EllipticSolverChain::FallbackPolicy Struct Reference

Stateful policy for the solver fallback chain. More...

#include <CD_EllipticSolverChain.H>

Public Member Functions
int	startIndex (const Vector< SolverType > &a_solvers)
	Start index into a_solvers for this solve, advancing the re-probe counter. Call exactly once per solve.
void	recordOutcome (const bool a_gmgAttempted, const bool a_gmgConverged, const int a_krylovIters=-1, const int a_verbosity=-1, const std::string &a_name="")
	Record the outcome of this solve and update the latch.
bool	isLatched () const
	Whether the policy is currently latched onto the Krylov fallback.

Public Attributes
int	retryInterval = 1
	Re-probe GMG every this many solves while latched. Clamped to >= 1; 1 reproduces always-retry behaviour.
int	reprobeIters = 0
	Adaptive early re-probe: re-probe GMG on the next solve whenever the Krylov fallback converged in this many iterations or fewer (a strong preconditioner is a sign that GMG alone may now succeed). 0 disables it, so only the periodic retryInterval applies.

Protected Member Functions
int	firstNonGmg (const Vector< SolverType > &a_solvers) const
	Index of the first non-GMG solver in the chain, or 0 if there is none.

Protected Attributes
bool	m_latched = false
	Whether a failed GMG attempt has latched us onto the Krylov fallback.
int	m_sinceProbe = 0
	Solves since the last GMG attempt while latched.
bool	m_reprobeNext = false
	Whether a cheap Krylov solve has armed an adaptive early re-probe of GMG on the next solve.

Detailed Description

Stateful policy for the solver fallback chain.

When the leading stand-alone multigrid (GMG) entry of a chain such as gmg gmres fails, the chain falls back to the Krylov solver. Without memory the next solve pays a full failed GMG attempt again, which is wasteful when the failure is persistent (e.g. a fixed, hard field operator). This policy latches onto the Krylov fallback after a GMG failure and only re-probes GMG every retryInterval solves, reclaiming multigrid's speed whenever the problem becomes tractable again. retryInterval == 1 reproduces the memoryless "always retry GMG first" behaviour; a large value approaches a permanent latch. Optionally, reprobeIters arms an adaptive early re-probe: if the Krylov fallback converges in very few iterations the preconditioner is clearly strong, so GMG alone is likely to succeed and is re-probed on the next solve without waiting for retryInterval.

Note

The state is deliberately decoupled from regridding: hold an instance as a host-solver member and do not reset it when the operators are rebuilt (regrids may be far more frequent than the desired re-probe cadence). The counter advances per solve, so the re-probe cadence is measured in solves, not necessarily time steps.

Member Function Documentation

firstNonGmg()

int EllipticSolverChain::FallbackPolicy::firstNonGmg ( const Vector< SolverType > &  a_solvers ) const

inlineprotected

Index of the first non-GMG solver in the chain, or 0 if there is none.

Parameters

[in]

a_solvers

The fallback chain.

Returns

Start index that skips a leading GMG.

isLatched()

bool EllipticSolverChain::FallbackPolicy::isLatched ( ) const

inline

Whether the policy is currently latched onto the Krylov fallback.

Returns

True if latched.

recordOutcome()

void EllipticSolverChain::FallbackPolicy::recordOutcome ( const bool  a_gmgAttempted, const bool  a_gmgConverged, const int  a_krylovIters = -1, const int  a_verbosity = -1, const std::string &  a_name = ""  )

inline

Record the outcome of this solve and update the latch.

Latching (a failed GMG attempt) is already announced by the host's fallback-chain message; this method additionally announces the release (a successful re-probe) at verbosity >= 1, since that transition is otherwise silent. While latched and skipping GMG, a Krylov solve that converged in <= reprobeIters iterations arms an adaptive early re-probe of GMG on the next solve.

Parameters

[in]	a_gmgAttempted	Whether a stand-alone GMG attempt was made this solve.
[in]	a_gmgConverged	Whether that GMG attempt converged.
[in]	a_krylovIters	Iterations taken by the Krylov fallback this solve, or < 0 if none ran.
[in]	a_verbosity	Host verbosity; a release message is printed when >= 1.
[in]	a_name	Host solver name, used to prefix the release message.

startIndex()

int EllipticSolverChain::FallbackPolicy::startIndex ( const Vector< SolverType > &  a_solvers )

inline

Start index into a_solvers for this solve, advancing the re-probe counter. Call exactly once per solve.

Returns 0 (include the leading GMG) when unlatched, when the periodic re-probe is due, or when an adaptive early re-probe was armed by a cheap Krylov solve; otherwise returns the index of the first non-GMG solver so the persistently-failing GMG attempt is skipped.

Parameters

[in]

a_solvers

The configured fallback chain.

Returns

Start index into a_solvers.

---

The documentation for this struct was generated from the following file:

• Source/Elliptic/CD_EllipticSolverChain.H