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.

Calculation of AeroMACS Spectrum Requirements Based on Traffic Simulator

In this paper, we propose a methodology for calculating the necessary spectrum requirements of aeronautical mobile airport communication system (AeroMACS) to provide various airport communication services. To accurately calculate the spectrum requirement, it is necessary to evaluate the AeroMACS traffic demand of the peak time and statistical data on the packet traffic generated at the airport. Because there is no AeroMACS traffic model and real trace data, we have developed the AeroMACS traffic simulator based on the report of Single European Sky Air Traffic Management Research (SESAR). To calculate the spectrum requirements, the AeroMACS traffic simulator is combined with the methodology of ITU-R M.1768-1. The developed traffic simulator reflects AeroMACS traffic priorities and can generate the required traffic according to its location in the airport. We observed the spectrum requirement by changing the number of sectors and the spectral efficiency. To show the feasibility of our methodology, we applied it to the case of Incheon International Airport in Korea. The simulation results show that the average bandwidth of 0.94 MHz is required in the ground area and 8.59 MHz is required in the entire airport.

CC2: Defending Hybrid Worm on Mobile Networks with Two-Dimensional Circulation Control

As the hybrid worm can propagate by both personal social interactions and wireless communications, it has been identified as one of the most severe threats to the mobile Internet. This problem is expected to become worse with the boom of social applications and mobile services. In this work, we study the propagation dynamics of hybrid worms and propose a systematic countermeasure. The system maintains a set of community structure which describes the high-speed infection zone of worms and contains worm propagation by distributing the worm signature to the guard nodes selected from the periphery of each community. For those nodes that are geographically close but located in different communities , we evaluate the communication security between them based on the observed infection history and limit communications between insecure ones to avoid the worm spreading across communities. We also design an efficient worm signature forwarding strategy that enables most nodes in the network to reach an immune state before being infected by the worm. Extensive real-trace driven simulations verify the feasibility and effectiveness of the proposed methods.

Simulation Speedup in OMNeT++ Using Contact Traces

The IoT influence is growing rapidly and it is expected that in the near future the number of connected IoT nodes will be in billions. Mobility of nodes is a key aspect in the IoT and network simulations are used to evaluate the performance of IoT networks with mobility. The increasing number of nodes will affect simulation environments by in- creasing the number of nodes per simulation and these simulations will require a long time to complete. So, techniques to reduce the time consumed in simulations are important. A possible technique is the use of contact traces when simulating mobility. A contact trace differs from a real trace or a synthetic mobility model in that, it deals with contact information instead of coordinate information. In this work, we have compared the per- formance of contact traces against coordinate based real traces and mobility models. For our evaluations we use the OPS framework of models built in OMNeT++. In the contact based approach, the use of real traces or mobility models will be removed and instead, nodes will use contact traces to obtain information about connections with other nodes. But this approach requires an additional step, before a simulation, to generate the contact traces. Even with this step, we show that using contact traces is advantageous in terms of simulation durations. When the scale of simulations increase (i.e., in terms of number of nodes, simulated times, etc.), the results presented show that the advantage increases considerably.

Estimating hotspots using a Gaussian mixture model from large-scale taxi GPS trace data

The data collected from taxi vehicles using the global positioning system (GPS) traces provides abundant temporal-spatial information, as well as information on the activity of drivers. Using taxi vehicles as mobile sensors in road networks to collect traffic information is an important emerging approach in efforts to relieve congestion. In this paper, we present a hybrid model for estimating driving paths using a density-based spatial clustering of applications with noise (DBSCAN) algorithm and a Gaussian mixture model (GMM). The first step in our approach is to extract the locations from pick-up and drop-off records (PDR) in taxi GPS equipment. Second, the locations are classified into different clusters using DBSCAN. Two parameters (density threshold and radius) are optimized using real trace data recorded from 1100 drivers. A GMM is also utilized to estimate a significant number of locations; the parameters of the GMM are optimized using an expectation-maximum (EM) likelihood algorithm. Finally, applications are used to test the effectiveness of the proposed model. In these applications, locations distributed in two regions (a residential district and a railway station) are clustered and estimated automatically.

A two-dimensional slice through the parameter space of two-generator Kleinian groups

We describe all real points of the parameter space of two-generator Kleinian groups with a parabolic generator, that is, we describe a certain two-dimensional slice through this space. In order to do this, we gather together known discreteness criteria for two-generator groups with a parabolic generator and present them in the form of conditions on parameters. We complete the description by giving discreteness criteria for groups generated by a parabolic and a [Formula: see text]-loxodromic elements whose commutator has real trace and present all orbifolds uniformized by such groups.

OFFLOADING LTE DATA TRAFFIC WITH NAMED-DATA NETWORKING INTEGRATION

Nowadays, mobile Internet becomes increasingly popular and the number of mobile users is growing exponentially. With a traditional clients-server connection, servers are usually in overload state by a huge number of users accessing the service at the same time. Moreover, in a century of “green” Internet technology, it should be more effective in content distribution for a huge number of users. Named-Data Networking (NDN) had been proposed as the promising solution for above problem, which is a content name-oriented approach to disseminate content to edge gateways/routers. In NDN, popular contents are cached at routers for a certain time, and the previously queried one can be reused for multiple times to save bandwidth. In this paper, we propose a solution for Long Term Evolution (LTE) mobile network based on the concept of NDN. By OPNET Modeler simulation, we carry out the evaluation in realistic mobile network with a huge number of mobility LTE mobile stations access to a single server, where content names and content sizes are obtained from real trace of Internet traffic. The obtained results show that Evolved Packet Core (EPC) caching scheme can helps to further increases the quality of service as well as offloads server traffic significantly.

OFFLOADING LTE DATA TRAFFIC WITH NAMED-DATA NETWORKING INTEGRATION

Nowadays, mobile Internet becomes increasingly popular and the number of mobile users is growing exponentially. With a traditional clients-server connection, servers are usually in overload state by a huge number of users accessing the service at the same time. Moreover, in a century of “green” Internet technology, it should be more effective in content distribution for a huge number of users. Named-Data Networking (NDN) had been proposed as the promising solution for above problem, which is a content name-oriented approach to disseminate content to edge gateways/routers. In NDN, popular contents are cached at routers for a certain time, and the previously queried one can be reused for multiple times to save bandwidth. In this paper, we propose a solution for Long Term Evolution (LTE) mobile network based on the concept of NDN. By OPNET Modeler simulation, we carry out the evaluation in realistic mobile network with a huge number of mobility LTE mobile stations access to a single server, where content names and content sizes are obtained from real trace of Internet traffic. The obtained results show that Evolved Packet Core (EPC) caching scheme can helps to further increases the quality of service as well as offloads server traffic significantly.