Empirical Research on Workload and Work Performance of University Faculty: Moderating Roles of Self-monitoring and Work Characteristics

종하몽 계명대학교 대학원 2023 국내박사

중학교 교사의 업무부담 및 교사효능감과 학교효과성 인식의 관계

김대안 전북대학교 교육대학원 2022 국내석사

The purpose of this study is to analyze perceived difference of workload, teacher efficacy, and school effectiveness of middle school teachers by their demographic variables, and to investigate the effect of workload and teacher efficacy on school effectiveness. To this end, four different research questions were examined as follows: First, What is the difference in perception of workload according to the background characteristics of middle school teachers? Second, What is the difference in perception of teacher efficacy according to the background characteristics of middle school teachers? Third, What is the difference in perception of school effectiveness according to the background characteristics of middle school teachers? Fourth, what is the relationship between workload, teacher efficacy and school effectiveness perception of middle school teacher? To achieve the purpose of this study, the relationship among workload, teacher efficacy, and school effectiveness was analyzed using the data 『Gyeonggi Education Panel Study』, which was conducted by the Gyeonggi Institute of Education. The major findings of this research are as follows: First, it has been shown that the workload of middle school teachers were at a slightly higher level than average. According to a thorough examination of the workload perceived by middle school teachers according to their background variables, teachers who has less experience, teachers who are on a fixed-term contract, teachers who are in charge of a homeroom, teachers who have a lower position, and younger teachers were suffering from a high level of perceived workload. Second, it has been shown that the level of class guidance teacher efficacy and life guidance teacher efficacy perceived by middle school teachers were high. According to the analysis of perceived class guidance teacher efficacy based on the teacher’s background variables, it has been found that teachers with more experience, higher position, higher level of education, higher age, and teachers who are not a homeroom teacher had a higher level of perceived class guidance teacher efficacy. Meanwhile, it has also been found that teachers with more experience, higher age, teachers who are on a fixed-term contract, and chief teachers had a higher level of perceived life guidance teacher efficacy. Third, middle school teachers generally had a high level of perceived job satisfaction, school satisfaction, teaching/learning activity, principal’s leadership, school culture, and cooperative network with local community, which are all sub-factors of school effectiveness. According to the analysis of job satisfaction perceived by middle teacher’s background variables, it has been found that teachers who have more than 30 years of experience, teachers who are not a homeroom teacher, teachers who are on a fixed-term contract, chief teachers, male teachers, and older teachers had a higher level of perceived job satisfaction. Fourth, the workload and school effectiveness of middle school teachers had a low, but significant positive correlation, while the correlation between class guidance teacher efficacy/life guidance teacher efficacy and school effectiveness were somewhat high. According to the result of a multiple regression analysis on the effect of workload on teacher efficacy, it has been found that workload did not have a significant influence on teacher efficacy. In addition, the result of a multiple regression analysis on the influence of workload, class guidance/life guidance teacher efficacy on school effectiveness has revealed that workload had a low but significant positive impact on school effectiveness, while class guidance/life guidance teacher efficacy had a signifiacnt positive impact on school effectiveness. The findings of this study suggests these implications. First, it is necessary to prepare measures to improve and support the treatment of fixed-term teachers and low-experience teachers. Second, supplemen tary measures are needed for institutional support implemented to redu ce teachers’ work. Third, there is a need for a plan to promote teacher efficacy in order to increase school effectiveness.

System Architecture and Optimization Strategies for Efficient AI Workload Management

권석민 이화여자대학교 대학원 2025 국내박사

인공지능(AI) 기술은 다양한 산업 분야에 혁신적인 변화를 가져오고 있으며, 이를 뒷받침하기 위한 컴퓨팅 시스템의 발전은 필수적이다. 특히, AI 워크로드는 고도의 연산 자원과 복잡한 데이터 흐름을 요구하기 때문에 이를 최적화하고 효율적으로 처리하기 위한 시스템 아키텍처 및 관리 기술의 개발은 중요한 연구 과제로 부각되고 있다. 본 논문은 온디바이스(On-device) 환경, 클라우드 클러스터 환경, 그리고 클라우드 컴퓨팅 환경이라는 세 가지 주요 컴퓨팅 환경에서 AI 워크로드의 효율성을 극대화하기 위한 새로운 접근법을 제시한다. 첫 번째 연구는 온디바이스 환경에서 제한된 자원을 효과적으로 활용하여 AI 워크로드의 성능 저하를 방지하는 데 초점을 맞춘다. AI 워크로드의 메모리 참조 특성을 심층적으로 분석하고, 이를 기반으로 비휘발성 메모리를 활용한 새로운 시스템 아키텍처를 제안한다. 분석 결과, AI 워크로드는 시간 지역성이 낮고 불규칙한 쓰기 패턴을 보이는 특징이 있으며, 이는 전통적인 시스템 설계로는 처리 성능에 심각한 영향을 미친다. 제안된 아키텍처는 비휘발성 메모리를 쓰기 가속기로 활용하여 이러한 문제를 효과적으로 해결하며, 시뮬레이션 결과 기존 시스템 대비 메모리 입출력 성능을 80% 이상 개선한다. 두 번째 연구는 다중 테넌트 클라우드 클러스터에서 GPU 자원의 활용률을 극대화하기 위한 스케줄링 전략을 탐구한다. 클라우드 환경은 이종 GPU와 같은 다양한 자원이 공존하며, 자원 단편화 문제로 인해 GPU 활용률이 저하되는 문제가 빈번히 발생한다. 본 연구는 유전 알고리즘 기반의 스케줄링 방법론을 제안하며, 이를 통해 클러스터 자원의 효율적 사용과 GPU 활용률 개선을 동시에 달성한다. 실제 클러스터 워크로드 데이터를 활용한 실험 결과, 제안된 방법은 기존 라운드 로빈 및 Tetris 알고리즘 대비 GPU 활용률을 12.8% 향상시키며, 작업 완료 시간에도 부정적인 영향을 미치지 않는다. 세 번째 연구는 클라우드 컴퓨팅 환경에서 자원 할당과 전력 관리를 통합적으로 최적화하기 위한 계층적 프레임워크를 제시한다. 기존의 자원 관리 기법은 동적인 AI 워크로드의 자원 요구를 효율적으로 처리하지 못하며, 전력 소비와 성능 간의 균형을 유지하는 데 한계를 가진다. 본 연구는 최신 딥 강화 학습(DRL) 기술을 기반으로 글로벌 자원 할당과 로컬 전력 관리를 통합한 프레임워크를 설계한다. 이를 통해 전력 소비를 최대 13.97% 절감하면서도 작업 지연 시간을 최소화하는 데 성공하며, 클라우드 환경에서의 자원 관리 효율성을 크게 향상시킨다. 본 논문은 온디바이스 환경에서의 메모리 관리, 클라우드 클러스터에서의 스케줄링 최적화, 그리고 클라우드 컴퓨팅 환경에서의 자원 및 전력 관리 통합이라는 세 가지 상이한 연구를 통해 AI 워크로드의 효율적 처리를 위한 통합적이고 체계적인 접근법을 제시한다. 이를 통해 다양한 컴퓨팅 환경에서 AI 워크로드의 성능과 자원 활용성을 극대화하며, 차세대 AI 응용 기술의 폭넓은 도입과 지속 가능성을 뒷받침하는 중요한 기반을 제공한다. AI 기술이 다양한 산업 분야에서 실질적이고 혁신적인 영향을 미칠 수 있도록 지원하는 데 기여할 것으로 기대된다. Artificial Intelligence (AI) technology is bringing about innovative changes across various industries, and the advancement of computing systems to support this is essential. In particular, AI workloads require high computational resources and complex data flows, making the development of system architectures and management technologies to optimize and efficiently process these workloads an important research task. This paper presents a new approach to maximize the efficiency of AI workloads in three main computing environments: On-device environments, cloud cluster environments, and cloud computing environments. The first study focuses on effectively utilizing limited resources in an On-device environment to prevent performance degradation of AI workloads. It conducts an in-depth analysis of the memory reference characteristics of AI workloads and proposes a new system architecture that leverages non-volatile memory based on this analysis. The results show that AI workloads exhibit low temporal locality and irregular write patterns, which can severely impact processing performance with traditional system designs. The proposed architecture effectively addresses these issues by using non-volatile memory as a write accelerator, and simulation results indicate an improvement of over 80% in memory input/output performance compared to existing systems. The second study explores scheduling strategies to maximize GPU resource utilization in multi-tenant cloud clusters. In cloud environments, various resources coexist, such as heterogeneous GPUs, and issues frequently arise due to resource fragmentation, leading to decreased GPU utilization. This study proposes a scheduling methodology based on genetic algorithms, which simultaneously achieves efficient use of cluster resources and improved GPU utilization. Experimental results using actual cluster workload data show that the proposed method enhances GPU utilization by 12.8% compared to existing round-robin and Tetris algorithms, without negatively impacting job completion times. The third study presents a hierarchical framework for the integrated optimization of resource allocation and power management in cloud computing environments. Existing resource management techniques struggle to efficiently handle the resource demands of dynamic AI workloads and have limitations in maintaining a balance between power consumption and performance. This research designs a framework that integrates global resource allocation and local power management based on the latest deep reinforcement learning (DRL) technology. As a result, it successfully reduces power consumption by up to 13.97% while minimizing job latency, significantly enhancing resource management efficiency in cloud environments. This paper presents an integrated and systematic approach for the efficient processing of AI workloads through three distinct studies: memory management in on-device environments, scheduling optimization in cloud clusters, and the integration of resource and power management in cloud computing environments. This approach aims to maximize the performance and resource utilization of AI workloads across various computing environments, providing a crucial foundation for the broad adoption and sustainability of next-generation AI application technologies. It is expected to contribute to supporting AI technologies in making practical and innovative impacts across various industrial sectors.

Prediction of multiple sources of mental workload under time pressure using ACT-R

Sungjin Park 고려대학교 대학원 2024 국내박사

Mental workload is the most critical element in the design of new systems or interfaces because it heavily influences the overall performance of operators. Due to its complexity, direct measurement is challenging. Prediction models offer an alternative solution, enabling proactive adjustments to interfaces and training procedures before real-world deployment. Despite their advantages, current prediction models may not adequately address all major sources of mental workload. Also, they might fail to capture the impact of time pressure on information processing and workload. This study proposes a mathematical method for quantitatively computing multiple sources of workload and reflecting the effect of time pressure with a cognitive architecture, Adaptive Control of Thought Rationale (ACT-R). Specifically, mathematical equations were developed to quantify each subscale of the National Aeronautics and Space Administration task load index (NASA-TLX) and the time pressure effect on information processing speed. The proposed method not only successfully predicts each subscale of the NASA-TLX but also models the change of human performance and workload by time pressure. This study compared predicted values with actual performance times and subjective ratings from the NASA-TLX, gathered from participants in two experiments. The first experiment focused on task-related sources of mental workload using simple laboratory tasks, such as menu selection and visual-manual tasks. The second experiment expanded on the first by considering all sources of mental workload and incorporating the effect of time pressure using the Kanfer-Ackerman Air Traffic Control (KA-ATC) task. In practical application, the proposed method offers a cost-effective alternative to traditional operator-in-the-loop assessments during the early stages of system design. It uses computer simulation, eliminating the need for prototypes and trained operators. Additionally, the method identifies specific workload bottlenecks within the operator. This allows for targeted interface adjustments. Furthermore, the method meets key criteria for mental workload measures, demonstrating sensitivity, diagnosticity, selectivity, and reliability. 정신적 작업 부하는 작업자의 전반적인 성능에 큰 영향을 미치기 때문에 새로운 시스템이나 인터페이스 설계 시 가장 중요한 요소입니다. 하지만 정신적 작업 부하는 복합적이고 다양한 요소에 의해 영향을 받기에 직접적으로 측정하기 어렵습니다. 예측 모델은 이러한 문제에 대한 대안의 해결책이 될 수 있으며, 다양한 시나리오에 적용되어 실제 사용 환경에 투입되기 전에 시스템 및 인터페이스를 개선할 수 있는 기회를 제공합니다. 그럼에도 불구하고 현재 대부분의 예측 모델은 작업자의 정신적 작업 부하에 영향을 미치는 주요 속성들을 충분히 다루지 못하거나, 정보 처리 및 정신적 작업 부하에 큰 영향을 미치는 시간 압박의 영향을 제대로 반영하지 못합니다. 본 연구에서는 Adaptive Control of Thought Rationale (ACT-R) 인지 아키텍처를 사용하여 여러 작업 부하 원인을 정량적으로 계산하고 시간 압박의 영향을 반영하는 수학적 방법을 제안합니다. 구체적으로, National Aeronautics and Space Administration task load index (NASA-TLX)의 각 하위 척도와 정보 처리 속도에 대한 시간 압박 효과를 정량적으로 예측하기 위한 수학적 방정식을 개발했습니다. 실험을 통해 제안된 방법이 NASA-TLX의 각 하위 척도를 성공적으로 추정할 수 있을 뿐만 아니라 시간 압박에 따른 작업 수행시간과 정신적 작업 부하의 변화를 예측할 수 있다는 것을 보여주었습니다. 이 방법을 검증하기 위해 두 가지 실험을 수행했습니다. 첫 번째 실험은 단순 실험실 환경의 작업을 통해 정신적 작업 부하의 작업 관련 속성에 대한 검증에 중점을 두었습니다. 두 번째 실험은 Kanfer-Ackerman 항공 교통 관제 (KA-ATC) 작업을 사용하여 정신적 작업 부하의 모든 속성과 시간 압박 효과를 포함했습니다. 실용적인 측면에서, 본 연구에서 제안된 방법은 초기 시스템 설계 단계에서 전통적인 작업자 참여 평가에 대한 비용 효율적인 대안을 제공합니다. 컴퓨터 시뮬레이션을 기반으로 하여 프로토타입이나 훈련된 피험자가 필요하지 않습니다. 또한, 작업자의 정신적 부하가 어디에 집중되는지를 파악할 수 있으며, 이를 통해 인터페이스 조정 및 디자인 개선을 가능하게 합니다. 더불어, 제안된 방법은 민감도, 진단성, 선택성, 신뢰성과 같은 정신적 작업 부하 측정 지표의 주요 기준을 충족합니다.

Team Workload in Action Teams

Johnson, Craig Arizona State University ProQuest Dissertations & 2023 해외박사(DDOD)

A key contribution of human factors engineering is the concept of workload: a construct that represents the relationship between an operator’s cognitive resources, the demands of their task, and performance. Understanding workload can lead to improvements in safety and performance for people working in critical environments, particularly within action teams. Recently, there has been interest in considering how the workload of a team as a whole may differ from that of an individual, prompting investigation into team workload as a distinct team-level construct. In empirical research, team-level workload is often considered as the sum or average of individual team members' workloads. However, the intrinsic characteristics of action teams—such as interdependence and heterogeneity—challenge this assumption, and traditional methods of measuring team workload might be unsuitable. This dissertation delves into this issue with a review of empirical work in action teams, pinpointing several gaps. Next, the development of a testbed is described and used to address two pressing gaps regarding the impact of interdependence and how team communications relate to team workload states and performance. An experiment was conducted with forty 3-person teams collaborating in an action team task. Results of this experiment suggest that the traditional way of measuring workload in action teams via subjective questionnaires averaged at the team level has some major shortcomings, particularly when demands are elevated, and action teams are highly interdependent. The results also suggested that several communication measures are associated with increases in demands, laying the groundwork for team-level communication-based measures of team workload. The results are synthesized with findings from the literature to provide a way forward for conceptualizing and measuring team workload in action teams.

AIR TRAFFIC CONTROLLER WORKLOAD EVALUATIONS FOR TERMINAL AIRSPACE CAPACITY ESTIMATION : 터미널 공역 용량 추정을 위한 항공 교통 관제사 업무량 부하 평가

JEMIMA LOUISE KAMESE Graduate School of Korea Aerospace University 2023 국내석사

The Aviation and Air Transportation Industry is a developing industry that continues to grow at an exponential rate. This growth is attributed to immense technological advancements and constant research into increasing optimization of the air transportation. In turn, safety precautions equivalent to the magnitude of growth are required to ensure that air transportation remains safe. And thus, safety is a paramount aspect and necessity of the Aviation and Air Transportation Industry. Among the various facilitators of Air Traffic Safety, an important one with a key role are the Air Traffic Controllers (ATCos). ATCos ought to make traffic control decisions promptly and swiftly to ensure traffic safety. They are responsible for numerous tasks and the rate at which these tasks are completed, and aircrafts safely guided through the various zones of the airspace to their destination, influences controllers’ workload. Hence, decision-making is undoubtedly a crucial part of air traffic control with a direct impact on Air Traffic Controllers’ workload. This study focuses on workload evaluation for terminal capacity estimation. There are multiple ways that can be applied to estimate terminal capacity and the methods chosen in this research utilize Human-in-the-loop Simulation to assess Air Traffic Controller’s workload through NASA task load index (TLX), the use of the encephalogram (EEG) and Well Clear Score (WCS) with air traffic conditions of varying traffic complexity and density. The results based on these methods are analyzed and used to investigate the most suitable ATCo workload indicators, the factors that influence workload, and the application of these results for increasing or decreasing capacity relative to the ATCo’s capabilities. 항공 운송 산업은 기하급수적인 속도로 성장하고 있는 산업이다. 이러한 성장은 항공 운송의 안전과 기술발전, 운송 산업의 성장 최적화를 증가시키기 위한 연구에 기인하게 된다. 특히, 항공 산업의 안전을 위해서는 성장의 크기에 상응하는 안전 예방 조치가 필요할 것이다. 항공 교통 안전의 핵심적인 역할을 하는 것은 항공 교통 관제사(ATCos)이다. 항공 교통 관제사는 항공기가 출발지부터 목적지까지 안전하게 도착할 수 있도록 신속하고 안전하게 항공기의 운항을 통제하여야한다. 이러한 항공기의 통제를 위해서는 수 많은 관제사의 업무 중, 의사 결정(decision-making)이 중요한 요인이 될 것이다. 본 연구는 단말기 용량 추정을 통해 항공 교통 관제사의 업무량을 추정하는 것에 초점을 맞추고 있다. 단말기를 통해 관제사의 업무량을 추정하는 방법은 여러 가지가 있으나, 본 연구에서는 HITL(Human-in-the-loop Simulation) 실험을 통해 NASA 업무 부하 지수(TLX), 뇌파(EEG) 및 WCS(Well Clear Score)를 분석하여 관제사의 업무량을 추정한다.

Measuring and Quantifying Driver Workload on Limited Access Roads

Liu, Ke ProQuest Dissertations & Theses University of Mich 2019 해외박사(DDOD)

Minimizing driver errors should improve driving safety. Driver errors are more common when workload is high than when it is low. Thus, it is of great importance to study driver workload. Knowing the amount of workload at any given time, take-over time can be determined, adaptive in-vehicle systems can be refined, and distracting in-vehicle secondary tasks can be regulated.In this dissertation, a model quantifying workload as a function of traffic, in which workload is proportional to inverse time headway (THW) and time to collision (TTC), was proposed. Two experiments were conducted to investigate how traffic affected driver workload and evaluate the proposed model. The driving scenarios were categorized into static (i.e., no relative movements among vehicles) and dynamic (i.e., there are relative velocities and lane change actions). Three categories of workload measures (i.e., workload rating, occlusion %, and driving performance statistics) were analyzed and compared. A GOMS model was built based upon a timeline model by using timerequired to represent mental resources demanded and timeavailable to represent mental resources available.In static traffic, the workload rating increased with increased number of vehicles around but was unaffected by participant age. The workload ratings decreased with increasing Distance Headways (DHWs) of each vehicle. From greatest to least, the effects were: DHWLead, DHWLeftLead, DHWLeftFollow, DHWFollow. Any surrounding vehicle that was 14.5 m away from the participant resulted in significant greater workload. Drivers tended to compromise longitudinal speed but still maintain lateral position when workload increased. Although occlusion% was less sensitive to scenarios having no lead vehicles, it can nonetheless be well predicted using the proposed workload model in sensitive scenarios. The resulting equations were occlusion% = 0.35 + 0.05/THWLead + 0.02/THWLeftLead - 0.08Age (Rocclusion2=0.91); rating = 1.74 + 1.74/THWLead + 0.20/THWFollow + 0.79/THWLeftLead + 0.28/THWLeftFollow (Rrating2=0.73). In dynamic traffic, drivers experienced greater workload in the faster lane; higher workload level was associated with greater relative velocity between two lanes. Both rating and occlusion% can be described using the proposed model: Anchored rating = 4.53 + 1.215/THWLeftLeadLead + 0.001/THWLeftFollow + 3.069/THWLeadLead + 0.524/THWLead + 0.240/(TTCLeadxTTCLeadLateral) + 30.487/(TTCLeftLeadxTTCLeftLeadLateral) (Rrating2=0.54); Occlusion% = 0.381 + 0.150/THWLeftLeadLead ˗ 0.117/THWLeadLead + 0.021/THWLead + 2.648/(TTCLeftLeadxTTCLeftLeadLateral) (Rocclusion2=0.58). In addition, it was shown that the GOMS model accounted for the observed differences of workload ratings from the empirical data (R2>0.83).In contrast to most previous studies that focus on average long-term traffic statistics (e.g., vehicles/lane/hour), this dissertation provided equations to predict two measures of workload using real-time traffic. The comparisons among three workload measures provided insights into how to choose the desired workload measures in their future research. In GOMS model, the procedural knowledge of rating workload while driving was developed. They should be transferrable to other workload studies and can serve as the primary tool to justify experimental design.Scientifically, the results of this dissertation offer insights into the mechanism of the way that humans perceive workload and the corresponding driving strategies. From the engineering application and practical value perspective, the proposed workload model would help future driving studies by providing a way to quantify driver workload and support the comparison of studies in different situations.

Optimizing GPU-accelerated applications using workload scheduling and memory management

박정호 서울대학교 대학원 2020 국내박사

매니코어 프로세서가 탑재된 가속기를 이종 시스템은 고성능이 필요한 다양한 분야에서 널리 활용되고 있다. 그 중 GPU는 가장 널리 사용되는 가속기 중 하나이다. GPU를 효율적으로 활용하기 위해서는 최적화와 메모리 관리가 매우 중요하다. GPU의 하드웨어 구조는 고전적인 CPU와 큰 차이가 있으므로 응용프로그램의 병렬성과 GPU의 구조를 모두 고려한 최적화가 필수적이다. 또한 GPU는 메인 메모리에 비해 매우 적은 메모리를 가지고 있으므로 효율적인 메모리 관리도 필요하다. 본 논문에서는 워크그룹 스케줄링 기법과 메모리 관리 기법을 제시한다. 이 기법은 두 가지 실제 GPU 가속 응용프로그램을 대상으로 구현되었다. 첫번째로 본 논문에서는 컨트롤 플로우 다이버전스(control flow divergence)와 워크로드 불균형을 최소화 하기 위한 워크로드 스케줄링 기법을 제안한다. 컨트롤 플로우 다이번전스와 워크로드 불균형은 GPU 가속 응용프로그램의 성능을 심각하게 저하시킨다. 제안하는 워크로드 스케줄링 기법은 여러개의 통(bin)을 이용하여 작업들을 같거나 비슷한 타입으로 분류한다. 분류된 작업들이 들어있는 통은 각각의 컴퓨트 유닛에서 실행되어 컨트롤 플로우 다이버전스와 작업 불균형을 최소화 한다. 본 논문에서 해당 기법은 저가의 내장형 APU를 활용한 고속의 IPsec 게이트웨이 상에서 예시로 구현된다. APU (Accelerated Processing Unit) 는 한 칩에 일반적인 CPU와 GPU가 탑재된 이종 멀티코어 프로세서이다. 이 프로세서는 물리적으로 CPU와 GPU 사이에 메모리를 공유하는 HSA (Heterogeneous System Architecture)를 지원한다. 기존의 일반적인 GPU에서는 제거하지 못했던 CPU와 GPU의 데이터 전송 오버헤드를 HSA를 이용하여 제거한다. 본 게이트웨이는 OpenCL을 사용하여 GPU 상에서 작동하도록 구현하였으며 제로카피 패킷 입출력을 위해 DPDK를 사용하였다. IPsec 게이트웨이는 다양한 종류의 실제 네트워크 트래픽을 처리해야한다. 이를 위해 본 논문에서는 이러한 패킷 처리 상황에서 GPU의 사용률을 획기적으로 높일 수 있는 패킷 스케줄링 알고리즘을 제안한다. 해당 알고리즘은 APU 뿐만 아니라 일반적인 분산 GPU에서도 작동한다. 두번째로 GPU 메모리 부족 문제를 해결하기 위해 스와핑 메모리 기법을 기반으로 한 GPU 메모리 관리 기법을 제안한다. 딥 러닝은 매우 많은 메모리를 필요로 하므로 제안된 메모리 관리 기법을 적용하기에 매우 좋은 어플리케이션이다. 해당 기법은 tvDNN이라고 이름지어진 공유 라이브러리 형태로 구현되어 딥 러닝시에 GPU 메모리 부족 문제를 해결한다. 이를 이용하여 GPU메모리 공간 보다 훨씬 큰 메모리를 필요로 하는 DNN 모델을 학습 할 수 있다. tvDNN은 두가지 효율적인 메모리 전송 스케줄링 알고리즘을 제시하는데 하나는 ILP를 이용한 알고리즘이고 다른 하나는 휴리스틱 알고리즘이다. 해당 알고리즘은 메모리 전송과 GPU 계산을 최대한 겹치게 하는 스케줄을 만드는 것이다. 또한 메모리 오브젝트 섹셔닝 기법을 제시하여 256이나 512같은 큰 배치 사이즈 에서도 DNN 학습을 가능하게 하였다. tvDNN은 기존 DNN 프레임워크의 소스코드 수정이나 재컴파일이 필요없이 투명하게 동작하는 라이브러리 형태로 작성되었다. Caffe를 이용한 다양한 실험 결과는 기존의 다른 소프트웨어 기반의 메모리 확장 솔루션에 비해 더 좋은 성능을 보여주었다. 또한 기존의 방식에서 불가능했던 256 배치 이상의 입력도 16 GB의 GPU 메모리로 잘 동작함을 보여준다. Heterogeneous systems that contain manycore processors as accelerators are widely used to achieve high performance. GPUs are the most well-known accelerators. The optimizations considering the GPU architecture and parallelism of applications are essential for the efficient use of GPUs. Moreover, GPU memory management is necessary because GPU memory capacity is much less than the main memory. In this thesis, we propose a workload scheduling and memory management technique. We implement two real-world GPU-accelerated applications with workload scheduling and memory management techniques and show the effectiveness of the techniques. First, this thesis proposes the workload scheduling technique to minimize control flow divergence and workload imbalance. The control flow divergence and workload imbalance significantly degrade the performance of the GPU-accelerated applications. The proposed workload scheduling technique uses multiple bins to classify the workloads of the same or similar types. Each bin is executed on a compute unit in a GPU. The technique is implemented on a high-performance IPsec gateway using a low-cost commodity embedded Accelerated processing unit (APU). The IPsec gateway processed different types of packets, which created tremendous control flow divergence and workload imbalance. An APU is a heterogeneous multicore processor that contains general-purpose CPU cores and a GPU in a single chip. It also supports Heterogeneous System Architecture (HSA) that provides coherent physically-shared memory between the CPU and GPU. The gateway is implemented in OpenCL to exploit the GPU and uses zero-copy packet I/O APIs in the DPDK. The IPsec gateway handles the real-world network traffic where each packet has a different workload. The proposed packet scheduling algorithm significantly improves GPU utilization for such traffic. It works not only for APUs but also for discrete GPUs. Second, this thesis proposes a GPU memory management technique, based on a memory swapping mechanism, to address the shortage of GPU memory capacity for applications that require more memory than the memory capacity of GPUs. The technique is implemented as a shared library, called tvDNN, to address the GPU memory restriction problem for deep learning. Deep learning requires a massive amount of GPU memory, so this is a suitable example of implementing the proposed technique. The tvDNN enables us to build a DNN model that requires a larger GPU memory space than that of the target GPU. The tvDNN provides two efficient memory transfer scheduling algorithms based on ILP and heuristics to generate an optimal memory transfer schedule that maximizes the overlap between memory transfers and GPU computations. We also propose a memory object sectioning technique that enables DNN configurations with a huge input batch size, such as 256 and 512. No source code modification or recompilation of the target DNN is required to use the tvDNN. The experimental result shows that the performance of Caffe with the tvDNN is much better than the existing software-based, non-transparent GPU memory management solution. Moreover, the tvDNN enables Caffe to build the VGG-16 with a batch size of 256 or 512 using 16GB of GPU memory, which is impossible with existing solutions.

Performance prediction model for data-parallel workload in distributed processing environment

명노영 Graduate School, Korea University 2021 국내박사

Enhancing performance of big data analytics in distributed environment has been issued because most of the big data related applications such as machine learning techniques and streaming services generally utilize distributed computing frameworks. Thus, optimizing performance of those applications at Spark has been actively researched. Optimizing the performance of the applications in a distributed environment is challenging because it not only requires optimizing the applications themselves but also requires tuning the distributed system configuration parameters. The application processing procedure of the Spark is analyzed and modeled. Through the analyzed results, performance optimization schemes for each step of the procedure: inner-stage and outer-stage are proposed. Also, an appropriate partitioning heuristic is proposed by analyzing the relationship between partitioning parallelism and the performance of the applications. A machine-learning-based prediction model is proposed that determines the efficient memory for a given workload and data. To determine the validity of the proposed model, the model is applied to data-parallel workloads which include a deep learning model. The predicted memory values are in close agreement with the actual amount of required memory. Additionally, the whole building time for the proposed model requires a maximum of 44% of the total execution time of a data-parallel workload. The proposed model can improve memory efficiency up to 1.89 times compared with the vanilla Spark setting. A completion time prediction model based on machine learning for the representative deep learning model convolutional neural network (CNN) is proposed by analyzing the effects of data, task, and resource characteristics on performance when executing the model in the Spark cluster. To reduce the time utilized in collecting the data for training the model, the causal relationship between the model features and the completion time based on Spark CNN's distributed data-parallel model is considered. The model features include the configurations of the Data Center OS Mesos environment, configurations of Apache Spark, and configurations of the CNN model. By applying the proposed model to famous CNN implementations, 99.98% prediction accuracy about estimating the job completion time is achieved. In addition to the downscale search area for the model features, extrapolation is leveraged, which significantly reduces the model build time at most to 89% with even better prediction accuracy in comparison to the actual work. Enhancing performance of big data analytics in distributed environment has been issued because most of the big data related applications such as machine learning techniques and streaming services generally utilize distributed computing frameworks. Thus, optimizing performance of those applications at Spark has been actively researched. Optimizing the performance of the applications in a distributed environment is challenging because it not only requires optimizing the applications themselves but also requires tuning the distributed system configuration parameters. The application processing procedure of the Spark is analyzed and modeled. Through the analyzed results, performance optimization schemes for each step of the procedure: inner-stage and outer-stage are proposed. Also, an appropriate partitioning heuristic is proposed by analyzing the relationship between partitioning parallelism and the performance of the applications. A machine-learning-based prediction model is proposed that determines the efficient memory for a given workload and data. To determine the validity of the proposed model, the model is applied to data-parallel workloads which include a deep learning model. The predicted memory values are in close agreement with the actual amount of required memory. Additionally, the whole building time for the proposed model requires a maximum of 44% of the total execution time of a data-parallel workload. The proposed model can improve memory efficiency up to 1.89 times compared with the vanilla Spark setting. A completion time prediction model based on machine learning for the representative deep learning model convolutional neural network (CNN) is proposed by analyzing the effects of data, task, and resource characteristics on performance when executing the model in the Spark cluster. To reduce the time utilized in collecting the data for training the model, the causal relationship between the model features and the completion time based on Spark CNN's distributed data-parallel model is considered. The model features include the configurations of the Data Center OS Mesos environment, configurations of Apache Spark, and configurations of the CNN model. By applying the proposed model to famous CNN implementations, 99.98% prediction accuracy about estimating the job completion time is achieved. In addition to the downscale search area for the model features, extrapolation is leveraged, which significantly reduces the model build time at most to 89% with even better prediction accuracy in comparison to the actual work.

중국 헤이룽장성(黑龍江) 교원업무경감 정책에 대한 교사들의 인식 분석연구

이함옥 중앙대학교 대학원 2023 국내석사

The purpose of this study was to investigate how middle school teachers perceive the policy of reducing teachers' workload and to derive implications for improving the practical implementation of workload reduction policies. Based on the work areas of Chinese teachers surveyed in previous studies, 42 items were created considering the content of workload reduction policies. A survey was conducted among teachers working in the Heilongjiang Province, with a total of 101 respondents. After excluding responses with a response time of less than 100 seconds, a total of 90 responses were analyzed. Using descriptive statistical methods such as frequency, percentage, mean, and standard deviation, the study analyzed the perceptions of elementary and middle school teachers regarding the intensity, difficulty, and level of reduction in administrative tasks. Cross-analysis was conducted to examine the differences in perception of workload intensity, difficulty, and reduction among different groups of teachers. Finally, t-tests, One-way ANOVA, and Scheffé post-hoc tests were used to explore the differences in perception of workload intensity, difficulty, and reduction among elementary and middle school teachers. The analysis revealed the following results: Firstly, elementary and middle school teachers perceived the intensity of workload at a level of 3.37, indicating a relatively high intensity of workload. Secondly, they perceived the difficulty of workload at a level of 3.20, indicating a high level of difficulty. Thirdly, after the implementation of workload reduction policies, teachers' perception of the level of reduction in workload was 3.75, indicating a minimal effect of workload reduction. Based on these findings, several implications can be drawn. Firstly, there is an urgent need for standardization of teachers' tasks by the Ministry of Education, educational authorities, and individual schools. Secondly, support for budget and human resources should be expanded to achieve substantial reduction in administrative tasks. Thirdly, there is a need to establish a new mechanism for managing and disseminating workload reduction policies. 본 연구는 교원 업무경감 정책에 대한 중등학교 교원의 업무경감 정책에 대해 어떻게 인식하고 있는지를 알아보는 한편, 정부가 의도적인 목적을 얼마나 달성하는지에 대한 문제 인식에서 출발한다. 헤이룽장성의 중등학교 교원을 대상으로 교원 업무경감 정책에 대한 인식을 분석하여 실질적인 업무경감 정책 개선을 위한 시사점을 도출하기 위해 수행되었다. 이를 위해 선행연구에 조사된 중국 교원의 업무 수행 영역을 바탕으로 업무경감 정책의 내용을 고려하여, 배경 변인이 포함된 42개 문항을 제작하며, 헤이룽장성 지역에 근무 중인 교원을 대상으로 설문을 실시하였다. 조사 결과 회수된 응답자가 101명이며, 응답 시간은 100초 이하인 응답을 제외하고, 총 90부를 분석에 활용하였다. 위의 분석 자료를 토대로, 먼저 초⦁중등학교 교원들이 행정업무의 업무 수행 강도, 곤란도, 감축된 정도에 대한 인식에 대해 빈도, 백분율, 평균, 표준편차와 같은 기술통계 방법을 사용하였고, 초⦁중등학교 교원들이 행정업무의 업무 수행 강도, 곤란도, 감축된 정도에 대한 인식에 대한 인식의 차이를 살펴보기 위하여 집단 간의 차이 검증인 교차분석을 실시하였다. 마지막으로 초⦁중등학교 교원들이 행정업무의 업무 수행 강도, 곤란도, 감축된 정도에 대한 인식에 대한 인식의 차이를 알아보기 위해 t-test, One way ANOVA를 사용하였으며, 사후검증으로 Scheffé 방법을 사용하였다. 분석 결과, 첫째, 초⦁중학교 교사가 업무 수행 강도에 대한 인식 수준은 3.37로 나타났다. 따라서 교사가 업무 수행 강도가 높은 편이다. 둘째, 초⦁중학교 교사가 곤란도에 대한 인식 수준은 3.20으로 나타났다. 따라서 교사가 업무에 대한 곤란도가 높은 편이다. 셋째, 업무경감 정책을 실시한 후에 초⦁중학교 교사가 업무 감축된 정도에 대한 인식 수준은 3.75로 나타났다. 따라서 업무경감 정책을 실시한 후에 업무에 대해 감축된 효과가 미미하다. 이러한 연구결과를 통대로 첫째, 교육부, 교육청, 단위 학교에서 교원 업무에 대해 업무 표준화가 시급하다. 둘째, 실질적인 행정업무 경감을 위한 예산 및 인적자원에 대한 지원을 확대해야 한다. 셋째, 새로운 업무경감 정책 관리⦁전달 메커니즘을 마련해야 할 필요가 있다.