GPU Domain Specialization via Composable On-Package Architecture

As GPUs scale their low-precision matrix math throughput to boost deep learning (DL) performance, they upset the balance between math throughput and memory system capabilities. We demonstrate that a converged GPU design trying to address diverging architectural requirements between FP32 (or larger)-based HPC and FP16 (or smaller)-based DL workloads results in sub-optimal configurations for either of the application domains. We argue that a C omposable O n- PA ckage GPU (COPA-GPU) architecture to provide domain-specialized GPU products is the most practical solution to these diverging requirements. A COPA-GPU leverages multi-chip-module disaggregation to support maximal design reuse, along with memory system specialization per application domain. We show how a COPA-GPU enables DL-specialized products by modular augmentation of the baseline GPU architecture with up to 4× higher off-die bandwidth, 32× larger on-package cache, and 2.3× higher DRAM bandwidth and capacity, while conveniently supporting scaled-down HPC-oriented designs. This work explores the microarchitectural design necessary to enable composable GPUs and evaluates the benefits composability can provide to HPC, DL training, and DL inference. We show that when compared to a converged GPU design, a DL-optimized COPA-GPU featuring a combination of 16× larger cache capacity and 1.6× higher DRAM bandwidth scales per-GPU training and inference performance by 31% and 35%, respectively, and reduces the number of GPU instances by 50% in scale-out training scenarios.

Geometrical investigation of microchannel with two trapezoidal blocks subjected to laminar convective flows with and without boiling

Microchannels are important devices to improve the heat exchange in several engineering applications as heat, ventilation and air conditioning, microelectronic cooling, power generation systems and others. The present work performs a numerical study of a microchannel with two trapezoidal blocks subjected to laminar flows, aiming to analyze the influence of the boiling process on the geometric configuration of the microchannel. Constructal Design and Exhaustive Search are used for the geometrical evaluation of the blocks. The Mixture multi-phase model and the Lee phase change model were both employed for the numerical simulation of the boiling process. In this study, the influence of the height and higher width of the first block (H11/L11) over the heat transfer rate and pressure drop for different magnitudes of the ratio between the lower width and higher width (L12/L11) was investigated. It is considered water in monophase cases and water/vapor mixture for multiphase flow. Two different Reynolds numbers (ReH = 0.1 and 10.0) were investigated. Results indicated that, for the present thermal conditions, the consideration of boiling flows were not significant for prediction of optimal configurations. Results also showed that in the cases where the boiling process was enabled, the multi-objective performance was higher than in the cases without boiling, especially for ReH = 0.1.

Transfer Learning for Operator Selection: A Reinforcement Learning Approach

In the past two decades, metaheuristic optimization algorithms (MOAs) have been increasingly popular, particularly in logistic, science, and engineering problems. The fundamental characteristics of such algorithms are that they are dependent on a parameter or a strategy. Some online and offline strategies are employed in order to obtain optimal configurations of the algorithms. Adaptive operator selection is one of them, and it determines whether or not to update a strategy from the strategy pool during the search process. In the filed of machine learning, Reinforcement Learning (RL) refers to goal-oriented algorithms, which learn from the environment how to achieve a goal. On MOAs, reinforcement learning has been utilised to control the operator selection process. Existing research, however, fails to show that learned information may be transferred from one problem-solving procedure to another. The primary goal of the proposed research is to determine the impact of transfer learning on RL and MOAs. As a test problem, a set union knapsack problem with 30 separate benchmark problem instances is used. The results are statistically compared in depth. The learning process, according to the findings, improved the convergence speed while significantly reducing the CPU time.

Life Augmentation of Turbine Exhaust System Compensators Through Integrated MADM Optimization Approach of Stress Based Fatigue Cycles

Exhaust expansion joints, also known as compensators, are found in a variety of applications such as gas turbine exhaust pipes, generators, marine propulsion systems, OEM engines, power units, and auxiliary equipment. The motion compensators employed must have accomplished the maximum expansion-contraction cycle life while imposing the least amount of stress. Discrepancies in the selecting of bellows expansion joint design parameters are corrected by evaluating stress-based fatigue life, which is challenging owing to the complicated form of convolutions. Meridional and circumferential convolution stress equations that influencing fatigue cycles are evaluated and verified with FEA. Fractional factorial Taguchi L25 matrix is used for finding the optimal configurations. The discrete design parameters for the selection of the suitable configuration of the compensators are analysed with the help of the MADM decision making techniques. The multi-response optimization methods GRA, AHP, and TOPSIS are used to determine the parametric selection on a priority basis. It is seen that weighing distribution among the responses plays an important role in these methods and GRA method integrated with principal components shows best optimal configurations. Multiple regression technique applied to these methods also shows that PCA-GRA gives better alternate solutions for the designer unlike the AHP and TOPSIS method. However, higher ranked Taguchi run obtained in these methods may enhance the suitable selection of different design configurations. Obtained PCA-GRG values by Taguchi, Regression and DOE are well matched and verified for the all alternate solutions. Further, it also shows that stress based fatigue cycles obtained in this analysis for the L25 run indicates the range varying from 1.13 × 104 cycles to 9.08 × 105 cycles, which is within 106 cycles. This work will assist the design engineer for selecting the discrete parameters of stiff compensators utilized in power plant thermal appliances.

Fast and Low-cost Search for Efficient Cloud Configurations for HPC Workloads

The wide variety of virtual machine types, network configurations, number of instances, among others configuration tweaks, in cloud computing, makes the finding of the best configuration a hard problem. Trying to reduce costs and resource underutilization while achieving acceptable performance can be a hard task even for specialists. Thus, many approaches to find these optimal or almost optimal configurations for a given program were proposed in the literature. Observing the performance of an application in the cloud takes time and money. Therefore, most of the approaches aim not only to find good solutions but also to reduce the search cost. One of those approaches relies on Bayesian Optimization, which analyzes fewer configurations, reducing the search cost while still finding good solutions. Another approach found in the literature is the use of a technique named Paramount Iteration, which enables users to reason about cloud configurations' cost and performance without executing the application to its completion (early-stopping) this approach reduces the cost of each observation. In this work, we show that both techniques can be used together to do fewer and lower-cost observations. We demonstrate that such an approach can recommend solutions that are 1.68x better on average than Random Searching and with a 6x cheaper search.

Switchover phenomenon induced by epidemic seeding on geometric networks

It is a fundamental question in disease modeling how the initial seeding of an epidemic, spreading over a network, determines its final outcome. One important goal has been to find the seed configuration, which infects the most individuals. Although the identified optimal configurations give insight into how the initial state affects the outcome of an epidemic, they are unlikely to occur in real life. In this paper we identify two important seeding scenarios, both motivated by historical data, that reveal a complex phenomenon. In one scenario, the seeds are concentrated on the central nodes of a network, while in the second one, they are spread uniformly in the population. Comparing the final size of the epidemic started from these two initial conditions through data-driven and synthetic simulations on real and modeled geometric metapopulation networks, we find evidence for a switchover phenomenon: When the basic reproduction number R0 is close to its critical value, more individuals become infected in the first seeding scenario, but for larger values of R0, the second scenario is more dangerous. We find that the switchover phenomenon is amplified by the geometric nature of the underlying network and confirm our results via mathematically rigorous proofs, by mapping the network epidemic processes to bond percolation. Our results expand on the previous finding that, in the case of a single seed, the first scenario is always more dangerous and further our understanding of why the sizes of consecutive waves of a pandemic can differ even if their epidemic characters are similar.

CubeSat formation architecture for small space debris surveillance and orbit determination

Introduction: Satellites which face space debris cannot track it throughout the whole orbit due to natural limitations of their optical sensors, sush as field of view, Earth occultation, or solar illumination. Besides, the time of continuous observations is usually very short. Therefore, we are trying to offer the most effective configuration of optical sensors in order to provide short-arc tracking of a target piece of debris, using a scalable Extended Information Filter. Purpose: The best scenario for short-arc tracking of a space debris orbit using multipoint optical sensors. Results: We have found optimal configurations for groups of satellites with optical sensors which move along a sun-synchronous orbit. Debris orbit determination using an Extended Information Filter and measurements from multipoint sensors was simulated, and mean squared errors of the target's position were calculated. Based on the simulation results for variouos configurations, inter-satellite distances and measurement time, the most reliable scenario (four satellites in tetrahedral configuration) was found and recommended for practical use in short-arc debris tracking.

UDO

UDO is a versatile tool for offline tuning of database systems for specific workloads. UDO can consider a variety of tuning choices, reaching from picking transaction code variants over index selections up to database system parameter tuning. UDO uses reinforcement learning to converge to near-optimal configurations, creating and evaluating different configurations via actual query executions (instead of relying on simplifying cost models). To cater to different parameter types, UDO distinguishes heavy parameters (which are expensive to change, e.g. physical design parameters) from light parameters. Specifically for optimizing heavy parameters, UDO uses reinforcement learning algorithms that allow delaying the point at which the reward feedback becomes available. This gives us the freedom to optimize the point in time and the order in which different configurations are created and evaluated (by benchmarking a workload sample). UDO uses a cost-based planner to minimize reconfiguration overheads. For instance, it aims to amortize the creation of expensive data structures by consecutively evaluating configurations using them. We evaluate UDO on Postgres as well as MySQL and on TPC-H as well as TPC-C, optimizing a variety of light and heavy parameters concurrently.

Optimizing the Size of Autonomous Hybrid Microgrids with Regard to Load Shifting

The article proposes a method of multipurpose optimization of the size of an autonomous hybrid energy system consisting of photovoltaic, wind, diesel, and battery energy storage systems, and including a load-shifting system. The classical iterative Gauss–Seidel method was applied to optimize the size of a hybrid energy system in a remote settlement on Sakhalin Island. As a result of the optimization according to the minimum net present value criterion, several optimal configurations corresponding to different component combinations were obtained. Several optimal configurations were also found, subject to a payback period constraint of 5, 6, and 7 years. Optimizing the size of the hybrid power system with electric load shifting showed that the share of the load not covered by renewable energy sources decreases by 1.25% and 2.1%, depending on the parameters of the load shifting model. Net present cost and payback period also decreased, other technical and economic indicators improved; however, CO2 emissions increased due to the reduction in the energy storage system.