BlastFunction: A Full-stack Framework Bringing FPGA Hardware Acceleration to Cloud-native Applications

“Cloud-native” is the umbrella adjective describing the standard approach for developing applications that exploit cloud infrastructures’ scalability and elasticity at their best. As the application complexity and user-bases grow, designing for performance becomes a first-class engineering concern. As an answer to these needs, heterogeneous computing platforms gained widespread attention as powerful tools to continue meeting SLAs for compute-intensive cloud-native workloads. We propose BlastFunction, an FPGA-as-a-Service full-stack framework to ease FPGAs’ adoption for cloud-native workloads, integrating with the vast spectrum of fundamental cloud models. At the IaaS level, BlastFunction time-shares FPGA-based accelerators to provide multi-tenant access to accelerated resources without any code rewriting. At the PaaS level, BlastFunction accelerates functionalities leveraging the serverless model and scales functions proactively, depending on the workload’s performance. Further lowering the FPGAs’ adoption barrier, an accelerators’ registry hosts accelerated functions ready to be used within cloud-native applications, bringing the simplicity of a SaaS-like approach to the developers. After an extensive experimental campaign against state-of-the-art cloud scenarios, we show how BlastFunction leads to higher performance metrics (utilization and throughput) against native execution, with minimal latency and overhead differences. Moreover, the scaling scheme we propose outperforms the main serverless autoscaling algorithms in workload performance and scaling operation amount.

Lattice Structure of Some Closed Classes for Three-Valued Logic and Its Applications

This paper provides a brief overview of modern applications of nonbinary logic models, where the design of heterogeneous computing systems with small computing units based on three-valued logic produces a mathematically better and more effective solution compared to binary models. For application, it is necessary to implement circuits composed of chipsets, the operation of which is based on three-valued logic. To be able to implement such schemes, a fundamentally important theoretical problem must be solved: the problem of completeness of classes of functions of three-valued logic. From a practical point of view, the completeness of the class of such functions ensures that circuits with the desired operations can be produced from an arbitrary (finite) set of chipsets. In this paper, the closure operator on the set of functions of three-valued logic that strengthens the usual substitution operator is considered. It is shown that it is possible to recover the sublattice of closed classes in the general case of closure of functions with respect to the classical superposition operator. The problem of the lattice of closed classes for the class of functions T2 preserving two is considered. The closure operators R1 for the functions that differ only by dummy variables are considered equivalent. This operator is withiin the scope of interest of this paper. A lattice is constructed for closed subclasses in T2={f|f(2,…,2)=2}, a class of functions preserving two.

List Scheduling Algorithm Based on Virtual Scheduling Length Table in Heterogeneous Computing System

Edge computing needs the close cooperation of cloud computing to better meet various needs. Therefore, ensuring the efficient implementation of applications in cloud computing is not only related to the development of cloud computing itself but also affects the promotion of edge computing. However, resource management and task scheduling strategy are important factors affecting the efficient implementation of applications. Therefore, aiming at the task scheduling problem in cloud computing environment, this paper proposes a new list scheduling algorithm, namely, based on a virtual scheduling length (BVSL) table algorithm. The algorithm first constructs the predicted remaining length table based on the prescheduling results, then constructs a virtual scheduling length table based on the predicted remaining length table, the current task execution cost, and the actual start time of the task, and calculates the task priority based on the virtual scheduling length table to make the overall path the longest task is scheduled first, thus effectively shorten the scheduling length. Finally, the processor is selected for the task based on the predicted remaining length table. The selected processor may not be the earliest for the current task, but it can shorten the finish time of the task in the next phase and reduce the scheduling length. To verify the effectiveness of the scheduling method, experiments were carried out from two aspects: randomly generated graphs and real-world application graphs. Experimental results show that the BVSL algorithm outperforms the latest Improved Predict Priority Task Scheduling (IPPTS) and RE-18 scheduling methods in terms of makespan, scheduling length ratio, speedup, and the number of occurrences of better quality of schedules while maintaining the same time complexity.

University of Warsaw Lagrangian Cloud Model (UWLCM) 2.0: Adaptation of a mixed Eulerian-Lagrangian numerical model for heterogeneous computing clusters

A numerical cloud model with Lagrangian particles coupled to an Eulerian flow is adapted for distributed memory systems. Eulerian and Lagrangian calculations can be done in parallell on CPUs and GPUs, respectively. Scaling efficiency and the amount of parallelization of CPU and GPU calculations both exceed 50 % for up to 40 nodes. A sophisticated Lagrangian microphysics model slows down simulation by only 50 % compared to a simplistic bulk microphysics model, thanks to the use of GPUs. Overhead of communications between cluster nodes is mostly related to the pressure solver. Presented method of adaptation for computing clusters can be used in any numerical model with Lagrangian particles coupled to an Eulerian fluid flow.

Optimization of a heterogeneous computing system architecture based on its coordination with a sensors group of a mobile unmanned platform

This article analyses the existing variety of sensors used in robotics and related fields and also proposes the architecture of a heterogeneous computing system designed to analyze data obtained from sensors of a mobile unmanned platform (MUP). A feature of unmanned platforms is the presence of tasks that require a significantly different level of performance of the on-board computing system for processing data from sensors of the corresponding type. Therefore, the adaptation of existing universal computing systems seems to be impractical, compared to the development of a specialized computing system with a heterogeneous architecture. The computing system is designed to solve problems of local navigation, stabilize the position of the MUP and control its movement, as well as control special equipment installed on the MUP. Often, if the goal is to ensure maximum efficiency, expressed in speed, accuracy and reliability, it is necessary to develop specialized devices. The article provides information on sensors of the main types used in robotics and indicates the requirements for the performance of a computing system necessary for processing data from sensors of this type. This, in turn, made it possible to propose a heterogeneous architecture containing processor subsystems focused on processing data from sensors requiring low, medium and high performance according to the considered classification.

Charged Particle Tracking via Edge-Classifying Interaction Networks

Recent work has demonstrated that geometric deep learning methods such as graph neural networks (GNNs) are well suited to address a variety of reconstruction problems in high-energy particle physics. In particular, particle tracking data are naturally represented as a graph by identifying silicon tracker hits as nodes and particle trajectories as edges, given a set of hypothesized edges, edge-classifying GNNs identify those corresponding to real particle trajectories. In this work, we adapt the physics-motivated interaction network (IN) GNN toward the problem of particle tracking in pileup conditions similar to those expected at the high-luminosity Large Hadron Collider. Assuming idealized hit filtering at various particle momenta thresholds, we demonstrate the IN’s excellent edge-classification accuracy and tracking efficiency through a suite of measurements at each stage of GNN-based tracking: graph construction, edge classification, and track building. The proposed IN architecture is substantially smaller than previously studied GNN tracking architectures; this is particularly promising as a reduction in size is critical for enabling GNN-based tracking in constrained computing environments. Furthermore, the IN may be represented as either a set of explicit matrix operations or a message passing GNN. Efforts are underway to accelerate each representation via heterogeneous computing resources towards both high-level and low-latency triggering applications.