detection rate (TPR).
As part of the proposed framework, salient ground-truth masks are generated from the given original ground-truth masks for a given dataset. Given an object detection dataset, this work constructs salient object location ground-truth data, referred to here as salient ground-truth data for short, that only denotes the locations of salient objects. This is obtained by first computing a saliency map for the input image and then using it to assign a saliency score to each object in the image. Objects whose saliency scores are sufficiently high are referred to as salient objects. The detection rates are analyzed for existing object proposal generators with respect to the original ground-truth masks and the generated salient ground-truth masks.
As part of this work, a salient object detection database with salient ground-truth masks was constructed from the PASCAL VOC 2007 dataset. Not only does this dataset aid in analyzing the performance of existing object detectors for salient object detection, but it also helps in the development of new object detection methods and evaluating their performance in terms of successful detection of salient objects.
This thesis proposes a novel evaluation framework that analyses the performance of popular existing object proposal generators in detecting the most salient objects. This work also shows that, by incorporating saliency constraints, the number of generated object proposals and thus the computational cost can be decreased significantly for a target true positive detection rate (TPR).
As part of the proposed framework, salient ground-truth masks are generated from the given original ground-truth masks for a given dataset. Given an object detection dataset, this work constructs salient object location ground-truth data, referred to here as salient ground-truth data for short, that only denotes the locations of salient objects. This is obtained by first computing a saliency map for the input image and then using it to assign a saliency score to each object in the image. Objects whose saliency scores are sufficiently high are referred to as salient objects. The detection rates are analyzed for existing object proposal generators with respect to the original ground-truth masks and the generated salient ground-truth masks.
As part of this work, a salient object detection database with salient ground-truth masks was constructed from the PASCAL VOC 2007 dataset. Not only does this dataset aid in analyzing the performance of existing object detectors for salient object detection, but it also helps in the development of new object detection methods and evaluating their performance in terms of successful detection of salient objects.
ContributorsKotamraju, Sai Prajwal (Author) / Karam, Lina J (Thesis advisor) / Yu, Hongbin (Committee member) / Jayasuriya, Suren (Committee member) / Arizona State University (Publisher)
Created2019
Description
With the massive multithreading execution feature, graphics processing units (GPUs) have been widely deployed to accelerate general-purpose parallel workloads (GPGPUs). However, using GPUs to accelerate computation does not always gain good performance improvement. This is mainly due to three inefficiencies in modern GPU and system architectures.
First, not all parallel threads have a uniform amount of workload to fully utilize GPU’s computation ability, leading to a sub-optimal performance problem, called warp criticality. To mitigate the degree of warp criticality, I propose a Criticality-Aware Warp Acceleration mechanism, called CAWA. CAWA predicts and accelerates the critical warp execution by allocating larger execution time slices and additional cache resources to the critical warp. The evaluation result shows that with CAWA, GPUs can achieve an average of 1.23x speedup.
Second, the shared cache storage in GPUs is often insufficient to accommodate demands of the large number of concurrent threads. As a result, cache thrashing is commonly experienced in GPU’s cache memories, particularly in the L1 data caches. To alleviate the cache contention and thrashing problem, I develop an instruction aware Control Loop Based Adaptive Bypassing algorithm, called Ctrl-C. Ctrl-C learns the cache reuse behavior and bypasses a portion of memory requests with the help of feedback control loops. The evaluation result shows that Ctrl-C can effectively improve cache utilization in GPUs and achieve an average of 1.42x speedup for cache sensitive GPGPU workloads.
Finally, GPU workloads and the co-located processes running on the host chip multiprocessor (CMP) in a heterogeneous system setup can contend for memory resources in multiple levels, resulting in significant performance degradation. To maximize the system throughput and balance the performance degradation of all co-located applications, I design a scalable performance degradation predictor specifically for heterogeneous systems, called HeteroPDP. HeteroPDP predicts the application execution time and schedules OpenCL workloads to run on different devices based on the optimization goal. The evaluation result shows HeteroPDP can improve the system fairness from 24% to 65% when an OpenCL application is co-located with other processes, and gain an additional 50% speedup compared with always offloading the OpenCL workload to GPUs.
In summary, this dissertation aims to provide insights for the future microarchitecture and system architecture designs by identifying, analyzing, and addressing three critical performance problems in modern GPUs.
First, not all parallel threads have a uniform amount of workload to fully utilize GPU’s computation ability, leading to a sub-optimal performance problem, called warp criticality. To mitigate the degree of warp criticality, I propose a Criticality-Aware Warp Acceleration mechanism, called CAWA. CAWA predicts and accelerates the critical warp execution by allocating larger execution time slices and additional cache resources to the critical warp. The evaluation result shows that with CAWA, GPUs can achieve an average of 1.23x speedup.
Second, the shared cache storage in GPUs is often insufficient to accommodate demands of the large number of concurrent threads. As a result, cache thrashing is commonly experienced in GPU’s cache memories, particularly in the L1 data caches. To alleviate the cache contention and thrashing problem, I develop an instruction aware Control Loop Based Adaptive Bypassing algorithm, called Ctrl-C. Ctrl-C learns the cache reuse behavior and bypasses a portion of memory requests with the help of feedback control loops. The evaluation result shows that Ctrl-C can effectively improve cache utilization in GPUs and achieve an average of 1.42x speedup for cache sensitive GPGPU workloads.
Finally, GPU workloads and the co-located processes running on the host chip multiprocessor (CMP) in a heterogeneous system setup can contend for memory resources in multiple levels, resulting in significant performance degradation. To maximize the system throughput and balance the performance degradation of all co-located applications, I design a scalable performance degradation predictor specifically for heterogeneous systems, called HeteroPDP. HeteroPDP predicts the application execution time and schedules OpenCL workloads to run on different devices based on the optimization goal. The evaluation result shows HeteroPDP can improve the system fairness from 24% to 65% when an OpenCL application is co-located with other processes, and gain an additional 50% speedup compared with always offloading the OpenCL workload to GPUs.
In summary, this dissertation aims to provide insights for the future microarchitecture and system architecture designs by identifying, analyzing, and addressing three critical performance problems in modern GPUs.
ContributorsLee, Shin-Ying (Author) / Wu, Carole-Jean (Thesis advisor) / Chakrabarti, Chaitali (Committee member) / Ren, Fengbo (Committee member) / Shrivastava, Aviral (Committee member) / Arizona State University (Publisher)
Created2017