Predictive Bounded-Deferral Checkpoint Scheduling Under Bandwidth Volatility: An LSTM-Guided Latency and Stability Analysis

Abstract

Checkpointing remains a central rollback-recovery mechanism for large-scale distributed execution, but the cost of a coordinated checkpoint is strongly coupled to the communication conditions at the instant at which checkpoint traffic is released. Fixed-period policies ignore this coupling, whereas reactive policies may act on a bandwidth observation that is already stale when the next checkpoint cycle begins. This study reformulates checkpoint initiation as a bounded-deferral control problem and proposes Predictive Bounded-Deferral Checkpointing (PBDC), an LSTM-guided scheduler that forecasts one-step-ahead bandwidth and either executes a due checkpoint, postpones it for a short re-evaluation interval, or forces execution when a reliability-preserving deferral budget is exhausted. A risk-sensitive objective combining expected cycle time, cycle-time variance, and deferral cost is developed, together with an analytical condition under which a one-step deferral is latency-beneficial. The approach is related to, but does not replace, classical Young-Daly checkpoint-period selection: the base checkpoint frequency remains reliability-driven, while PBDC shifts checkpoint initiation only within bounded temporal slack. Using the reported 1,000-process MPI simulation summaries of 1,000 cycles per strategy and 2,000,000 timing observations, the forecast-guided policy achieves a mean cycle time of 2.45 on the simulator timing scale, compared with 2.96 for reactive adaptation and 2.81 for static scheduling. Relative to the reactive baseline, mean cycle time decreases by 17.2%, standard deviation by 41.9%, variance by 66.3%, and maximum observed cycle time by 24.8%. Under an independent-cycle approximation, summary-statistics analysis gives Welch t ≈ 44.99 and Cohen's d ≈ 2.01 for PBDC versus the reactive baseline. The results support forecast-guided checkpoint timing as a practical mechanism for reducing both latency and instability; they also identify prediction calibration, wall-clock validation, and multi-feature network sensing as necessary next steps for production deployment..