Skip to contents

Configure and run HMC/NUTS safely in true parallel (one chain per worker)

Source: R/diagnostics.R
configure_hmc_safely_parallel.Rd

Parallel counterpart of configure_hmc_safely_bis, implementing true parallelism by running one MCMC chain per PSOCK worker (parallel::makeCluster). Each chain is executed in an isolated R session, and results are aggregated on the master into a single coda::mcmc.list.

Usage

configure_hmc_safely_parallel(
  build_fn,
  niter,
  nburnin,
  thin,
  nchains,
  monitors,
  inits = NULL,
  enable_WAIC = FALSE,
  show_compiler_output = FALSE,
  log_dir = "log",
  seed = 1L
)

Arguments

build_fn

Function; model builder used to construct the NIMBLE model and base MCMC configuration (e.g. build_M). If .fresh_build() is available on workers, it is called as .fresh_build(build_fn, monitors, thin); otherwise, build_fn() must return a list with at least model and conf. If build_fn accepts a chain_id argument, it is passed automatically.

niter

Integer; total number of MCMC iterations per chain.

nburnin

Integer (>= 0); number of burn-in iterations.

thin

Integer (>= 1); thinning interval.

nchains

Integer (>= 1); number of chains (one PSOCK worker per chain).

monitors

Character vector or NULL; nodes to monitor, passed to model building and/or HMC configuration.

inits

NULL or a list of per-chain initial values. If provided, its length must match nchains; each element is passed to the corresponding worker.

enable_WAIC

Logical; whether to enable WAIC computation (if supported).

show_compiler_output

Logical; forwarded to nimbleOptions(showCompilerOutput = ...) on each worker.

log_dir

Character; directory for worker-level logs (if enabled).

seed

Integer; base RNG seed. Chain i uses seed + i - 1.

Value

A list with components:

  • conf: final MCMC configuration (from chain 1),

  • cmcmc: compiled MCMC object (from chain 1),

  • samples: combined coda::mcmc.list (one chain per worker),

  • runtime_s: wall-clock time in seconds (maximum over chains),

  • build: build object returned by build_fn or .fresh_build() (from chain 1).

Details

Relative to the sequential implementation, this function:

  • launches a dedicated PSOCK worker for each chain, ensuring full memory and RNG isolation (no forking),

  • initializes each worker by attaching nimble (and nimbleHMC when available) and defining internal helpers that compiled or lifted NIMBLE code may call unqualified,

  • runs nimble::runMCMC(..., nchains = 1L) independently on each worker and combines chains on the master,

  • reports runtime_s as the maximum elapsed time across chains, corresponding to the effective wall-clock time of the parallel execution,

  • returns compiled objects (conf, cmcmc, build) from the first chain only, avoiding the transfer of large compiled objects from all workers.

In contrast to mixed or fallback strategies, this function performs a strict HMC/NUTS configuration: no per-node fallback (e.g. addSampler(type = "slice")) is attempted. Any failure during HMC configuration or execution results in an error.