Installing and configuring BLAS thread control

What the error means

An explicit request such as mmes(..., blasThreads="auto") can produce:

Error: Runtime BLAS control is unavailable. Install flexiblas for a FlexiBLAS R installation, or RhpcBLASctl for a supported BLAS; alternatively configure BLAS before launching R and leave blasThreads=NULL.

This is a thread-control availability error, not an error in the model formula or covariance structures. Three separate components are involved:

The BLAS library used by R. A threaded implementation, such as OpenBLAS or Intel MKL, performs dense numerical work. FlexiBLAS is a wrapper that selects a BLAS backend; it is not itself a guarantee of a threaded backend.
An R package that controls that library. The current sommer implementation uses flexiblas for an available FlexiBLAS installation, or RhpcBLASctl for a recognized supported BLAS installation.
Allocation-aware CPU detection. parallelly is needed specifically for blasThreads="auto". An explicit positive integer does not require it.

These packages are optional dependencies. Installing sommer does not necessarily install them, and installing a control package does not replace R’s BLAS library. No separate CHOLMOD installation is needed for solver="cholmod"; that solver uses the CHOLMOD bundled with R’s Matrix package.

Identify the R installation and BLAS

Run these commands in the same R session that you use to fit the model:

R.version.string
#> [1] "R version 4.6.1 (2026-06-24)"
R.home()
#> [1] "/opt/R/4.6.1/lib/R"
.libPaths()
#> [1] "/home/runner/work/_temp/Library" "/opt/R/4.6.1/lib/R/site-library"
#> [3] "/opt/R/4.6.1/lib/R/library"
extSoftVersion()["BLAS"]
#>                                                      BLAS 
#> "/usr/lib/x86_64-linux-gnu/openblas-pthread/libblas.so.3"
sessionInfo()[c("BLAS", "LAPACK")]
#> $BLAS
#> [1] "/usr/lib/x86_64-linux-gnu/openblas-pthread/libblas.so.3"
#> 
#> $LAPACK
#> [1] "/usr/lib/x86_64-linux-gnu/openblas-pthread/libopenblasp-r0.3.26.so"

Look for flexiblas, openblas, or mkl in the BLAS path or description. On a cluster, a different R module, terminal, RStudio session, or batch job can use a different R installation and package library. Installing the adapter into one installation does not make it available to another.

The current implementation recognizes non-FlexiBLAS paths containing openblas, mkl, acml, or goto for the RhpcBLASctl route. A generic name such as libRblas.so or Rblas.dll may conceal the actual backend; the current detection does not infer its identity from that name. It also does not currently recognize a standalone BLIS or Apple Accelerate backend through this route. A control package’s capabilities alone do not override these detection rules.

Install the appropriate R packages

Route A: R uses FlexiBLAS

In the R installation used for fitting, run:

install.packages(c("flexiblas", "parallelly"),
                 repos="https://cloud.r-project.org")

Restart R after installation, then check:

library(flexiblas)
stopifnot(flexiblas_avail())
flexiblas_current_backend()
flexiblas_get_num_threads()
parallelly::availableCores()

flexiblas_avail() must be TRUE. If it is FALSE, loading this R package has not made the running R installation use FlexiBLAS. Use the correct R installation or ask your administrator for an R build linked to FlexiBLAS.

If FlexiBLAS is available but its selected backend is a serial reference implementation, list the available backends:

flexiblas::flexiblas_list()
flexiblas::flexiblas_list_loaded()

# Only use OPENBLAS if that name appears in the available backend list.
backend <- flexiblas::flexiblas_load_backend("OPENBLAS")
flexiblas::flexiblas_switch(backend)
flexiblas::flexiblas_current_backend()
flexiblas::flexiblas_get_num_threads()

Backend names and availability depend on the system configuration. Switching backends changes BLAS/LAPACK for the whole R process; do it before fitting models and record the choice for reproducibility. mmes() changes the thread setting when requested, but does not choose or install a FlexiBLAS backend.

For a FlexiBLAS R installation, installing only RhpcBLASctl is not a substitute for this route: the current sommer controller deliberately requires the FlexiBLAS interface when the reported BLAS is FlexiBLAS.

Route B: R uses a recognized OpenBLAS or MKL library directly

Run:

install.packages(c("RhpcBLASctl", "parallelly"),
                 repos="https://cloud.r-project.org")

Restart R, then check:

library(RhpcBLASctl)
extSoftVersion()["BLAS"]
blas_get_num_procs()
parallelly::availableCores()

The getter alone does not prove that the backend supports changing its thread count. Some unsupported configurations return one. The small mmes() example below checks the complete path; an explicit request that the backend does not accept results in an error rather than a silent fallback.

If R still uses a reference BLAS

Installing RhpcBLASctl, flexiblas, or an operating-system OpenBLAS package does not by itself guarantee that the running R process uses a threaded BLAS. R must actually load or be linked to that library.

On Linux, use a distribution-supported R/BLAS configuration or an administrator provided R module linked to OpenBLAS, MKL, or FlexiBLAS. Some distributions use a BLAS alternatives mechanism, while other installations require a different R build. Follow the instructions for your distribution and R installation; there is no universal library-switch command. For source builds, consult the R Installation and Administration manual.

On Windows, RhpcBLASctl may be available as a binary, but installing it does not replace the BLAS shipped with R. Source packages need the Rtools version matching your R release. On macOS, source packages need the applicable command line compiler tools and, when required, a compatible Fortran compiler. Use a supported R distribution or build with a known threaded BLAS if the current backend cannot be controlled by this implementation. Do not manually overwrite R’s BLAS/LAPACK files with an arbitrary downloaded library.

If installation fails with missing headers or libraries, install the development files for the already selected numerical stack, or request them from your administrator. On a managed cluster, using its matching compiler and R modules is preferable to mixing incompatible system and user-installed libraries.

Verify control with a small model

First confirm that this session is loading a sommer build containing the argument. Older installed builds can coexist with the current source folder:

library(sommer)
packageVersion("sommer")
find.package("sommer")
stopifnot("blasThreads" %in% names(formals(mmes)))

The following example runs only when the required controller and parallelly are available. It uses small synthetic data, rather than requiring your DT or MMTi objects. It tests a one-thread request first, then allocation-aware automatic sizing. Both are short verification fits, not convergence or performance benchmarks.

blas <- unname(extSoftVersion()["BLAS"])
hasFlexiBLAS <- requireNamespace("flexiblas", quietly=TRUE) &&
  isTRUE(tryCatch(flexiblas::flexiblas_avail(), error=function(error) FALSE))
hasDirectControl <- !is.na(blas) &&
  !grepl("flexiblas", blas, ignore.case=TRUE) &&
  grepl("openblas|mkl|acml|goto", blas, ignore.case=TRUE) &&
  requireNamespace("RhpcBLASctl", quietly=TRUE)

if(requireNamespace("sommer", quietly=TRUE) &&
   "blasThreads" %in% names(formals(sommer::mmes)) &&
   requireNamespace("parallelly", quietly=TRUE) &&
   (hasFlexiBLAS || hasDirectControl)){
  local({
    suppressPackageStartupMessages(library(sommer))
    getThreads <- if(hasFlexiBLAS) flexiblas::flexiblas_get_num_threads else
      RhpcBLASctl::blas_get_num_procs
    before <- getThreads()
    exampleData <- expand.grid(Name=factor(seq_len(12L)),
      Env=factor(letters[1:3]), replicate=seq_len(2L))
    exampleData$Yield <- as.integer(exampleData$Env) +
      sin(as.integer(exampleData$Name)) + cos(seq_len(nrow(exampleData)))
    for(request in list(1L, "auto")){
      fit <- sommer::mmes(Yield~Env, random=~Name, rcov=~units,
        data=exampleData, solver="cholmod", nIters=2,
        dateWarning=FALSE, verbose=FALSE, blasThreads=request)
      print(fit$threadingDiagnostics[c("blasBackend", "controller",
        "blasThreadsRequested", "blasThreadsConfigured", "availableCPUs",
        "openmpMaxThreads", "numericalPath", "averageCPUs")])
      stopifnot(getThreads() == before)
    }
  })
}else{
  message("Verification skipped: install the matching controller and parallelly, and load a sommer build with blasThreads.")
}
#> Solver selected: cholmod
#> $blasBackend
#> [1] "/usr/lib/x86_64-linux-gnu/openblas-pthread/libblas.so.3"
#> 
#> $controller
#> [1] "RhpcBLASctl"
#> 
#> $blasThreadsRequested
#> [1] 1
#> 
#> $blasThreadsConfigured
#> [1] 1
#> 
#> $availableCPUs
#> [1] 4
#> 
#> $openmpMaxThreads
#> [1] 4
#> 
#> $numericalPath
#> [1] "block Schur"
#> 
#> $averageCPUs
#> [1] 1
#> Solver selected: cholmod
#> $blasBackend
#> [1] "/usr/lib/x86_64-linux-gnu/openblas-pthread/libblas.so.3"
#> 
#> $controller
#> [1] "RhpcBLASctl"
#> 
#> $blasThreadsRequested
#> [1] 4
#> 
#> $blasThreadsConfigured
#> [1] 4
#> 
#> $availableCPUs
#> [1] 4
#> 
#> $openmpMaxThreads
#> [1] 4
#> 
#> $numericalPath
#> [1] "block Schur"
#> 
#> $averageCPUs
#> [1] 1.375

blasThreadsConfigured should show the setting used during the fit. The controller restores the previous BLAS setting on return, including on errors. Therefore, querying BLAS threads after mmes() returns can show the old count; that does not mean the request was ignored. Use the returned diagnostics to inspect the fit’s configuration.

Retry your model

After completing the route matching your R installation, the original model can be fitted without changing its statistical specification:

library(sommer)
mixm2 <- mmes(value~envf, nIters=5,
  random=~vsm(fam(envf, 1), ism(id), Gu=MMTi) +
           vsm(dsm(envf), ism(repf)),
  rcov=~vsm(dsm(envf), ism(units)),
  data=DT, blasThreads="auto")

mixm2$threadingDiagnostics

Here DT and MMTi are your existing data and inverse relationship matrix. Alternatively, request an explicit count such as blasThreads=4L. Installing parallelly alone only resolves automatic CPU detection; it does not supply BLAS control.

Cluster allocations and startup configuration

"auto" uses parallelly::availableCores() rather than blindly using every CPU visible on the host. Scheduler limits, affinity and container quotas can make this count smaller than parallel::detectCores(). It does not request additional CPUs from a scheduler or distribute a fit across nodes.

For example, a Slurm job requesting eight CPUs for one R process can start with:

#!/bin/bash
#SBATCH --nodes=1
#SBATCH --ntasks=1
#SBATCH --cpus-per-task=8

# Load your site's compatible R/numerical-library modules here.
export OMP_NUM_THREADS=1
Rscript fit.R

With a working runtime controller, fit.R can use blasThreads="auto". Setting OpenMP to one isolates BLAS threading and avoids many OpenMP workers each launching a BLAS team; it is a starting point to benchmark, not a universally optimal setting. BLAS and Sommer’s OpenMP loops are separate threading layers. Some BLAS builds themselves use OpenMP, so startup variables and runtime setters can interact.

If runtime control is unavailable but your BLAS supports startup thread configuration, leave blasThreads=NULL and configure it before starting R. For a pthread-based OpenBLAS backend, for example:

OPENBLAS_NUM_THREADS=8 OMP_NUM_THREADS=1 Rscript fit.R

For MKL, use its documented MKL_NUM_THREADS setting; OpenMP-based OpenBLAS uses OMP_NUM_THREADS. Follow the selected backend’s documentation. These variables cannot make a serial reference BLAS threaded, and changing them with Sys.setenv() after BLAS initializes may have no effect. Without a supported query interface, the BLAS thread count can remain unknown to Sommer even when startup configuration works.

Troubleshooting and interpreting usage

Symptom	Check or action
`Runtime BLAS control is unavailable`	Check the BLAS identity in this R session; install its matching adapter in an active `.libPaths()` library.
`blasThreads='auto' requires ... parallelly`	Install `parallelly`, or supply a positive integer instead.
`flexiblas_avail()` is `FALSE`	Use R linked to FlexiBLAS; installing its R interface does not relink R.
Adapter installed but BLAS has a generic name	Current detection may not recognize that installation; use startup settings with `blasThreads=NULL` or an explicitly identified supported R build.
Backend did not accept the requested setting	Check whether it is serial or has a thread limit; select a supported threaded backend or use an accepted count.
`unused argument (blasThreads=...)`	The loaded Sommer is older than the source tree; install the current tree and restart R.
Packages available interactively but missing in a job	Check the job’s R module, `.libPaths()` and `R_LIBS_USER`; install or expose packages for that R installation.

Configured counts are not actual utilization. openmpMaxThreads is an OpenMP limit, not a CHOLMOD or BLAS usage count. averageCPUs measures aggregate process CPU seconds divided by elapsed seconds, not an instantaneous number of active threads. Factorization measurements are in threadingDiagnostics$factorizations; their measurement fields distinguish numeric CHOLMOD/LDLT work from the broader block-Schur workload.

Even solver="cholmod" can select a structure-aware Schur engine for an eligible model. Small blocks, serial work, memory bandwidth and thread startup overhead can prevent speedups. Confirm the numerical path and benchmark several counts on representative data rather than assuming that "auto" is fastest. The repository’s benchmarks/thread-scaling.R and accompanying README provide a repeated scaling benchmark with numerical-agreement checks.

sommer development team

2026-10-06