The Threading Layers
This section is about the Numba threading layer, this is the library that is
used internally to perform the parallel execution that occurs through the use of
parallel targets for CPUs, namely:
The use of the
The use of the
If a code base does not use the
modules (or any other sort of parallelism) the defaults for the threading
layer that ship with Numba will work well, no further action is required!
Which threading layers are available?
There are three threading layers available and they are named as follows:
tbb- A threading layer backed by Intel TBB.
omp- A threading layer backed by OpenMP.
workqueue-A simple built-in work-sharing task scheduler.
In practice, the only threading layer guaranteed to be present is
omp layer requires the presence of a suitable OpenMP runtime library.
tbb layer requires the presence of Intel’s TBB libraries, these can be
obtained via the conda command:
$ conda install tbb
If you installed Numba with
pip, TBB can be enabled by running:
$ pip install tbb
Due to compatibility issues with manylinux1 and other portability concerns, the OpenMP threading layer is disabled in the Numba binary wheels on PyPI.
The default manner in which Numba searches for and loads a threading layer is tolerant of missing libraries, incompatible runtimes etc.
Setting the threading layer
The threading layer is set via the environment variable
NUMBA_THREADING_LAYER or through assignment to
numba.config.THREADING_LAYER. If the programmatic approach to setting the
threading layer is used it must occur logically before any Numba based
compilation for a parallel target has occurred. There are two approaches to
choosing a threading layer, the first is by selecting a threading layer that is
safe under various forms of parallel execution, the second is through explicit
selection via the threading layer name (e.g.
Setting the threading layer selection priority
By default the threading layers are searched in the order of
'workqueue'. To change this search order whilst
maintaining the selection of a threading layer based on availability, the
NUMBA_THREADING_LAYER_PRIORITY can be used.
Note that it can also be set via
it must occur logically before any Numba based
compilation for a parallel target has occurred.
For example, to instruct Numba to choose
omp first if available,
tbb and so on, set the environment variable as
NUMBA_THREADING_LAYER_PRIORITY="omp tbb workqueue".
numba.config.THREADING_LAYER_PRIORITY = ["omp", "tbb", "workqueue"].
Selecting a threading layer for safe parallel execution
Parallel execution is fundamentally derived from core Python libraries in four forms (the first three also apply to code using parallel execution via other means!):
spawning processes from the
spawn(default on Windows, only available in Python 3.4+ on Unix)
forking processes from the
fork(default on Unix).
forking processes from the
multiprocessingmodule through the use of a
forkserver(only available in Python 3 on Unix). Essentially a new process is spawned and then forks are made from this new process on request.
Any library in use with these forms of parallelism must exhibit safe behaviour under the given paradigm. As a result, the threading layer selection methods are designed to provide a way to choose a threading layer library that is safe for a given paradigm in an easy, cross platform and environment tolerant manner. The options that can be supplied to the setting mechanisms are as follows:
defaultprovides no specific safety guarantee and is the default.
safeis both fork and thread safe, this requires the
tbbpackage (Intel TBB libraries) to be installed.
forksafeprovides a fork safe library.
threadsafeprovides a thread safe library.
To discover the threading layer that was selected, the function
numba.threading_layer() may be called after parallel execution. For example,
on a Linux machine with no TBB installed:
from numba import config, njit, threading_layer import numpy as np # set the threading layer before any parallel target compilation config.THREADING_LAYER = 'threadsafe' @njit(parallel=True) def foo(a, b): return a + b x = np.arange(10.) y = x.copy() # this will force the compilation of the function, select a threading layer # and then execute in parallel foo(x, y) # demonstrate the threading layer chosen print("Threading layer chosen: %s" % threading_layer())
Threading layer chosen: omp
and this makes sense as GNU OpenMP, as present on Linux, is thread safe.
Selecting a named threading layer
Advanced users may wish to select a specific threading layer for their use case, this is done by directly supplying the threading layer name to the setting mechanisms. The options and requirements are as follows:
Threading Layer Name
GNU OpenMP libraries (very likely this will already exist)
MS OpenMP libraries (very likely this will already exist)
Should the threading layer not load correctly Numba will detect this and provide
a hint about how to resolve the problem. It should also be noted that the Numba
numba -s has a section
__Threading Layer Information__ that reports on the availability of
threading layers in the current environment.
The threading layers have fairly complex interactions with CPython internals and system level libraries, some additional things to note:
The installation of Intel’s TBB libraries vastly widens the options available in the threading layer selection process.
On Linux, the
ompthreading layer is not fork safe due to the GNU OpenMP runtime library (
libgomp) not being fork safe. If a fork occurs in a program that is using the
ompthreading layer, a detection mechanism is present that will try and gracefully terminate the forked child and print an error message to
On systems with the
fork(2)system call available, if the TBB backed threading layer is in use and a
forkcall is made from a thread other than the thread that launched TBB (typically the main thread) then this results in undefined behaviour and a warning will be displayed on
execit is safe to
spawnfrom a non-main thread, but as this cannot be differentiated from just a
forkcall the warning message will still be displayed.
On OSX, the
intel-openmppackage is required to enable the OpenMP based threading layer.
Setting the Number of Threads
The number of threads used by numba is based on the number of CPU cores
numba.config.NUMBA_DEFAULT_NUM_THREADS), but it can be
overridden with the
NUMBA_NUM_THREADS environment variable.
The total number of threads that numba launches is in the variable
For some use cases, it may be desirable to set the number of threads to a lower value, so that numba can be used with higher level parallelism.
The number of threads can be set dynamically at runtime using
numba.set_num_threads(). Note that
set_num_threads() only allows
setting the number of threads to a smaller value than
NUMBA_NUM_THREADS. Numba always launches
numba.config.NUMBA_NUM_THREADS threads, but
causes it to mask out unused threads so they aren’t used in computations.
The current number of threads used by numba can be accessed with
numba.get_num_threads(). Both functions work inside of a jitted
Example of Limiting the Number of Threads
In this example, suppose the machine we are running on has 8 cores (so
numba.config.NUMBA_NUM_THREADS would be
8). Suppose we want to run
some code with
@njit(parallel=True), but we also want to run our code
concurrently in 4 different processes. With the default number of threads,
each Python process would run 8 threads, for a total in 4*8 = 32 threads,
which is oversubscription for our 8 cores. We should rather limit each process
to 2 threads, so that the total will be 4*2 = 8, which matches our number of
There are two ways to do this. One is to set the
environment variable to
$ NUMBA_NUM_THREADS=2 python ourcode.py
However, there are two downsides to this approach:
NUMBA_NUM_THREADSmust be set before Numba is imported, and ideally before Python is launched. As soon as Numba is imported the environment variable is read and that number of threads is locked in as the number of threads Numba launches.
If we want to later increase the number of threads used by the process, we cannot.
NUMBA_NUM_THREADSsets the maximum number of threads that are launched for a process. Calling
set_num_threads()with a value greater than
numba.config.NUMBA_NUM_THREADSresults in an error.
The advantage of this approach is that we can do it from outside of the process without changing the code.
Another approach is to use the
numba.set_num_threads() function in our code
from numba import njit, set_num_threads @njit(parallel=True) def func(): ... set_num_threads(2) func()
If we call
set_num_threads(2) before executing our parallel code, it has
the same effect as calling the process with
NUMBA_NUM_THREADS=2, in that
the parallel code will only execute on 2 threads. However, we can later call
set_num_threads(8) to increase the number of threads back to the default
size. And we do not have to worry about setting it before Numba gets imported.
It only needs to be called before the parallel function is run.
Getting a Thread ID
In some cases it may be beneficial to have access to a unique identifier for the
current thread that is executing a parallel region in Numba. For that purpose,
Numba provides the
numba.get_thread_id() function. This function is the
corollary of OpenMP’s function
omp_get_thread_num and returns an integer
between 0 (inclusive) and the number of configured threads as described above
The total (maximum) number of threads launched by numba.
The number of usable CPU cores on the system (as determined by
len(os.sched_getaffinity(0)), if supported by the OS, or
multiprocessing.cpu_count()if not). This is the default value for
NUMBA_NUM_THREADSenvironment variable is set.
Set the number of threads to use for parallel execution.
By default, all
numba.config.NUMBA_NUM_THREADSthreads are used.
This functionality works by masking out threads that are not used. Therefore, the number of threads n must be less than or equal to
NUMBA_NUM_THREADS, the total number of threads that are launched. See its documentation for more details.
This function can be used inside of a jitted function.
- n: The number of threads. Must be between 1 and NUMBA_NUM_THREADS.
Get the number of threads used for parallel execution.
This number is less than or equal to the total number of threads that are launched,
This function can be used inside of a jitted function.
- The number of threads.