Resumen de Exploiting task-based programming models for resilience
Hardware errors become more common as silicon technologies shrink and become more vulnerable, especially in memory cells, which are the most exposed to errors. Permanent and intermittent faults are caused by manufacturing variability and circuits ageing. While these can be mitigated once they are identified, their continuous rate of appearance throughout the lifetime of memory devices will always cause unexpected errors. In addition, transient faults are caused by effects such as radiation or small voltage/frequency margins, and there is no efficient way to shield against these events.
Other constraints related to the diminishing sizes of transistors, such as power consumption and memory latency have caused the microprocessor industry to turn to increasingly complex processor architectures. To solve the difficulties arising from programming such architectures, programming models have emerged that rely on runtime systems. These systems form a new intermediate layer on the hardware-software abstraction stack, that performs tasks such as distributing work across computing resources: processor cores, accelerators, etc. These runtime systems dispose of a lot of information, both from the hardware and the applications, and offer thus many possibilities for optimisations.
This thesis proposes solutions to the increasing fault rates in memory, across multiple resilience disciplines, from algorithm-based fault tolerance to hardware error correcting codes, through OS reliability strategies. These solutions rely for their efficiency on the opportunities presented by runtime systems.
The first contribution of this thesis is an algorithmic-based resilience technique, allowing to tolerate detected errors in memory. This technique allows to recover data that is lost by performing computations that rely on simple redundancy relations identified in the program. The recovery is demonstrated for a family of iterative solvers, the Krylov subspace methods, and evaluated for the conjugate gradient solver. The runtime can transparently overlap the recovery with the computations of the algorithm, which allows to mask the already low overheads of this technique.
The second part of this thesis proposes a metric to characterise the impact of faults in memory, which outperforms state-of-the-art metrics in precision and assurances on the error rate. This metric reveals a key insight into data that is not relevant to the program, and we propose an OS-level strategy to ignore errors in such data, by delaying the reporting of detected errors. This allows to reduce failure rates of running programs, by ignoring errors that have no impact.
The architectural-level contribution of this thesis is a dynamically adaptable Error Correcting Code (ECC) scheme, that can increase protection of memory regions where the impact of errors is highest. A runtime methodology is presented to estimate the fault rate at runtime using our metric, through performance monitoring tools of current commodity processors. Guiding the dynamic ECC scheme online using the methodology's vulnerability estimates allows to decrease error rates of programs at a fraction of the redundancy cost required for a uniformly stronger ECC.
This provides a useful and wide range of trade-offs between redundancy and error rates.
The work presented in this thesis demonstrates that runtime systems allow to make the most of redundancy stored in memory, to help tackle increasing error rates in DRAM. This exploited redundancy can be an inherent part of algorithms that allows to tolerate higher fault rates, or in the form of dead data stored in memory. Redundancy can also be added to a program, in the form of ECC. In all cases, the runtime allows to decrease failure rates efficiently, by diminishing recovery costs, identifying redundant data, or targeting critical data. It is thus a very valuable tool for the future computing systems, as it can perform optimisations across different layers of abstractions.
