feat(triggers): added ConvergenceFailureTrigger

added new trigger which throws on convergence failures. This also required adding a new "step" method for other triggers which gets called every step instead of every trigger throw. The ConvergenceFailureTrigger has sucsessgully let teh engine evolve to 10Gyr without any meaningful growth of convergence failures.
This commit is contained in:
2025-11-04 13:22:57 -05:00
parent 7f19143ff6
commit 72a3f5bf4c
6 changed files with 285 additions and 21 deletions

View File

@@ -175,14 +175,24 @@ namespace gridfire::solver {
const size_t num_steps; ///< Number of CVODE steps taken so far.
const DynamicEngine& engine; ///< Reference to the engine.
const std::vector<fourdst::atomic::Species>& networkSpecies; ///< Species layout.
const size_t currentConvergenceFailures; ///< Total number of convergence failures
const size_t currentNonlinearIterations; ///< Total number of non linear iterations
/**
* @brief Construct a context snapshot.
*/
TimestepContext(
double t, const N_Vector& state, double dt, double last_step_time,
double t9, double rho, size_t num_steps, const DynamicEngine& engine,
const std::vector<fourdst::atomic::Species>& networkSpecies
double t,
const N_Vector& state,
double dt,
double last_step_time,
double t9,
double rho,
size_t num_steps,
const DynamicEngine& engine,
const std::vector<fourdst::atomic::Species>& networkSpecies,
size_t currentConvergenceFailure,
size_t currentNonlinearIterations
);
/**
@@ -281,7 +291,7 @@ namespace gridfire::solver {
SUNLinearSolver m_LS = nullptr; ///< Dense linear solver.
TimestepCallback m_callback; ///< Optional per-step callback.
std::optional<TimestepCallback> m_callback; ///< Optional per-step callback.
int m_num_steps = 0; ///< CVODE step counter (used for diagnostics and triggers).
bool m_stdout_logging_enabled = true; ///< If true, print per-step logs and use CV_ONE_STEP.

View File

@@ -71,6 +71,8 @@ namespace gridfire::trigger::solver::CVODE {
* (ctx.t - last_trigger_time) - interval for diagnostics.
*/
void update(const gridfire::solver::CVODESolverStrategy::TimestepContext &ctx) override;
void step(const gridfire::solver::CVODESolverStrategy::TimestepContext &ctx) override;
/**
* @brief Reset counters and last trigger bookkeeping (time and delta) to zero.
*/
@@ -149,6 +151,7 @@ namespace gridfire::trigger::solver::CVODE {
* @param ctx CVODE timestep context (unused except for symmetry with interface).
*/
void update(const gridfire::solver::CVODESolverStrategy::TimestepContext &ctx) override;
void step(const gridfire::solver::CVODESolverStrategy::TimestepContext &ctx) override;
/** @brief Reset counters to zero. */
void reset() override;
@@ -233,6 +236,7 @@ namespace gridfire::trigger::solver::CVODE {
* @param ctx CVODE timestep context.
*/
void update(const gridfire::solver::CVODESolverStrategy::TimestepContext &ctx) override;
void step(const gridfire::solver::CVODESolverStrategy::TimestepContext &ctx) override;
/** @brief Reset counters and clear the dt window. */
void reset() override;
@@ -268,6 +272,50 @@ namespace gridfire::trigger::solver::CVODE {
std::deque<double> m_timestep_window;
};
class ConvergenceFailureTrigger final : public Trigger<gridfire::solver::CVODESolverStrategy::TimestepContext> {
public:
explicit ConvergenceFailureTrigger(size_t totalFailures, float relativeFailureRate, size_t windowSize);
bool check(const gridfire::solver::CVODESolverStrategy::TimestepContext &ctx) const override;
void update(const gridfire::solver::CVODESolverStrategy::TimestepContext &ctx) override;
void step(const gridfire::solver::CVODESolverStrategy::TimestepContext &ctx) override;
void reset() override;
[[nodiscard]] std::string name() const override;
[[nodiscard]] std::string describe() const override;
[[nodiscard]] TriggerResult why(const gridfire::solver::CVODESolverStrategy::TimestepContext &ctx) const override;
[[nodiscard]] size_t numTriggers() const override;
[[nodiscard]] size_t numMisses() const override;
private:
/** @brief Logger used for trace/error diagnostics. */
quill::Logger* m_logger = fourdst::logging::LogManager::getInstance().getLogger("log");
/** @name Diagnostics counters */
///@{
mutable size_t m_hits = 0;
mutable size_t m_misses = 0;
mutable size_t m_updates = 0;
mutable size_t m_resets = 0;
size_t m_totalFailures;
float m_relativeFailureRate;
size_t m_windowSize;
std::deque<size_t> m_window;
private:
float current_mean() const;
bool abs_failure(const gridfire::solver::CVODESolverStrategy::TimestepContext& ctx) const;
bool rel_failure(const gridfire::solver::CVODESolverStrategy::TimestepContext& ctx) const;
};
/**
* @brief Compose a trigger suitable for deciding engine re-partitioning during CVODE solves.
*

View File

@@ -42,6 +42,12 @@ namespace gridfire::trigger {
* @param ctx Context snapshot used to update state.
*/
virtual void update(const TriggerContextStruct& ctx) = 0;
/**
* @brief similar to update but intended to be run on every step not just those where the trigger triggered
* @param ctx Context snapshot used to update state.
*/
virtual void step(const TriggerContextStruct& ctx) = 0;
/**
* @brief Reset internal state and diagnostics counters.
*/

View File

@@ -52,6 +52,8 @@ namespace gridfire::trigger {
* @brief Update both sub-triggers and increment update counter.
*/
void update(const TriggerContextStruct& ctx) override;
void step(const TriggerContextStruct& ctx) override;
/**
* @brief Reset both sub-triggers and local counters.
*/
@@ -84,6 +86,7 @@ namespace gridfire::trigger {
mutable size_t m_misses = 0;
mutable size_t m_updates = 0;
mutable size_t m_resets = 0;
mutable size_t m_steps = 0;
};
/**
@@ -101,6 +104,7 @@ namespace gridfire::trigger {
bool check(const TriggerContextStruct& ctx) const override;
void update(const TriggerContextStruct& ctx) override;
void step(const TriggerContextStruct& ctx) override;
void reset() override;
std::string name() const override;
TriggerResult why(const TriggerContextStruct& ctx) const override;
@@ -116,6 +120,7 @@ namespace gridfire::trigger {
mutable size_t m_misses = 0;
mutable size_t m_updates = 0;
mutable size_t m_resets = 0;
mutable size_t m_steps = 0;
};
/**
@@ -133,6 +138,7 @@ namespace gridfire::trigger {
bool check(const TriggerContextStruct& ctx) const override;
void update(const TriggerContextStruct& ctx) override;
void step(const TriggerContextStruct& ctx) override;
void reset() override;
std::string name() const override;
@@ -148,7 +154,7 @@ namespace gridfire::trigger {
mutable size_t m_misses = 0;
mutable size_t m_updates = 0;
mutable size_t m_resets = 0;
mutable size_t m_steps = 0;
};
/**
@@ -168,6 +174,7 @@ namespace gridfire::trigger {
bool check(const TriggerContextStruct& ctx) const override;
void update(const TriggerContextStruct& ctx) override;
void step(const TriggerContextStruct& ctx) override;
void reset() override;
std::string name() const override;
@@ -185,6 +192,7 @@ namespace gridfire::trigger {
mutable size_t m_misses = 0;
mutable size_t m_updates = 0;
mutable size_t m_resets = 0;
mutable size_t m_steps = 0;
};
///////////////////////////////
@@ -216,6 +224,13 @@ namespace gridfire::trigger {
m_updates++;
}
template <typename TriggerContextStruct>
void AndTrigger<TriggerContextStruct>::step(const TriggerContextStruct &ctx) {
m_A->step(ctx);
m_B->step(ctx);
m_steps++;
}
template <typename TriggerContextStruct>
void AndTrigger<TriggerContextStruct>::reset() {
m_A->reset();
@@ -301,6 +316,13 @@ namespace gridfire::trigger {
m_updates++;
}
template <typename TriggerContextStruct>
void OrTrigger<TriggerContextStruct>::step(const TriggerContextStruct &ctx) {
m_A->step(ctx);
m_B->step(ctx);
m_steps++;
}
template <typename TriggerContextStruct>
void OrTrigger<TriggerContextStruct>::reset() {
m_A->reset();
@@ -383,6 +405,12 @@ namespace gridfire::trigger {
m_updates++;
}
template <typename TriggerContextStruct>
void NotTrigger<TriggerContextStruct>::step(const TriggerContextStruct &ctx) {
m_A->step(ctx);
m_steps++;
}
template <typename TriggerContextStruct>
void NotTrigger<TriggerContextStruct>::reset() {
m_A->reset();
@@ -457,6 +485,12 @@ namespace gridfire::trigger {
m_updates++;
}
template <typename TriggerContextStruct>
void EveryNthTrigger<TriggerContextStruct>::step(const TriggerContextStruct &ctx) {
m_A->step(ctx);
m_steps++;
}
template <typename TriggerContextStruct>
void EveryNthTrigger<TriggerContextStruct>::reset() {
m_A->reset();
@@ -508,7 +542,4 @@ namespace gridfire::trigger {
size_t EveryNthTrigger<TriggerContextStruct>::numMisses() const {
return m_misses;
}
}