GDAL API: re-entrant, but not thread-safe
The exact meaning of the terms
re-entrant is not fully
standardized. We will use here the QT definitions.
In particular, a C function or C++ method is said to be re-entrant if it can
be called simultaneously from multiple threads, but only if each invocation
uses its own data.
All GDAL public C functions and C++ methods are re-entrant, except:
the general initialization functions, like
the general cleanup functions like
Those functions should not be called concurrently from several threads, and it is general best practice to call them from the main thread of the program at program initialization and termination.
Unless otherwise stated, no GDAL public C functions and C++ methods should be
assumed to be thread-safe. That is you should not call simultaneously GDAL
functions from multiple threads on the same data instance, or even instances
that are closely related through ownership relationships. For example, for a
multi-band raster dataset, it is not safe to call concurrently GDAL functions
GDALRasterBand instances owned by the same
GDALDataset instance (each thread should instead manipulate a
distinct GDALDataset). Similarly for a GDALDataset owning several
Those restrictions apply to the C and C++ ABI, and all languages bindings (unless they would take special precautions to serialize calls)
GDAL block cache and multi-threading
The current design of the GDAL raster block cache make it appropriate to read several datasets from several threads. However performance issues may arise when writing several datasets from several threads, due to lock contention in the global structures of the block cache mechanism.
RAM fragmentation and multi-threading
It has been observed that scenarios that involve multi-threading reading or writing of raster datasets are prone to cause a high RAM usage, in particular when using the default dynamic memory allocator of Linux. Using the alternate tcmalloc memory allocator helps reducing the amount of virtual and resident memory used.
For example, with Debian/Ubuntu distributions, this can be done by
libtcmalloc-minimal4 package and running the binary that
executes GDAL with:
GDAL and multi-processing
POSIX fork() API should not be called during the middle of a GDAL operation, otherwise some structures like mutexes might appear to be locked forever in the forked process. If multi-processing is done, we recommend that processes are forked before any GDAL operation is done. Operating on the same GDALDataset instance in several sub-processes will generally lead to wrong results due to the underlying file descriptors being shared.