Generative AI poses both opportunities and risks for solving inverse design problems in the sciences. Generative tools provide the ability to expand and refine a search space autonomously, but do so at the cost of exploring low-quality regions until sufficiently fine tuned. Here, we propose a queue prioritization algorithm that combines generative modeling and active learning in the context of a distributed workflow for exploring complex design spaces. We find that incorporating an active learning model to prioritize top design candidates can prevent a generative AI workflow from expending resources on nonsensical candidates and halt potential generative model decay. For an existing generative AI workflow for discovering novel molecular structure candidates for carbon capture, our active learning approach significantly increases the number of high-quality candidates identified by the generative model. We find that, out of 1000 novel candidates, our workflow without active learning can generate an average of 281 high-performing candidates, while our proposed prioritization with active learning can generate an average 604 high-performing candidates.

Language models (LMs) can “memorize” information, i.e., encode training data in their weights in such a way that inference-time queries can lead to verbatim regurgitation of that data. This ability to extract training data can be problematic, for example, when data are private or sensitive. In this work, we investigate methods to mitigate memorization: three regularizer-based, three fine-tuning-based, and eleven machine unlearning-based methods, with five of the latter being new methods that we introduce. We also introduce TinyMem, a suite of small, computationally-efficient LMs for the rapid development and evaluation of memorization-mitigation methods. We demonstrate that the mitigation methods that we develop using TinyMem can successfully be applied to production-grade LMs, and we determine via experiment that: regularizer-based mitigation methods are slow and ineffective at curbing memorization; fine-tuning-based methods are effective at curbing memorization, but overly expensive, especially for retaining higher accuracies; and unlearning-based methods are faster and more effective, allowing for the precise localization and removal of memorized information from LM weights prior to inference. We show, in particular, that our proposed unlearning method BalancedSubnet outperforms other mitigation methods at removing memorized information while preserving performance on target tasks.

The high-performance computing (HPC) community has adopted incentive structures to motivate reproducible research, with major conferences awarding badges to papers that meet reproducibility requirements. Yet, many papers do not meet such requirements. The uniqueness of HPC infrastructure and software, coupled with strict access requirements, may limit opportunities for reproducibility. In the absence of resource access, we believe that regular documented testing, through continuous integration (CI), coupled with complete provenance information, can be used as a substitute. Here, we argue that better HPC-compliant CI solutions will improve reproducibility of applications. We present a survey of reproducibility initiatives and describe the barriers to reproducibility in HPC. To address existing limitations, we present a GitHub Action, CORRECT, that enables secure execution of tests on remote HPC resources. We evaluate CORRECT’s usability across three different types of HPC applications, demonstrating the effectiveness of using CORRECT for automating and documenting reproducibility evaluations.

Serverless computing has revolutionized application development and deployment by abstracting infrastructure management, allowing developers to focus on writing code. To do so, serverless platforms dynamically create execution environments, often using containers. The cost to create and deploy these environments is known as "cold start" latency, and this cost can be particularly detrimental to scientific computing workloads characterized by sporadic and dynamic demands. We investigate methods to mitigate cold start issues in scientific computing applications by pre-installing Python packages in container images. Using data from Globus Compute and Binder, we empirically analyze cold start behavior and evaluate four strategies for building containers, including fully pre-built environments and dynamic, on-demand installations. Our results show that pre-installing all packages reduces initial cold start time but requires significant storage. Conversely, dynamic installation offers lower storage requirements but incurs repetitive delays. Additionally, we implemented a simulator and assessed the impact of different warm times, finding that moderate warm times significantly reduce cold starts without the excessive overhead of maintaining always-hot states.

Task-based execution frameworks, such as parallel programming libraries, computational workflow systems, and function-as-a-service platforms, enable the composition of distinct tasks into a single, unified application designed to achieve a computational goal and abstract the parallel and distributed execution of those tasks on arbitrary hardware. Research into these task executors has accelerated as computational sciences increasingly need to take advantage of parallel compute and/or heterogeneous hardware. However, the lack of evaluation standards makes it challenging to compare and contrast novel systems against existing implementations. Here, we introduce TaPS, the Task Performance Suite, to support continued research in distributed task executor frameworks. TaPS provides (1) a unified, modular interface for writing and evaluating applications using arbitrary execution frameworks and data management systems and (2) an initial set of reference synthetic and real-world science applications. We discuss how the design of TaPS supports the reliable evaluation of frameworks and demonstrate TaPS through a survey of benchmarks using the provided reference applications.

As vehicles have become increasingly automated, novel vehicular applications have emerged to enhance the safety and security of the vehicles and improve user experience. This brings ever-increasing data and resource requirements for timely computation by the vehicle’s on-board computing systems. To meet these demands, prior work proposes deploying vehicular edge computing (VEC) resources in road-side units (RSUs) in the traffic infrastructure with which the vehicles can communicate and offload compute-intensive tasks. Due to the limited communication range of these RSUs, the communication link between the vehicles and the RSUs — and, therefore, the response times of the offloaded applications — are significantly impacted by vehicle mobility through road traffic. Existing task offloading strategies do not consider the influence of traffic lights on vehicular mobility while offloading workloads onto the RSUs. This causes deadline misses and quality-of-service (QoS) reduction for the offloaded tasks. In this article, we present a novel task model that captures time and location-specific requirements for vehicular applications. We then present a deadline-based strategy that incorporates traffic light data to opportunistically offload tasks. Our approach allows up to 33% more tasks to be offloaded onto RSUs compared with existing work without causing deadline misses, maximizing the resource utilization of RSUs.

Deep learning methods are transforming research, enabling new techniques, and ultimately leading to new discoveries. As the demand for more capable AI models continues to grow, we are now entering an era of Trillion Parameter Models (TPM), or models with more than a trillion parameters---such as Huawei's PanGu-Σ. We describe a vision for the ecosystem of TPM users and providers that caters to the specific needs of the scientific community. We then outline the significant technical challenges and open problems in system design for serving TPMs to enable scientific research and discovery. Specifically, we describe the requirements of a comprehensive software stack and interfaces to support the diverse and flexible requirements of researchers.

Advances in networks, accelerators, and cloud services encourage programmers to reconsider where to compute---such as when fast networks make it cost-effective to compute on remote accelerators despite added latency. Workflow and cloud-hosted serverless computing frameworks can manage multi-step computations spanning federated collections of cloud, high-performance computing (HPC), and edge systems, but passing data among computational steps via cloud storage can incur high costs. Here, we overcome this obstacle with a new programming paradigm that decouples control flow from data flow by extending the pass-by-reference model to distributed applications. We describe ProxyStore, a system that implements this paradigm by providing object proxies that act as wide-area object references with just-in-time resolution. This proxy model enables data producers to communicate data unilaterally, transparently, and efficiently to both local and remote consumers. We demonstrate the benefits of this model with synthetic benchmarks and real-world scientific applications, running across various computing platforms.

Mass data generation occurring in the Internet-of-Things (IoT) requires processing to extract meaningful information. Deep learning is commonly used to perform such processing. However, due to the sensitive nature of these data, it is important to consider data privacy. As such, federated learning (FL) has been proposed to address this issue. FL pushes training to the client devices and tasks a central server with aggregating collected model weights to update a global model. However, the transmission of these model weights can be costly, gradually. The trade-off between communicating model weights for aggregation and the loss provided by the global model remains an open problem. In this work, we cast this trade-off problem of client selection in FL as an optimization problem. We then design a Distributed Client Selection (DCS) algorithm that allows client devices to decide to participate in aggregation in hopes of minimizing overall communication cost — while maintaining low loss. We evaluate the performance of our proposed client selection algorithm against standard FL and a state-of-the-art client selection algorithm, called Power-of-Choice (PoC), using CIFAR-10, FMNIST, and MNIST datasets. Our experimental results confirm that our DCS algorithm is able to closely match the loss provided by the standard FL and PoC, while on average reducing the overall communication cost by nearly 32.67% and 44.71% in comparison to standard FL and PoC, respectively.

Federated learning (FL) is a technique for distributed machine learning that enables the use of siloed and distributed data. With FL, individual machine learning models are trained separately and then only model parameters (e.g., weights in a neural network) are shared and aggregated to create a global model, allowing data to remain in its original environment. While many applications can benefit from FL, existing frameworks are incomplete, cumbersome, and environment-dependent. To address these issues, we present FLoX, an FL framework built on the funcX federated serverless computing platform. FLoX decouples FL model training/inference from infrastructure management and thus enables users to easily deploy FL models on one or more remote computers with a single line of Python code. We evaluate FLoX using three benchmark datasets deployed on ten heterogeneous and distributed compute endpoints. We show that FLoX incurs minimal overhead, especially with respect to the large communication overheads between endpoints for data transfer. We show how balancing the number of samples and epochs with respect to the capacities of participating endpoints can significantly reduce training time with minimal reduction in accuracy. Finally, we show that global models consistently outperform any single model on average by 8%.

Traffic congestion is a costly phenomenon of every-day life. Reinforcement Learning (RL) is a promising solution due to its applicability to solving complex decision-making problems in highly dynamic environments. To train smart traffic lights using RL, large amounts of data is required. Recent RL-based approaches consider training to occur on some nearby server or a remote cloud server. However, this requires that traffic lights all communicate their raw data to some central location. For large road systems, communication cost can be impractical, particularly if traffic lights collect heavy data (e.g., video, LIDAR). As such, this work pushes training to the traffic lights directly to reduce communication cost. However, completely independent learning can reduce the performance of trained models. As such, this work considers the recent advent of Federated Reinforcement Learning (FedRL) for edge-enabled traffic lights so they can learn from each other's experience by periodically aggregating locally-learned policy network parameters rather than share raw data, hence keeping communication costs low. To do this, we propose the SEAL framework which uses an intersection-agnostic representation to support FedRL across traffic lights controlling heterogeneous intersection types. We then evaluate our FedRL approach against Centralized and Decentralized RL strategies. We compare the reward-communication trade-offs of these strategies. Our results show that FedRL is able to reduce the communication costs associated with Centralized training by 36.24%; while only seeing a 2.11 % decrease in average reward (i.e., decreased traffic congestion).

Recent advances in distributed data processing and machine learning provide new opportunities to enable critical, time-sensitive functionalities of smart distribution grids in a secure and reliable fashion. Combining the recent advents of edge computing (EC) and edge intelligence (EI) with existing advanced metering infrastructure (AMI) has the potential to reduce overall communication cost, preserve user privacy, and provide improved situational awareness. In this paper, we provide an overview for how EC and EI can supplement applications relevant to AMI systems. Additionally, using such systems in tandem can enable distributed deep learning frameworks (e.g., federated learning) to empower distributed data processing and intelligent decision making for AMI. Finally, to demonstrate the efficacy of this considered architecture, we approach the non-intrusive load monitoring (NILM) problem using federated learning to train a deep recurrent neural network architecture in a 2-tier and 3-tier manner. In this approach, smart homes locally train a neural network using their metering data and only share the learned model parameters with AMI components for aggregation. Our results show this can reduce communication cost associated with distributed learning, as well as provide an immediate layer of privacy, due to no raw data being communicated to AMI components. Further, we show that FL is able to closely match the model loss provided by standard centralized deep learning where raw data is communicated for centralized training.

Mobile edge computing pushes computationally-intensive services closer to the user to provide reduced delay due to physical proximity. This has led many to consider deploying deep learning models on the edge – commonly known as edge intelligence (EI). EI services can have many model implementations that provide different QoS. For instance, one model can perform inference faster than another (thus reducing latency) while achieving less accuracy when evaluated. In this paper, we study joint service placement and model scheduling of EI services with the goal to maximize Quality-of-Servcice (QoS) for end users where EI services have multiple implementations to serve user requests, each with varying costs and QoS benefits. We cast the problem as an integer linear program and prove that it is NP-hard. We then prove the objective is equivalent to maximizing a monotone increasing, submodular set function and thus can be solved greedily while maintaining a (1 – 1/e)-approximation guarantee. We then propose two greedy algorithms: one that theoretically guarantees this approximation and another that empirically matches its performance with greater efficiency. Finally, we thoroughly evaluate the proposed algorithm for making placement and scheduling decisions in both synthetic and real-world scenarios against the optimal solution and some baselines. In the real-world case, we consider real machine learning models using the ImageNet 2012 data-set for requests. Our numerical experiments empirically show that our more efficient greedy algorithm is able to approximate the optimal solution with a 0.904 approximation on average, while the next closest baseline achieves a 0.607 approximation on average.

Smart cities are a growing paradigm in the design of systems that interact with one another for informed and efficient decision making, empowered by data and technology, of resources in a city. The diffusion of information to citizens in a smart city will rely on social trends and smart advertisement. Online social networks (OSNs) are prominent and increasingly important platforms to spread information, observe social trends, and advertise new products. To maximize the benefits of such platforms in sharing information, many groups invest in finding ways to maximize the expected number of clicks as a proxy of these platform's performance. As such, the study of click-through rate (CTR) prediction of advertisements, in environments like online social media, is of much interest. Prior works build machine learning (ML) using user-specific data to classify whether a user will click on an advertisement or not. For our work, we consider a large set of Facebook advertisement data (with no user data) and categorize targeted interests into thematic groups we call conceptual nodes. ML models are trained using the advertisement data to perform CTR prediction with conceptual node combinations. We then cast the problem of finding the optimal combination of conceptual nodes as an optimization problem. Given a certain budget k, we are interested in finding the optimal combination of conceptual nodes that maximize the CTR. We discuss the hardness and possible NP-hardness of the optimization problem. Then, we propose a greedy algorithm and a genetic algorithm to find near-optimal combinations of conceptual nodes in polynomial time, with the genetic algorithm nearly matching the optimal solution. We observe that simple ML models can exhibit the high Pearson correlation coefficients w.r.t. click predictions and real click values. Additionally, we find that the conceptual nodes of “politics”, “celebrity”, and “organization” are notably more influential than other considered conceptual nodes.