Impact of GPUs Parallelism Management on Safety-Critical and HPC Applications Reliability

Abstract

Graphics Processing Units (GPUs) offer high computational power but require high scheduling strain to manage parallel processes, which increases the GPU cross section. The results of extensive neutron radiation experiments performed on NVIDIA GPUs confirm this hypothesis. Reducing the application Degree Of Parallelism (DOP) reduces the scheduling strain but also modifies the GPU parallelism management, including memory latency, thread registers number, and the processors occupancy, which influence the sensitivity of the parallel application. An analysis on the overall GPU radiation sensitivity dependence on the code DOP is provided and the most reliable configuration is experimentally detected. Finally, modifying the parallel management affects the GPU cross section but also the code execution time and, thus, the exposure to radiation required to complete computation. The Mean Workload and Executions Between Failures metrics are introduced to evaluate the workload or the number of executions computed correctly by the GPU on a realistic application.

GPGPUs ECC efficiency and efficacy

“…The ECC seems effective in reducing SDC occurrences. Some SDC still occur even if ECC is enabled as SET on logic resources or scheduler failures are left undetected, and both these resources have been demonstrated to contribute significantly to GPGPUs SDC rate [7], [18].…”

“…A SDC is typically produced when radiation corrupts memory elements storing variables or data used for computation, or when the logic executing an operation experience a Single Event Transient (SET) [9]. Additionally, on GPGPUs a SDC can occur when the scheduler fails in synchronizing threads, assigning a thread to a proper CUDA core, or presents results that are still incomplete [18]. A FI happens, for instance, when radiation induces a control flow error and prevents the application running on the GPGPU from being completed, when a scheduler failure hangs the GPGPU kernel or when the PCI-Express bus controller is corrupted.…”

Radiation-Induced Error Criticality in Modern HPC Parallel Accelerators

“…The scheduler on NVIDIA devices is implemented in hardware and has already been demonstrated to contribute to the device radiation sensitivity [34]. Intel Xeon Phi relies on the operating system to manage execution [22] which may be less susceptible to radiation-induced failures.…”

“…It is worth noting that while the K40 thread management seems to increase its sensitivity, it may be more efficient. The K40 may then produce more correct data before experiencing a failure [34].…”

Microarchitecture level reliability comparison of modern GPU designs: First findings

“…Recently, the research community has started tackling the challenging problem of characterizing the reliability of GPGPU based systems, i.e., their vulnerability to soft and hard errors [1] [2]. This challenging problem requires the development of accurate and fast reliability assessment techniques to deal with the delicate trade-off between analysis time and accuracy of the reported measurements and the ability to provide results that can guide system designers in the choice and development of efficient error resilience mechanisms.…”