Federated Learning (FL) is an emerging machine learning paradigm that enables the collaborative training of a shared global model across distributed clients while keeping the data decentralized. Recent works on designing systems for efficient FL have shown that utilizing serverless computing technologies, particularly Function-as-a-Service (FaaS) for FL, can enhance resource efficiency, reduce training costs, and alleviate the complex infrastructure management burden on data holders. However, current serverless FL systems still suffer from the presence of stragglers, i.e., slow clients that impede the collaborative training process. While strategies aimed at mitigating stragglers in these systems have been proposed, they overlook the diverse hardware resource configurations among FL clients. To this end, we present Apodotiko, a novel asynchronous training strategy designed for serverless FL. Our strategy incorporates a scoring mechanism that evaluates each client’s hardware capacity and dataset size to intelligently prioritize and select clients for each training round, thereby minimizing the effects of stragglers on system performance. We comprehensively evaluate Apodotiko across diverse datasets, considering a mix of CPU and GPU clients, and compare its performance against five other FL training strategies. Results from our experiments demonstrate that Apodotiko outperforms other FL training strategies, achieving an average speedup of 2.75x and a maximum speedup of 7.03x. Furthermore, our strategy significantly reduces cold starts by a factor of four on average, demonstrating suitability in serverless environments.

With the advent of AWS Lambda in 2014, Serverless Computing, particularly Function-as-a-Service (FaaS), has witnessed growing popularity across various application domains. FaaS enables an application to be decomposed into fine-grained functions that are executed on a FaaS platform. It offers several advantages such as no infrastructure management, a pay-per-use billing policy, and on-demand fine-grained autoscaling. However, despite its advantages, developers today encounter various challenges while adopting FaaS solutions that reduce productivity. These include FaaS platform lock-in, support for diverse function deployment parameters, and diverse interfaces for interacting with FaaS platforms. To address these challenges, we present gFaaS, a novel framework that facilitates the holistic development and management of functions across diverse FaaS platforms. Our framework enables the development of generic functions in multiple programming languages that can be seamlessly deployed across different platforms without modifications. Results from our experiments demonstrate that gFaaS functions perform similarly to native platform-specific functions across various scenarios. A video demonstrating the functioning of gFaaS is available from https://youtu.be/STbb6ykJFf0.

Serverless computing (FaaS) has been extensively utilized for deep learning (DL) inference due to the ease of deployment and pay-per-use benefits. However, existing FaaS platforms utilize GPUs in a coarse manner for DL inferences, without taking into account spatio-temporal resource multiplexing and isolation, which results in severe GPU under-utilization, high usage expenses, and SLO (Service Level Objectives) violation. There is an imperative need to enable an efficient and SLO-aware GPU-sharing mechanism in serverless computing to facilitate cost-effective DL inferences. In this paper, we propose FaST-GShare, an efficient FaaS-oriented Spatio-Temporal GPU Sharing architecture for deep learning inferences. In the architecture, we introduce the FaST-Manager to limit and isolate spatio-temporal resources for GPU multiplexing. In order to realize function performance, the automatic and flexible FaST-Profiler is proposed to profile function throughput under various resource allocations. Based on the profiling data and the isolation mechanism, we introduce the FaST-Scheduler with heuristic auto-scaling and efficient resource allocation to guarantee function SLOs. Meanwhile, FaST-Scheduler schedules function with efficient GPU node selection to maximize GPU usage. Furthermore, model sharing is exploited to mitigate memory contention. Our prototype implementation on the OpenFaaS platform and experiments on MLPerf-based benchmark prove that FaST-GShare can ensure resource isolation and function SLOs. Compared to the time sharing mechanism, FaST-GShare can improve throughput by 3.15x, GPU utilization by 1.34x, and SM (Streaming Multiprocessor) occupancy by 3.13x on average.

Containerization approaches based on namespaces offered by the Linux kernel have seen an increasing popularity in the HPC community both as a means to isolate applications and as a format to package and distribute them. However, their adoption and usage in HPC systems faces several challenges. These include difficulties in unprivileged running and building of scientific application container images directly on HPC resources, increasing heterogeneity of HPC architectures, and access to specialized networking libraries available only on HPC systems. These challenges of container-based HPC application development closely align with the several advantages that a new universal intermediate binary format called WebAssembly (Wasm) has to offer. These include a lightweight userspace isolation mechanism and portability across operating systems and processor architectures. In this paper, we explore the usage of Wasm as a distribution format for MPI-based HPC applications. To this end, we present MPIWasm, a novel Wasm embedder for MPI-based HPC applications that enables high-performance execution of Wasm code, has low-overhead for MPI calls, and supports high-performance networking interconnects present on HPC systems. We evaluate the performance and overhead of MPIWasm on a production HPC system and AWS Graviton2 nodes using standardized HPC benchmarks. Results from our experiments demonstrate that MPIWasm delivers competitive native application performance across all scenarios. Moreover, we observe that Wasm binaries are 139.5x smaller on average as compared to the statically-linked binaries for the different standardized benchmarks.

Serverless computing paradigm has become more ingrained into the industry, as it offers a cheap alternative for application development and deployment. This new paradigm has also created new kinds of problems for the developer, who needs to tune memory configurations for balancing cost and performance. Many researchers have addressed the issue of minimizing cost and meeting Service Level Objective (SLO) requirements for a single FaaS function, but there has been a gap for solving the same problem for an application consisting of many FaaS functions, creating complex application workflows.In this work, we designed a tool called SLAM to address the issue. SLAM uses distributed tracing to detect the relationship among the FaaS functions within a serverless application. By modeling each of them, it estimates the execution time for the application at different memory configurations. Using these estimations, SLAM determines the optimal memory configuration for the given serverless application based on the specified SLO requirements and user-specified objectives (minimum cost or minimum execution time). We demonstrate the functionality of SLAM on AWS Lambda by testing on four applications. Our results show that the suggested memory configurations guarantee that more than 95% of requests are completed within the predefined SLOs.

Federated Learning (FL) is a machine learning paradigm that enables the training of a shared global model across distributed clients while keeping the training data local. While most prior work on designing systems for FL has focused on using stateful always running components, recent work has shown that components in an FL system can greatly benefit from the usage of serverless computing and Function-as-a-Service technologies. To this end, distributed training of models with severless FL systems can be more resource-efficient and cheaper than conventional FL systems. However, serverless FL systems still suffer from the presence of stragglers, i.e., slow clients due to their resource and statistical heterogeneity. While several strategies have been proposed for mitigating stragglers in FL, most methodologies do not account for the particular characteristics of serverless environments, i.e., cold-starts, performance variations, and the ephemeral stateless nature of the function instances. Towards this, we propose FedLesScan, a novel clustering-based semi-asynchronous training strategy, specifically tailored for serverless F L. FedLesScan dynamically adapts to the behavior of clients and minimizes the effect of stragglers on the overall system. We implement our strategy by extending an open-source serverless FL system called FedLess. Moreover, we comprehensively evaluate our strategy using the 2ndgeneration Google Cloud Functions with four datasets and varying percentages of stragglers. Results from our experiments show that compared to other approaches FedLesScan reduces training time and cost by an average of 8% and 20% respectively while utilizing clients better with an average increase in the effective update ratio of 17.75%.

In the recent past, characterizing workloads has been attempted to gain a foothold in the emerging serverless cloud market, especially in the large production cloud clusters of Google, AWS, and so forth. While analyzing and characterizing real workloads from a large production cloud cluster benefits cloud providers, researchers, and daily users, analyzing the workload traces of these clusters has been an arduous task due to the heterogeneous nature of data. This article proposes a scalable infrastructure based on Google's dataproc for analyzing the workload traces of cloud environments. We evaluated the functioning of the proposed infrastructure using the workload traces of Google cloud cluster-usage-traces-v3. We perform the workload characterization on this dataset, focusing on the heterogeneity of the workload, the variations in job durations, aspects of resources consumption, and the overall availability of resources provided by the cluster. The findings reported in the paper will be beneficial for cloud infrastructure providers and users while managing the cloud computing resources, especially serverless platforms.

Function-as-a-Service (FaaS) is an attractive cloud computing model that simplifies application development and deployment. However, current serverless compute platforms do not consider data placement when scheduling functions. With the growing demand for edge-cloud continuum, multi-cloud, and multi-serverless applications, this flaw means serverless technologies are still ill-suited to latency-sensitive operations like media streaming. This work proposes a solution by presenting a tool called FaDO: FaaS Functions and Data Orchestrator, designed to allow data-aware functions scheduling across multi-serverless compute clusters present at different locations, such as at the edge and in the cloud. FaDO works through header-based HTTP reverse proxying and uses three load-balancing algorithms: 1) The Least Connections, 2) Round Robin, and 3) Random for load balancing the invocations of the function across the suitable serverless compute clusters based on the set storage policies. FaDO further provides users with an abstraction of the serverless compute cluster’s storage, allowing users to interact with data across different storage services through a unified interface. In addition, users can configure automatic and policy-aware granular data replications, causing FaDO to spread data across the clusters while respecting location constraints. Load testing results show that it is capable of load balancing high-throughput workloads, placing functions near their data without contributing any significant performance overhead.

Microservice architecture is the common choice for cloud applications these days since each individual microservice can be independently modified, replaced, and scaled. However, the complexity of microservice applications requires automated testing with a focus on the interactions between the services. While this is achievable with end-to-end tests, they are error-prone, brittle, expensive to write, time-consuming to run, and require the entire application to be deployed. Integration tests are an alternative to end-to-end tests since they have a smaller test scope and require the deployment of a significantly fewer number of services. The de-facto standard for deploying microservice applications in the cloud is containers with Kubernetes being the most widely used container orchestration platform. To support the integration testing of microservices in Kubernetes, several tools such as Octopus, Istio, and Jenkins exist. However, each of these tools either lack crucial functionality or lead to a substantial increase in the complexity and growth of the tool landscape when introduced into a project. To this end, we present Bunk8s, a tool for integration testing of microservice applications in Kubernetes that overcomes the limitations of these existing tools. Bunk8s is independent of the test framework used for writing integration tests, independent of the used CI/CD infrastructure, and supports test result publishing. A video demonstrating the functioning of our tool is available from https://www.youtube.com/watch?v=e8wbS25O4Bo.

FaaS allows an application to be decomposed into functions that are executed on a FaaS platform. The FaaS platform is responsible for the resource provisioning of the functions. Recently, there is a growing trend towards the execution of compute-intensive FaaS functions that run for several seconds. However, due to the billing policies followed by commercial FaaS offerings, the execution of these functions can incur significantly higher costs. Moreover, due to the abstraction of underlying processor architectures on which the functions are executed, the performance optimization of these functions is challenging. As a result, most FaaS functions use pre-compiled libraries generic to x86-64 leading to performance degradation. In this paper, we examine the underlying processor architectures for Google Cloud Functions (GCF) and determine their prevalence across the 19 available GCF regions. We modify, adapt, and optimize three compute-intensive FaaS workloads written in Python using Numba, a JIT compiler based on LLVM, and present results wrt performance, memory consumption, and costs on GCF. Results from our experiments show that the optimization of FaaS functions can improve performance by 12.8x (geometric mean) and save costs by 73.4% on average for the three functions. Our results show that optimization of the FaaS functions for the specific architecture is very important. We achieved a maximum speedup of 1.79x by tuning the function especially for the instruction set of the underlying processor architecture.

The traditional cloud-centric approach for Deep Learning (DL) requires training data to be collected and processed at a central server which is often challenging in privacy-sensitive domains like healthcare. Towards this, a new learning paradigm called Federated Learning (FL) has been proposed that brings the potential of DL to these domains while addressing privacy and data ownership issues. FL enables clients to learn a shared ML model while keeping the data local. However, conventional FL systems face challenges such as scalability, complex infrastructure management, and wasted compute and incurred costs due to idle clients. These challenges of FL systems closely align with the core problems that serverless computing and Function-as-a-Service (FaaS) platforms aim to solve. These include rapid scalability, no infrastructure management, automatic scaling to zero for idle clients, and a pay-per-use billing model. To this end, we present a novel system and framework for serverless FL, called FedLess. Our system supports multiple commercial and self-hosted FaaS providers and can be deployed in the cloud, on-premise in institutional data centers, and on edge devices. To the best of our knowledge, we are the first to enable FL across a large fabric of heterogeneous FaaS providers while providing important features like security and Differential Privacy. We demonstrate with comprehensive experiments that the successful training of DNNs for different tasks across up to 200 client functions and more is easily possible using our system. Furthermore, we demonstrate the practical viability of our methodology by comparing it against a traditional FL system and show that it can be cheaper and more resource-efficient.

EdgeAI (Edge computing based Artificial Intelligence) has been most actively researched for the last few years to handle variety of massively distributed AI applications to meet up the strict latency requirements. Meanwhile, many companies have released edge devices with smaller form factors (low power consumption and limited resources) like the popular Raspberry Pi and Nvidia's Jetson Nano for acting as compute nodes at the edge computing environments. Although the edge devices are limited in terms of computing power and hardware resources, they are powered by accelerators to enhance their performance behavior. Therefore, it is interesting to see how AI-based Deep Neural Networks perform on such devices with limited resources. In this work, we present and compare the performance in terms of inference time and power consumption of the four SoCs: Asus Tinker Edge R, Raspberry Pi 4, Google Coral Dev Board, Nvidia Jetson Nano, and one microcontroller: Arduino Nano 33 BLE, on different deep learning models and frameworks. We also provide a method for measuring power consumption, inference time and accuracy for the devices, which can be easily extended to other devices. Our results showcase that, for Tensorflow based quantized model, the Google Coral Dev Board delivers the best performance, both for inference time and power consumption. For a low fraction of inference computation time, i.e. less than 29.3% of the time for MobileNetV2, the Jetson Nano performs faster than the other devices.

Several of today's cloud applications are spread over heterogeneous connected computing resources and are highly dynamic in their structure and resource requirements. However, serverless computing and Function-as-a-Service (FaaS) platforms are limited to homogeneous clusters and homogeneous functions. We introduce an extension of FaaS to heterogeneous computing and to support heterogeneous functions through a network of distributed heterogeneous target platforms called Function Delivery Network (FDN). A target platform is a combination of a cluster of a homogeneous computing system and a FaaS platform on top of it. FDN provides Function-Delivery-as-a-Service (FDaaS), delivering the function invocations to the right target platform. We showcase the opportunities such as collaborative execution between multiple target platforms and varied target platform's characteristics that the FDN offers in fulfilling two objectives: Service Level Objective (SLO) requirements and energy efficiency when scheduling functions invocations by evaluating over five distributed target platforms.

Summary Serverless computing has rapidly grown following the launch of Amazon's Lambda platform. Function‐as‐a‐Service (FaaS) a key enabler of serverless computing allows an application to be decomposed into simple, standalone functions that are executed on a FaaS platform. The FaaS platform is responsible for deploying and facilitating resources to the functions. Several of today's cloud applications spread over heterogeneous connected computing resources and are highly dynamic in their structure and resource requirements. However, FaaS platforms are limited to homogeneous clusters and homogeneous functions and do not account for the data access behavior of functions before scheduling. We introduce an extension of FaaS to heterogeneous clusters and to support heterogeneous functions through a network of distributed heterogeneous target platforms called Function Delivery Network (FDN). A target platform is a combination of a cluster of homogeneous nodes and a FaaS platform on top of it. FDN provides Function‐Delivery‐as‐a‐Service (FDaaS), delivering the function to the right target platform. We showcase the opportunities such as varied target platform's characteristics, possibility of collaborative execution between multiple target platforms, and localization of data that the FDN offers in fulfilling two objectives: Service Level Objective (SLO) requirements and energy efficiency when scheduling functions by evaluating over five distributed target platforms using the FDNInspector, a tool developed by us for benchmarking distributed target platforms. Scheduling functions on an edge target platform in our evaluation reduced the overall energy consumption by 17× without violating the SLO requirements in comparison to scheduling on a high‐end target platform.

With the growing constraints on power budget and increasing hardware failure rates, the operation of future exascale systems faces several challenges. Towards this, resource awareness and adaptivity by enabling malleable jobs has been actively researched in the HPC community. Malleable jobs can change their computing resources at runtime and can significantly improve HPC system performance. However, due to the rigid nature of popular parallel programming paradigms such as MPI and lack of support for dynamic resource management in batch systems, malleable jobs have been largely unrealized. In this paper, we extend the SLURM batch system to support the execution and batch scheduling of malleable jobs. The malleable applications are written using a new adaptive parallel paradigm called Invasive MPI which extends the MPI standard to support resource-adaptivity at runtime. We propose two malleable job scheduling strategies to support performance-aware and power-aware dynamic reconfiguration decisions at runtime. We implement the strategies in SLURM and evaluate them on a production HPC system. Results for our performance-aware scheduling strategy show improvements in makespan, average system utilization, average response, and waiting times as compared to other scheduling strategies. Moreover, we demonstrate dynamic power corridor management using our power-aware strategy.

Energy efliciency and energy conservation are one of the most crucial constraints for meeting the 20MW power envelope desired for exascale systems. Towards this, most of the research in this area has been focused on the utilization of user-controllable hardware switches such as per-core dynamic voltage frequency scaling (DVFS) and software controlled clock modulation at the application level. In this paper, we present a tuning plugin for the Periscope Tuning Framework which integrates line-grained autotuning at the region level with DVFS and uncore frequency scaling (UFS). The tuning is based on a feed-forward neural network which is formulated using Performance Monitoring Counters (PMC) supported by x86 systems and trained using standardized benchmarks. Experiments on live standardized hybrid benchmarks show an energy improvement of 16.1% on average when the applications are tuned according to our methodology as compared to 7.8% for static tuning.

Power and energy optimization of applications running on a High Performance Computing infrastructure is an important research area in today's HPC driven world. Considerable amount of investments are being made for the same, emphasizing on the importance of such a research. The above mentioned optimization is a two step process. First, one must measure and analyze the power consumption accurately and then optimize the application by modifying the relevant sections of the application software responsible for it. In this paper we propose an API, called the Unified Power Profiling API (UPPAPI), which allows a parallel application developer to measure the power and energy consumed by the heterogeneous system during the application execution. The first important feature of this API is the abstraction created over different co-processors. Second, we have incorporated a corrective power model to remove the discrepancy arising out of the on-board GPU sensors. We illustrate the working of the above mentioned tool using Rodinia and SHOC Benchmarks. We further verify the correctness of the API generated measurement by comparing the results returned by the API with the actual power measured with a Krykard ALM10 power analyzer. This paper is an early attempt to develop a well rounded software tool, which enables the developer to optimize the resource usage and create efficient applications.