DYRE: a DYnamic REconfigurable solution to increase GPGPU’s reliability

The Journal of Supercomputing - General-purpose graphics processing units (GPGPUs) are extensively used in high-performance computing. However, it is well known that these devices’...     

Evaluating Different Solutions to Design Fault Tolerant Systems with SRAM-based FPGAs

The latest SRAM-based FPGA devices are making the development of low-cost, high-performance, re-configurable systems feasible, paving the way for innovative architectures suitable for mission- or safety-critical applications, such as those dominating the space or avionic fields. Unfortunately, SRAM-based FPGAs are extremely sensitive to Single Event Upsets (SEUs) induced by radiation. SEUs may alter the logic value stored in the memory elements the FPGAs embed. A large part of the FPGA memory elements is dedicated to the configuration memory, whose content dictates how the resources inside the FPGA have to be used to implement any given user circuit, SEUs affecting configuration memory cells can be extremely critics. Facing the effects of SEUs through radiation-hardened FPGAs is not cost-effective. Therefore, various fault-tolerant design techniques have been devised for developing dependable solutions, starting from Commercial-Off-The-Shelf (COTS) SRAM-based FPGAs. These techniques present advantages and disadvantages that must be evaluated carefully to exploit them successfully. In this paper we mainly adopted an empirical analysis approach. We evaluated the reliability of a multiplier, a digital FIR filter, and an 8051 microprocessor implemented in SRAM-based FPGA’s, by means of extensive fault-injection experiments, assessing the capability provided by different design techniques of tolerating SEUs within the FPGA configuration memory. Experimental results demonstrate that by combining architecture-level solutions (based on redundancy) with layout-level solutions (based on reliability-oriented place and route) designers may implement reliable re-configurable systems choosing the best solution that minimizes the penalty in terms of area and speed degradation.     

Recovery Time and Fault Tolerance Improvement for Circuits mapped on SRAM-based FPGAs

The rapid adoption of FPGA-based systems in space and avionics demands dependability rules from the design to the layout phases to protect against radiation effects. Triple Modular Redundancy is a widely used fault tolerance methodology to protect circuits against radiation-induced Single Event Upsets implemented on SRAM-based FPGAs. The accumulation of SEUs in the configuration memory can cause the TMR replicas to fail, requiring a periodic write-back of the configuration bit-stream. The associated system downtime due to scrubbing and the probability of simultaneous failures of two TMR domains are increasing with growing device densities. We propose a methodology to reduce the recovery time of TMR circuits with increased resilience to Cross-Domain Errors. Our methodology consists of an automated tool-flow for fine-grain error detection, error flags convergence and non-overlapping domain placement. The fine-grain error detection logic identifies the faulty domain using gate-level functions while the error flag convergence logic reduces the overwhelming number of flag signals. The non-overlapping placement enables selective domain reconfiguration and greatly reduces the number of Cross-Domain Errors. Our results demonstrate an evident reduction of the recovery time due to fast error detection time and selective partial reconfiguration of faulty domains. Moreover, the methodology drastically reduces Cross-Domain Errors in Look-Up Tables and routing resources. The improvements in recovery time and fault tolerance are achieved at an area overhead of a single LUT per majority voter in TMR circuits.     

Software and Hardware Techniques for SEU Detection in IP Processors

In the recent years both software and hardware techniques have been adopted to carry out reliable designs, aimed at autonomously detecting the occurrence of faults, to allow discarding erroneous data and possibly performing the recovery of the system. The aim of this paper is the introduction of a combined use of software and hardware approaches to achieve a complete fault coverage in generic IP processors, with respect to SEU faults. Software techniques are preferably adopted to reduce the necessity and costs of modifying the processor architecture; since a complete fault coverage cannot be achieved, partial hardware redundancy techniques are then introduced to deal with the remaining, not covered, faults. The paper presents the methodological approach adopted to achieve the complete fault coverage, the proposed resulting architecture, and the experimental results gathered from the analysis of the fault injection campaigns.     

Evaluating low-level software-based hardening techniques for configurable GPU architectures

The high processing power of GPUs makes them attractive for safety-critical applications, where transient effects are a major concern, and resilience must be enforced without compromising performance. Configurable softcore GPUs are a recent technology that allows detailed reliability assessment capable of bringing directions to the design of reliable GPU applications. This work investigates the reliability of the register files and the pipeline of a softcore GPU under radiation-induced faults. It proposes software-based fault tolerance techniques to mitigate errors. Faults are simulated at the register transfer level in four case-study algorithms, and the Architectural Vulnerability Factor (AVF) and Mean Workload to Failure (MWTF) are checked over different GPU configurations. Results indicate that software-based techniques efficiently reduce AVF. In terms of MWTF, results show that the best cases depend on an optimized balance between GPU configuration, application runtime, and AVF.

     

Water jump reorientation: from theoretical prediction to experimental observation

Liquid water is remarkably labile in reorganizing its hydrogen-bond (HB) network through the breaking and forming of HBs. This rapid restructuring, which occurs on the picosecond time scale, is critical not only for many of the pure liquid's special features but also for a range of aqueous media phenomena, including chemical reactions and protein activity. An essential part of the HB network reorganization is water molecule reorientation, which has long been described as Debye rotational diffusion characterized by very small angular displacements. Recent theoretical work, however, has presented a starkly contrasting picture: a sudden, large-amplitude jump mechanism, in which the reorienting water molecule rapidly exchanges HB partners in what amounts to an activated chemical reaction. In this Account, we first briefly review the jump mechanism and then discuss how it is supported by a series of experiments. These studies range from indirect indications to direct characterization of the jumps through pioneering two-dimensional infrared spectroscopy (2D-IR), the power of which accords it a special focus here. The scenarios in which experimental signatures of the jump mechanism are sought increase in complexity throughout the Account, beginning with pure water. Here 2D-IR in combination with theory can give a glimpse of the jumps, but the tell-tale markers are not pronounced. A more fruitful arena is provided by aqueous ionic solutions. The difference between water-water and water-anion HB strengths provides the experimental handle of differing OH stretch frequencies; in favorable cases, the kinetic exchange of a water between these two sites can be monitored. Sole observation of this exchange, however, is insufficient to establish the jump mechanism. Fortunately, 2D-IR with polarized pulses has demonstrated that HB exchange is accompanied by significant angular displacement, with an estimated jump angle similar to theoretical estimates. The Janus-like character of amphiphilic solutes, with their hydrophobic and hydrophilic faces, presents a special challenge for theory and experiment. Here a consensus on the 2D-IR interpretation has not yet been achieved; this lack of accord impedes the understanding of, for example, biochemical solutes and interfaces. We argue that the influence of hydrophobic groups on water jumps is only modest and well accounted for by an excluded volume effect in the HB exchange process. Conversely, hydrophilic groups have an important influence when their HB strength with water differs significantly from that of the water-water HB. The power of 2D-IR is argued to be accompanied by subtleties that can lead to just the opposite and, in our view, erroneous conclusion. We close with a prediction that a hydrophobic surface offers an arena in which the dynamics of "dangling" water OHs, bereft of a HB, could provide a 2D-IR confirmation of water jumps.     

Coarse-grained and All-atom Simulations towards the Early and Late Steps of Amyloid Fibril Formation

Alzheimer's disease is the most common neurodegenerative disease. Experiments and computer simulations can complement one another to provide a full and in-depth understanding of many aspects in the amyloid field at the atomistic level. Here, we review results of our coarse-grained and all-atom simulations in aqueous solution aimed at determining: 1) early aggregation steps of short linear peptides; 2) nucleation size number; 3) solution structure of the Aβ_1–40/Aβ_1–42 wild-type dimers; 4) impact of FAD (familial forms of Alzheimer's disease) mutations on the structure of Aβ_1–40/Aβ_1–42 dimers; and 5) impact of protective mutations on the structure of Aβ_1–40/Aβ_1–42 dimers.