GPU Process Exception Location via Hierarchical Task Simulation

The patent segments the task verification process into two distinct phases: a fast execution trace phase that processes all tasks to identify suspicious segments, and a slow simulation phase that only verifies suspicious segments in detail. This segmentation allows the system to maintain high reliability by thoroughly checking problematic areas while avoiding the time cost of simulating all tasks comprehensively.

The patent applies partial action by performing complete simulation verification only on suspicious task segments identified by the execution trace, rather than verifying all tasks. The execution trace provides a quick overview that flags only the segments needing detailed simulation, thus applying verification effort partially where most needed rather than excessively across all tasks.

The verification process is segmented into two complementary components: an execution trace mechanism that quickly identifies suspicious segments, and a simulation mechanism that provides accurate verification of those segments. This segmentation reduces overall process complexity by dividing the workload into a simple filtering stage and a detailed verification stage, rather than requiring one complex all-encompassing verification process.

The execution trace acts as an intermediary between the task execution and the simulation verification. It processes task execution data to identify suspicious segments, which then become the input for the simulation phase. This intermediary layer simplifies the overall process by pre-filtering tasks, so the simulation mechanism only needs to handle a small subset of tasks rather than all tasks directly.

The patent extracts only the essential execution trace information needed for identifying suspicious segments, rather than storing complete task execution data. The execution trace captures key metrics and state changes that enable quick identification of problematic segments, extracting only the necessary data for verification purposes while discarding redundant information that would consume storage resources.

Instead of storing all task data and filtering it during verification, the patent inverts the approach by generating execution traces that contain only the necessary verification information. The trace data is constructed specifically for verification purposes, storing information in a format that enables quick suspicious segment identification without requiring storage of the original complete task datasets.