PyTables parameter files

PyTables issues warnings when certain limits are exceeded. Those limits are not intrinsic limitations of the underlying software, but rather are proactive measures to avoid large resource consumptions. The default limits should be enough for most of cases, and users should try to respect them. However, in some situations, it can be convenient to increase (or decrease) these limits.

Also, and in order to get maximum performance, PyTables implements a series of sophisticated features, like I/O buffers or different kind of caches (for nodes, chunks and other internal metadata). These features comes with a default set of parameters that ensures a decent performance in most of situations. But, as there is always a need for every case, it is handy to have the possibility to fine-tune some of these parameters.

Because of these reasons, PyTables implements a couple of ways to change the values of these parameters. All the tunable parameters live in the tables/parameters.py. The user can choose to change them in the parameter files themselves for a global and persistent change. Moreover, if he wants a finer control, he can pass any of these parameters directly to the tables.open_file() function, and the new parameters will only take effect in the corresponding file (the defaults will continue to be in the parameter files).

A description of all of the tunable parameters follows. As the defaults stated here may change from release to release, please check with your actual parameter files so as to know your actual default values.

Warning

Changing the next parameters may have a very bad effect in the resource consumption and performance of your PyTables scripts.

Please be careful when touching these!

Tunable parameters in parameters.py

Cache limits

tables.parameters.CHUNK_CACHE_NELMTS = 521

Number of elements for HDF5 chunk cache.

tables.parameters.CHUNK_CACHE_PREEMPT = 0.0

Chunk preemption policy. This value should be between 0 and 1 inclusive and indicates how much chunks that have been fully read are favored for preemption. A value of zero means fully read chunks are treated no differently than other chunks (the preemption is strictly LRU) while a value of one means fully read chunks are always preempted before other chunks.

tables.parameters.CHUNK_CACHE_SIZE = 2097152

Size (in bytes) for HDF5 chunk cache.

tables.parameters.COND_CACHE_SLOTS = 128

Maximum number of conditions for table queries to be kept in memory.

tables.parameters.METADATA_CACHE_SIZE = 1048576

Size (in bytes) of the HDF5 metadata cache.

tables.parameters.NODE_CACHE_SLOTS = 64

Maximum number of nodes to be kept in the metadata cache.

It is the number of nodes to be kept in the metadata cache. Least recently used nodes are unloaded from memory when this number of loaded nodes is reached. To load a node again, simply access it as usual. Nodes referenced by user variables and, in general, all nodes that are still open are registered in the node manager and can be quickly accessed even if they are not in the cache.

Negative value means that all the touched nodes will be kept in an internal dictionary. This is the faster way to load/retrieve nodes. However, and in order to avoid a large memory comsumption, the user will be warned when the number of loaded nodes will reach the -NODE_CACHE_SLOTS value.

Finally, a value of zero means that any cache mechanism is disabled.

Parameters for the different internal caches

tables.parameters.BOUNDS_MAX_SIZE = 1048576

The maximum size for bounds values cached during index lookups.

tables.parameters.BOUNDS_MAX_SLOTS = 4096

The maximum number of slots for the BOUNDS cache.

tables.parameters.ITERSEQ_MAX_ELEMENTS = 1024

The maximum number of iterator elements cached in data lookups.

tables.parameters.ITERSEQ_MAX_SIZE = 1048576

The maximum space that will take ITERSEQ cache (in bytes).

tables.parameters.ITERSEQ_MAX_SLOTS = 128

The maximum number of slots in ITERSEQ cache.

tables.parameters.LIMBOUNDS_MAX_SIZE = 262144

The maximum size for the query limits (for example, (lim1, lim2) in conditions like lim1 <= col < lim2) cached during index lookups (in bytes).

tables.parameters.LIMBOUNDS_MAX_SLOTS = 128

The maximum number of slots for LIMBOUNDS cache.

tables.parameters.TABLE_MAX_SIZE = 1048576

The maximum size for table chunks cached during index queries.

tables.parameters.SORTED_MAX_SIZE = 1048576

The maximum size for sorted values cached during index lookups.

tables.parameters.SORTEDLR_MAX_SIZE = 8388608

The maximum size for chunks in last row cached in index lookups (in bytes).

tables.parameters.SORTEDLR_MAX_SLOTS = 1024

The maximum number of chunks for SORTEDLR cache.

Parameters for general cache behaviour

Warning

The next parameters will not take any effect if passed to the open_file() function, so they can only be changed in a global way. You can change them in the file, but this is strongly discouraged unless you know well what you are doing.

tables.parameters.DISABLE_EVERY_CYCLES = 10

The number of cycles in which a cache will be forced to be disabled if the hit ratio is lower than the LOWEST_HIT_RATIO (see below). This value should provide time enough to check whether the cache is being efficient or not.

tables.parameters.ENABLE_EVERY_CYCLES = 50

The number of cycles in which a cache will be forced to be (re-)enabled, irregardingly of the hit ratio. This will provide a chance for checking if we are in a better scenario for doing caching again.

tables.parameters.LOWEST_HIT_RATIO = 0.6

The minimum acceptable hit ratio for a cache to avoid disabling (and freeing) it.

Parameters for the I/O buffer in Leaf objects

tables.parameters.IO_BUFFER_SIZE = 1048576

The PyTables internal buffer size for I/O purposes. Should not exceed the amount of highest level cache size in your CPU.

tables.parameters.BUFFER_TIMES = 100

The maximum buffersize/rowsize ratio before issuing a tables.PerformanceWarning.

Miscellaneous

tables.parameters.EXPECTED_ROWS_EARRAY = 1000

Default expected number of rows for EArray objects.

tables.parameters.EXPECTED_ROWS_TABLE = 10000

Default expected number of rows for Table objects.

tables.parameters.PYTABLES_SYS_ATTRS = True

Set this to False if you don’t want to create PyTables system attributes in datasets. Also, if set to False the possible existing system attributes are not considered for guessing the class of the node during its loading from disk (this work is delegated to the PyTables’ class discoverer function for general HDF5 files).

tables.parameters.MAX_NUMEXPR_THREADS = 2

The maximum number of threads that PyTables should use internally in Numexpr. If None, it is automatically set to the number of cores in your machine. In general, it is a good idea to set this to the number of cores in your machine or, when your machine has many of them (e.g. > 8), perhaps stay at 8 at maximum. In general, 2 threads is a good tradeoff.

tables.parameters.MAX_BLOSC_THREADS = 2

The maximum number of threads that PyTables should use internally in Blosc. If None, it is automatically set to the number of cores in your machine. In general, it is a good idea to set this to the number of cores in your machine or, when your machine has many of them (e.g. > 8), perhaps stay at 8 at maximum. In general, 2 threads is a good tradeoff.

HDF5 driver management

tables.parameters.DRIVER = None

The HDF5 driver that should be used for reading/writing to the file.

Following drivers are supported:

  • H5FD_SEC2: this driver uses POSIX file-system functions like read and write to perform I/O to a single, permanent file on local disk with no system buffering. This driver is POSIX-compliant and is the default file driver for all systems.
  • H5FD_DIRECT: this is the H5FD_SEC2 driver except data is written to or read from the file synchronously without being cached by the system.
  • H5FD_WINDOWS: this driver was modified in HDF5-1.8.8 to be a wrapper of the POSIX driver, H5FD_SEC2. This change should not affect user applications.
  • H5FD_STDIO: this driver uses functions from the standard C stdio.h to perform I/O to a single, permanent file on local disk with additional system buffering.
  • H5FD_CORE: with this driver, an application can work with a file in memory for faster reads and writes. File contents are kept in memory until the file is closed. At closing, the memory version of the file can be written back to disk or abandoned.
  • H5FD_SPLIT: this file driver splits a file into two parts. One part stores metadata, and the other part stores raw data. This splitting a file into two parts is a limited case of the Multi driver.

The following drivers are not currently supported:

  • H5FD_LOG: this is the H5FD_SEC2 driver with logging capabilities.
  • H5FD_FAMILY: with this driver, the HDF5 file’s address space is partitioned into pieces and sent to separate storage files using an underlying driver of the user’s choice. This driver is for systems that do not support files larger than 2 gigabytes.
  • H5FD_MULTI: with this driver, data can be stored in multiple files according to the type of the data. I/O might work better if data is stored in separate files based on the type of data. The Split driver is a special case of this driver.
  • H5FD_MPIO: this is the standard HDF5 file driver for parallel file systems. This driver uses the MPI standard for both communication and file I/O.
  • H5FD_MPIPOSIX: this parallel file system driver uses MPI for communication and POSIX file-system calls for file I/O.
  • H5FD_STREAM: this driver is no longer available.

See also

the Drivers section of the HDF5 User’s Guide for more information.

Note

not all supported drivers are always available. For example the H5FD_WINDOWS driver is not available on non Windows platforms.

If the user try to use a driver that is not available on the target platform a RuntimeError is raised.

New in version 3.0.

tables.parameters.DRIVER_DIRECT_ALIGNMENT = 0

Specifies the required alignment boundary in memory.

A value of 0 (zero) means to use HDF5 Library’s default value.

New in version 3.0.

tables.parameters.DRIVER_DIRECT_BLOCK_SIZE = 0

Specifies the file system block size.

A value of 0 (zero) means to use HDF5 Library’s default value of 4KB.

New in version 3.0.

tables.parameters.DRIVER_DIRECT_CBUF_SIZE = 0

Specifies the copy buffer size.

A value of 0 (zero) means to use HDF5 Library’s default value.

New in version 3.0.

tables.parameters.DRIVER_CORE_INCREMENT = 65536

Core driver memory increment.

Specifies the increment by which allocated memory is to be increased each time more memory is required.

New in version 3.0.

tables.parameters.DRIVER_CORE_BACKING_STORE = 1

Enables backing store for the core driver.

With the H5FD_CORE driver, if the DRIVER_CORE_BACKING_STORE is set to 1 (True), the file contents are flushed to a file with the same name as this core file when the file is closed or access to the file is terminated in memory.

The application is allowed to open an existing file with H5FD_CORE driver. In that case, if the DRIVER_CORE_BACKING_STORE is set to 1 and the flags for tables.open_file() is set to H5F_ACC_RDWR, any change to the file contents are saved to the file when the file is closed. If backing_store is set to 0 and the flags for tables.open_file() is set to H5F_ACC_RDWR, any change to the file contents will be lost when the file is closed. If the flags for tables.open_file() is set to H5F_ACC_RDONLY, no change to the file is allowed either in memory or on file.

New in version 3.0.

tables.parameters.DRIVER_CORE_IMAGE = None

String containing an HDF5 file image.

If this oprion is passed to the tables.open_file() function then the returned file object is set up using the specified image.

A file image can be retrieved from an existing (and opened) file object using the tables.File.get_file_image() method.

Note

requires HDF5 >= 1.8.9.

New in version 3.0.

tables.parameters.DRIVER_SPLIT_META_EXT = '-m.h5'

The extension for the metadata file used by the H5FD_SPLIT driver.

If this option is passed to the tables.openFile() function along with driver=’H5FD_SPLIT’, the extension is appended to the name passed as the first parameter to form the name of the metadata file. If the string ‘%s’ is used in the extension, the metadata file name is formed by replacing ‘%s’ with the name passed as the first parameter instead.

New in version 3.1.

tables.parameters.DRIVER_SPLIT_RAW_EXT = '-r.h5'

The extension for the raw data file used by the H5FD_SPLIT driver.

If this option is passed to the tables.openFile() function along with driver=’H5FD_SPLIT’, the extension is appended to the name passed as the first parameter to form the name of the raw data file. If the string ‘%s’ is used in the extension, the raw data file name is formed by replacing ‘%s’ with the name passed as the first parameter instead.

New in version 3.1.