Cost Efficient GPU Cluster Management for Training and Inference of Deep Learning

Expanding the scale of GPU-based deep learning (DL) clusters would bring not only accelerated AI services but also significant energy consumption costs. In this paper, we propose a cost efficient deep learning job allocation (CE-DLA) approach minimizing the energy consumption cost for the DL cluster operation while guaranteeing the performance requirements of user requests. To do this, we first categorize the DL jobs into two classes: training jobs and inference jobs. Through the architecture-agnostic modeling, our CE-DLA approach is able to conduct the delicate mapping of heterogeneous DL jobs to GPU computing nodes. Second, we design the electricity price-aware DL job allocation so as to minimize the energy consumption cost of the cluster. We show that our approach efficiently avoids the peak-rate time slots of the GPU computing nodes by using the sophisticated mixed-integer nonlinear problem (MINLP) formulation. We additionally integrate the dynamic right-sizing (DRS) method with our CE-DLA approach, so as to minimize the energy consumption of idle nodes having no running job. In order to investigate the realistic behavior of our approach, we measure the actual output from the NVIDIA-based GPU devices with well-known deep neural network (DNN) models. Given the real trace data of the electricity price, we show that the CE-DLA approach outperforms the competitors in views of both the energy consumption cost and the performance for DL job processing.

Improving the Accuracy of Estimates of Indoor Distance Moved Using Deep Learning-Based Movement Status Recognition

As a result of the development of wireless indoor positioning techniques such as WiFi, Bluetooth, and Ultra-wideband (UWB), the positioning traces of moving people or objects in indoor environments can be tracked and recorded, and the distances moved can be estimated from these data traces. These estimates are very useful in many applications such as workload statistics and optimized job allocation in the field of logistics. However, due to the uncertainties of the wireless signal and corresponding positioning errors, accurately estimating movement distance still faces challenges. To address this issue, this paper proposes a movement status recognition-based distance estimating method to improve the accuracy. We divide the positioning traces into segments and use an encoder–decoder deep learning-based model to determine the motion status of each segment. Then, the distances of these segments are calculated by different distance estimating methods based on their movement statuses. The experiments on the real positioning traces demonstrate the proposed method can precisely identify the movement status and significantly improve the distance estimating accuracy.

Analysis of Security Based Accessibility Control Models for Cloud Computing

Access control has become the most necessary requirement to limit unauthorized and privileged access to information systems in cloud computing. Access control models counter the additional security challenges like rules, domain names, job allocation, multi hosting and separation of tasks. This paper classifies the conventional and modern access control models which has been utilized to restrain these access flaws by employing a variety of practices and methodologies. It examine the frequent security threats to information confidentiality, integrity, data accessibility and their approach used for cloud solutions. This paper proposed a priority based task scheduling access control (PbTAC) model to secure and scheduled access of resources & services rendered to cloud user. PbTAC model will ensure the job allocation, tasks scheduling and security of information through its rule policies during transmission between parties. It also help in reducing system overhead by minimize the computation and less storage cost.

MO1038JOB BURNOUT IN RARE DISEASE RESEARCH PHYSICIANS DURING COVID-19 PANDEMIC

Background and Aims Doctors are exposed to high levels of stress in their profession and are particularly susceptible to experiencing burnout. Rare disease researches are enlightening, with more workload to clinicians, especially during the Covid-19 pandemic. We aim to explore the mental influence of participating in rare disease researches on clinicians. Method Doctors received electronic questionnaires regarding job-burnout in October 2020. The modified Maslach Burnout Inventory-General Survey (MBI-GS) was used to evaluate job burnout state. The MBI-GS consisting of three dimensions, emotional exhaustion (five questions), cynicism (five questions), and reduced personal accomplishment (six questions). The 7-grade Likert scale is adopted in each question, from 0 point (never) to 6 points (very frequently). Job burnout was considered if the average score of any dimension is no less than three. Results Questionnaires from all 203 doctors were analysed in this study, with females (70.0%, n=140). Age ranging from 25 to 39, 40 to 54, and above 55 were 41.4%, 50.7%, 7.9%, respectively. Nearly half of the subjects (50.2%, n=102) fulfil the definition of job-burnout, which was fewer than that in the residency program (50.2% vs. 62.9%, p=0.02). An inappropriate evaluation system (36.0%) and lack of private time (35.5%) were the leading cause of job-burnout. The pressure of scientific researches (79.3%) and career promotion (58.1%) was the major source of mental pressure. Doctors who participated in rare disease researches (46.8%, n=95) did not show significant differences in burnout rate than individuals who did not (44.2% vs 55.6%, p=0.123), nor as in three dimensions (27.3% vs 36.1%, p=0.183 for emotional exhaustion, 21.1% vs 20.4%, p=0.905 for cynicism, 21.1% vs 27.8%, p=0.267 for reduced personal accomplishment). Logistic analysis revealed that high requirement from superior (22.5% vs 6.9%, p= 0.001), pressure from family (33.3% vs 17.8%, p=0.010), inappropriate job allocation (47.1% vs 29.7%, p= 0.019) as well as delayed off-work time (p=0.013) were independent risk factors of job-burnout. Physicians who participate in rare disease research had better job allocation (75.8% vs. 49.1%, p<0.001), but not in the other three risk factors. Conclusion More workload did not increase the job-burnout of physicians participating in the rare disease research, which might be contributed by the appropriate job allocation.

Optimal Stopping in the Balls-and-Bins Problem

This paper considers a multistage balls-and-bins problem with optimal stopping connected with the job allocation model. There are N steps. The player drops balls (tasks) randomly one at a time into available bins (servers). The game begins with only one empty bin. At each step, a new bin can appear with probability p. At step n (n = 1, . . . ,N), the player can choose to stop and receive the payoff or continue the process and move to the next step. If the player stops, then he/she gets 1 for every bin with exactly one ball and loses 1/2 for every bin with two or more balls. Empty bins do not count. At the last step, the player must stop the process. The player's aim is to find the stopping rule which maximizes the expected payoff. The optimal payoff at each step are calculated. An approximate strategy depending on the number of steps is proposed. It is demonstrated that the payo when using this strategy is close to the optimal payoff.