Heterogeneous Multi-Processor Systems-on-Chip (MPSoCs) that combine multiple, heterogeneous processing units are becoming increasingly popular for a wide range of applications, including industrial applications, where complex real-time applications can benefit from the performance and flexibility they offer. However, deploying real-time applications with low latency requirements across multiple processing units on such MPSoCs remains a challenging problem, particularly when communication between processors is required on a time-critical path. Existing solutions generally rely on the presence of at least one full-featured, general-purpose operating system on the device, and do not cater to the requirements of distributed, low-latency real-time applications. In this paper, we investigate the performance, with a focus on latency, of different options for communication between CPUs, including inter-processor interrupts and shared memory communication via different memories, on the popular Xilinx Zynq UltraScale+ platform and propose a novel solution for communication between heterogeneous processing units that relies only on the availability of shared memory. Our solution is capable of achieving sub-microsecond latencies for signaling and the transfer of small amounts of data between processing units, making it suitable for deploying distributed, low-latency real-time applications.
Precision-timed (PRET) machines are an alternative to modern processors that provide precise control over the timing of software execution. This paper describes a platform for developing predictable real-time embedded systems that pair PRET machines with Lingua Franca (LF), a recent reactor-based coordination language with temporal semantics. Specifically, we port LF to FlexPRET, a PRET machine with flexible hardware thread scheduling. We evaluate single-threaded LF with a tight control loop style application on four embedded platforms, including the FlexPRET. The results reveal the underlying platform’s timing variability and how LF plus FlexPRET can remedy this timing variability. Finally, we compare single-threaded to multithreaded LF, again concerning timing. The four embedded platforms used are FlexPRET (bare-metal), RP2040 (bare-metal), nRF52 (with Zephyr), and Raspberry Pi 3b+ (with Linux). Our results indicate that FlexPRET with LF is attractive when precise timing is essential.
The number of applications subject to safety-critical regulations is on the rise, and consequently, the computing requirements for such applications are increasing as well. This trend has led to the integration of General-Purpose Graphics Processing Units (GPGPUs) into these systems. However, the inherent characteristics of GPGPUs, including their black-box nature, dynamic allocation mechanisms, and frequent use of pointers, present challenges in certifying these applications for safety-critical systems. This paper aims to shed light on the unique characteristics of GPU programs and how they impact the certification process. To achieve this goal, several allocation methods are rigorously evaluated to determine which one is best suited to an application, regarding the program characteristics within the safety-critical domain. By conducting this evaluation, we seek to provide insights into the complexities of GPU memory accesses and its compatibility with safety-critical requirements. The ultimate objective is to offer recommendations on the most appropriate allocation method based on the unique needs of each application, thus contributing to the safe and reliable integration of GPGPUs into safety-critical systems.
The rise in complexity of the algorithms run on space systems, largely attributable to higher resolution instruments which generate a large amount of the data to be processed, as well as to the need for increased autonomy, which relies on Neural Network inference systems in future missions, demand the adoption of more powerful on-board hardware, such as multicores. At the same time, the correctness and reliability of critical on-board software is of paramount importance for the success of space missions. However, developing such complex software in low-level languages can have a negative impact on these aspects. For this reason, this paper evaluates the role that the Rust programming language can have in this change, given its memory safety and built in support for parallelism, which allows to better utilise more powerful hardware, in particular multicore cpus, without compromising the programmability and safety of the code. To this end, the GPU4S benchmarking suite, part of the open source OBPMark benchmarking suite of the European Space Agency (ESA), is ported to Rust, with sequential and parallel implementations. The performance of the ported benchmarks is compared to the existing sequential and parallel implementations in low-level languages to evaluate the trade-offs of the different solutions, and it is evaluated on several multicore platforms which are candidates for future on-board processing systems. A particular focus is put on parallel versions of the benchmarks, where Rust offers solid native support, as well as library support for fast parallelization similar to OpenMP. Finally, in terms of correctness, the Rust implementations are free of recently detected defects in the low-level implementations of the GPU4S benchmarks.
The evolution of Graphics Processing Unit (GPU) compilers has facilitated the support for general-purpose programming languages across various architectures. The NVIDIA CUDA Compiler (NVCC) employs multiple compilation levels prior to generating machine code, implementing intricate optimizations to enhance performance. These optimizations influence the manner in which software is mapped to the underlying hardware, which can also impact GPU reliability. TASA is a source-to-source code randomization tool designed to alter the mapping of software onto the underlying hardware. It achieves this by generating random permutations of variable and function declarations, thereby introducing random padding between declarations of different types and modifying the program memory layout. Since this modifies their location in the memory, it also modifies their cache placement, affecting both their execution time (due to the different conflicts between them, which result in a different amount of cache misses in every execution), as well as their lifetime in the cache. In this work, which is part of the HiPEAC Student Challenge 2025, we first examine the reproducibility of a subset of data presented in the ACM TACO paper "Assessing the Impact of Compiler Optimizations on GPU Reliability" [10], and second we extend it by combining it with our proposal of software randomization. The paper indicates that the -O3 optimization flag facilitates an increased workload before failures occur within the application. By employing TASA, we investigate the impact of GPU randomization on reliability and performance metrics. By reproducing the results of the paper on a different GPU platform, we observe the same trend as reported in the original publication. Moreover, our preliminary results with the application of software randomization show in several cases an improved Mean Waiting Before Failure (MWBF) compared to the original source code.
The rapidly increasing complexity of embedded systems and the critical impact of non-functional requirements demand the adoption of an appropriate system-level HW/SW co-design methodology. This methodology tries to satisfy all design requirements by simultaneously considering several alternative HW/SW implementations. In this context, early performance estimation approaches are crucial in reducing the design space, thereby minimizing design time and cost. To address the challenge of system-level performance estimation, this work presents and formalizes a novel approach based on a unifying HW/SW performance metric for early execution time estimation. The proposed approach estimates the execution time of a C function when executed by different HW/SW processor technologies. The approach is validated through an extensive experimental study, demonstrating its effectiveness and efficiency in terms of estimation error (i.e., lower than 10%) and estimation time (close to zero) when compared to existing methods in the literature.
Mean Field Analysis and Markovian Agents are powerful techniques for modeling complex systems of distributed interacting objects, for which efficient analytical and numerical solution algorithms can be implemented through linear systems of ordinary differential equations (ODEs). Solving such ODE systems on Field Programmable Gate Arrays (FPGAs) is a promising alternative to traditional CPU-and GPU-based approaches, especially in terms of energy consumption; however, the floating-point computations required are generally thought to be slow and inefficient when implemented on FPGA. In this paper, we demonstrate the use of High-Level Synthesis with automated customization of low-precision floating-point calculations, obtaining hardware accelerators for ODE solvers with improved quality of results and minimal output error. The proposed methodology does not require any manual rewriting of the solver code, but it remains prohibitively slow to evaluate any possible floating-point configuration through logic synthesis; in the future, we will thus implement automated design space exploration methods able to suggest promising configurations under user-defined accuracy and performance constraints.
Cyber-physical systems (CPS) attempt to meet real-time and safety requirements by using hypervisors that provide isolation via virtualisation and Real-Time Operating Systems that manage the concurrency of system tasks. However, the operating system's (OS) decisions may hinder the efficiency of tasks because it needs more awareness of their specific intricacies. Hence, one critical limitation to efficiently developing CPSs is the lack of tailored parallel programming models that can harness the capabilities of advanced heterogeneous architectures while meeting the requirements integral to CPSs, such as real-time behaviour and safety requirements. While conventional HPC languages, like OpenMP and CUDA, cannot accommodate critical non-functional properties, safety languages, like Rust and Ada, are limited in their capabilities to exploit complex systems efficiently. On top of that, accessibility to the programming task is essential to making the system usable to different domain experts. HiPART tackles these challenges by developing a comprehensive framework holistically addressing efficiency, interoperability, reliability, and sustainability. The HiPART framework, based on OpenMP, provides tailored support for (1) real-time behaviour and safety requirements and (2) the efficient exploitation of advanced parallel and heterogeneous processor architectures. This support is exposed to users through extensions to the OpenMP specification and its implementation in the LLVM framework, including the compiler and the OpenMP runtime library. With this framework, HiPART will contribute to realising more capable and reliable autonomous systems across various domains, from autonomous mobility to space exploration.
Recent advancements in fields such as automotive and aerospace have driven a growing demand for robust computational resources. Applications that were once designed for basic MCUs are now deployed on highly heterogeneous SoC platforms. While these platforms deliver the necessary computational performance, they also present challenges related to resource sharing and predictability. These challenges are particularly pronounced when consolidating safety and non-safety-critical systems, the so-called Mixed-Criticality Systems (MCS) to adhere to strict SWaP-C requirements. MCS consolidation on shared platforms requires stringent spatial and temporal isolation to comply with functional safety standards. Virtualization, mainly leveraged by hypervisors, is a key technology that ensures spatial isolation across multiple OSes and applications; however, ensuring temporal isolation remains challenging due to contention on shared hardwar resources, which impacts real-time performance and predictability. To mitigate this problem, several strategies as cache coloring and memory bandwidth reservation have been proposed. Although cache coloring is typically implemented on state-of-the-art hypervisors, memory bandwidth reservation approaches are commonly implemented at the Linux kernel level or rely on dedicated hardware and typically do not consider the concept of VMs that can run different OSes. To fill the gap between current memory bandwidth reservation solutions and the deployment of MCSs that operate on a hypervisor, this work introduces H-MBR, an open-source VM-centric memory bandwidth reservation mechanism. H-MBR features (i) VM-centric bandwidth reservation, (ii) OS and platform agnosticism, and (iii) reduced overhead. Empirical results evidenced no overhead on non-regulated workloads, and negligible overhead (<1 for regulated workloads for regulation periods of 2 us or higher.
Modern embedded systems are evolving toward complex, heterogeneous architectures to accommodate increasingly demanding applications. Driven by SWAP-C constraints, this shift has led to consolidating multiple systems onto single hardware platforms. Static Partitioning Hypervisors offer a promising solution to partition hardware resources and provide spatial isolation between critical workloads. However, shared resources like the Last-Level Cache and system bus can introduce temporal interference between virtual machines (VMs), negatively impacting performance and predictability. Over the past decade, academia and industry have developed interference mitigation techniques, such as cache partitioning and memory bandwidth reservation. However, configuring these techniques is complex and time-consuming. Cache partitioning requires balancing cache sections across VMs, while memory bandwidth reservation needs tuning bandwidth budgets and periods. Testing all configurations is impractical and often leads to suboptimal results. Moreover, understanding how these techniques interact is limited, as their combined use can produce compounded or conflicting effects on performance. Static analysis tools estimating worst-case execution times offer guidance for configuring mitigation techniques but often fail to capture the complexity of modern multi-core systems. They typically focus on limited shared resources while neglecting others, such as IOMMUs and interrupt controllers. To address these challenges, we present SP-IMPact, an open-source framework for analyzing and guiding interference mitigation configurations. SP-IMPact supports (i) cache coloring and (ii) memory bandwidth reservation, while evaluating their interactions and cumulative impact. By providing insights on real hardware, SP-IMPact helps optimize configurations for mixed-criticality systems, ensuring performance and predictability.
We propose a new live-migration paradigm for virtual machines called zero-copy migration. By making the working set of the virtual machine available on the destination host through transparently byteaddressable disaggregated memory, we remove the need for a pre-copy phase while simultaneously reducing the performance impact of the post-copy phase. We describe an open-source implementation of the proposed paradigm based on QEMU-KVM and libvirt, and we evaluate the efficiency of the approach with a deployment on a functional hardware prototype of a memory disaggregation system realized using ThymesisFlow. Using a series of configurable benchmarks, we show that the lead time and blackout time of the migration are equal to best-case scenarios of traditional pre-copy, post-copy and hybrid approaches. Key performance metrics from the perspective of applications running in the virtual machine, such as memory latency and throughput, are improved by up to three orders of magnitude, increasing both flexibility and responsiveness of live-migrations in the datacenter.
In this paper, we present a method for optimizing C code for embedded multi-core systems using cross-layer parallelization. The method has two phases. The first is to develop the algorithm without any optimization for the target platform. Then, the second step is to optimize and parallelize the code across four defined layers which are the algorithm, code, task, and data layers, for efficient execution on the target hardware. Each layer is focused on selected hardware characteristics. By using an iterative approach, individual kernels and composite algorithms can be very well adapted to execution on the hardware without further adaptation of the algorithm itself. The realization of this cross-layer parallelization consists of algorithm recognition, code transformations, task distribution, and insertion of synchronization and communication statements. The method is evaluated first on a common kernel and then on a sample image processing algorithm to showcase the benefits of the approach. Compared to other methods that only rely on two or three of these layers, 20 to 30 % of additional performance gain can be achieved.
Processing graphs on a large scale presents a range of difficulties, including irregular memory access patterns, device memory limitations, and the need for effective partitioning in distributed systems, all of which can lead to performance problems on traditional architectures such as CPUs and GPUs. To address these challenges, recent research emphasizes the use of Field-Programmable Gate Arrays (FPGAs) within distributed frameworks, harnessing the power of FPGAs in a distributed environment for accelerated graph processing. This paper examines the effectiveness of a multi-FPGA distributed architecture in combination with a partitioning system to improve data locality and reduce inter-partition communication. Utilizing Hadoop at a higher level, the framework maps the graph to the hardware, efficiently distributing pre-processed data to FPGAs. The FPGA processing engine, integrated into a cluster framework, optimizes data transfers, using offline partitioning for large-scale graph distribution. A first evaluation of the framework is based on the popular PageRank algorithm, which assigns a value to each node in a graph based on its importance. In the realm of large-scale graphs, the single FPGA solution outperformed the GPU solution that were restricted by memory capacity and surpassing CPU speedup by 26x compared to 12x. Moreover, when a single FPGA device was limited due to the size of the graph, our performance model showed that a distributed system with multiple FPGAs could increase performance by around 12x. This highlights the effectiveness of our solution for handling large datasets that surpass on-chip memory restrictions.
The increasing complexity of transformer models in artificial intelligence expands their computational costs, memory usage, and energy consumption. Hardware acceleration tackles the ensuing challenges by designing processors and accelerators tailored for transformer models, supporting their computation hotspots with high efficiency. However, memory bandwidth can hinder improvements in hardware accelerators. Against this backdrop, in this paper we propose a novel memory arrangement strategy, governed by the hardware accelerator's kernel size, which effectively minimizes off-chip data access. This arrangement is particularly beneficial for end-to-end transformer model inference, where most of the computation is based on general matrix multiplication (GEMM) operations. Additionally, we address the overhead of non-GEMM operations in transformer models within the scope of this memory data arrangement. Our study explores the implementation and effectiveness of the proposed accelerator-driven data arrangement approach in both single- and multi-core systems. Our evaluation demonstrates that our approach can achieve up to a 2.8x speed increase when executing inferences employing state-of-the-art transformers.
European efforts to boost competitiveness in the space services sector promote the research and development of advanced software and hardware solutions. The EU-funded HERMES project contributes to the effort by qualifying radiation-hardened, high-performance programmable microprocessors and developing a software ecosystem that facilitates the deployment of complex applications on such platforms. The main objectives of the project include reaching a technology readiness level of 6 (i.e., validated and demonstrated in relevant environment) for the rad-hard NG-ULTRA FPGA with its ceramic hermetic package CGA 1752, developed within projects of the European Space Agency, French National Centre for Space Studies and the European Union. An equally important share of the project is dedicated to the development and validation of tools that support multicore software programming and FPGA acceleration. The HERMES project selected the Bambu High-Level Synthesis tool to integrate capabilities to translate C/C++ code into Verilog/VHDL in its development ecosystem. In HERMES, Bambu has been and will be extended to support new FPGA targets, architectural models, model-based design, and input applications. The increased performance offered by FPGAs is thus made available also to software developers who do not have hardware design expertise.
Precision tuning is an increasingly common approach for exploiting the tradeoff between energy efficiency or speedup, and accuracy. Its effectiveness is particularly strong whenever the maximum performance must be extracted from a computing system, such as embedded platforms. In these contexts, current engineering practice sees a dominance of memory-unsafe programming languages such as C and C++. However, the unsafe nature of these languages has come under great scrutiny as it leads to significant software vulnerabilities. Hence, safer programming languages which prevent memory-related bugs by design have been proposed as a replacement. Amongst these safer programming languages, one of the most popular has been Rust. In this work we adapt a state-of-the-art precision tuning tool, TAFFO, to operate on Rust code. By porting the PolyBench/C benchmark suite to Rust, we show that the effectiveness of the precision tuning is not affected by the use of a safer programming language, and moreover the safety properties of the language can be successfully preserved. Specifically, using TAFFO and Rust we achieved up to a 15x speedup over the base Rust code, thanks to the use of precision tuning.
Embedded system applications normally come with a set of non-functional requirements defined over properties (e.g., latency), expressed as a corridor of correct values via a lower and an upper bound per requirement. These requirements should be guaranteed during each execution of an application program on a given MPSoC platform. This can be achieved using a reactive control loop, where an enforcer controls a set of properties to be enforced, e.g., by adapting the number of cores allocated to a program or by scaling the voltage/frequency mode of active processors. An enforcement strategy may react on a requirement response differently, depending on (a) satisfying a requirement, or violating (b) a lower bound or (c) an upper bound. A better strategy might be to differentiate the reaction taken according to the amount of violation of a lower or upper bound, thus to react in a finer granular way. In this paper, we propose a design space exploration (DSE) method called Co-explore that automatically partitions the requirement corridors into so-called response ranges (i.e., sub-corridors) such that formulated verification goals of simultaneously generated enforcer FSMs, e.g., the number of consecutive violations of a requirement, are optimized. The evaluation shows that the explored enforcement FSMs can achieve higher probabilities of meeting a given set of requirements compared to reacting solely based on the ternary information (a), (b), or (c).
