org.nd4j.config

Class ND4JEnvironmentVars

java.lang.Object

org.nd4j.config.ND4JEnvironmentVars

public class ND4JEnvironmentVars
extends Object

Field Summary

Fields

Modifier and Type	Field and Description
static String	BACKEND_DYNAMIC_LOAD_CLASSPATH Applicability: always - but only if an ND4J backend cannot be found/loaded via standard ServiceLoader mechanisms Description: Set this environment variable to a set fully qualified JAR files to attempt to load before failing on not loading a backend.
static String	BACKEND_PRIORITY_CPU Applicability: nd4j-native, when multiple backends are on classpath Description: Defines the priority that the CPU/Native backend should be loaded (or attempt to be loaded).
static String	BACKEND_PRIORITY_GPU Applicability: nd4j-cuda-xx, when multiple backends are on classpath Description: Defines the priority that the CUDA (GPU) backend should be loaded (or attempt to be loaded).
static String	DL4J_VOID_IP Applicability: nd4j-parameter-server Usage: A fallback for determining the local IP the parameter server, if other approaches fail to determine the local IP
static String	ND4J_CUDA_FORCE_SINGLE_GPU Applicability: nd4j-cuda-xx used on multi-GPU systems Description: If set, only a single GPU will be used by ND4J, even if multiple GPUs are available in the system
static String	ND4J_CUDA_MAX_BLOCK_SIZE Applicability: nd4j-cuda-xx Description:
static String	ND4J_CUDA_MAX_DEVICE_ALLOCATION Applicability: nd4j-cuda-xx Description:
static String	ND4J_CUDA_MAX_DEVICE_CACHE Applicability: nd4j-cuda-xx Description:
static String	ND4J_CUDA_MAX_GRID_SIZE Applicability: nd4j-cuda-xx Description:
static String	ND4J_CUDA_MAX_HOST_CACHE Applicability: nd4j-cuda-xx Description:
static String	ND4J_CUDA_MIN_BLOCK_SIZE Applicability: nd4j-cuda-xx Description:
static String	ND4J_CUDA_USE_PREALLOCATION Applicability: nd4j-cuda-xx Description:
static String	ND4J_FALLBACK Applicability: nd4j-native backend Description: Whether build-in BLAS matrix multiplication (GEMM) should be used instead of the native BLAS library such as MKL or OpenBLAS.
static String	ND4J_SKIP_BLAS_THREADS Applicability: nd4j-native backend Description: Skips the setting of the OMP_NUM_THREADS property for ND4J ops.
static String	OMP_NUM_THREADS Applicability: nd4j-native backend Description: Sets the number of OpenMP parallel threads for ND4J native operations (and also native BLAS libraries such as Intel MKL and OpenBLAS).

Method Summary

Methods inherited from class java.lang.Object

clone, equals, finalize, getClass, hashCode, notify, notifyAll, toString, wait, wait, wait

Field Detail

BACKEND_PRIORITY_CPU

public static final String BACKEND_PRIORITY_CPU

Applicability: nd4j-native, when multiple backends are on classpath
Description: Defines the priority that the CPU/Native backend should be loaded (or attempt to be loaded). If this is set to a higher value than BACKEND_PRIORITY_GPU (which has default value 100) the native backend will be loaded in preference to the CUDA backend, when both are on the classpath. Default value: 0

See Also:

Constant Field Values

BACKEND_PRIORITY_GPU

public static final String BACKEND_PRIORITY_GPU

Applicability: nd4j-cuda-xx, when multiple backends are on classpath
Description: Defines the priority that the CUDA (GPU) backend should be loaded (or attempt to be loaded). If this is set to a higher value than BACKEND_PRIORITY_CPU (which has default value 0) the GPU backend will be loaded in preference to the CUDA backend, when both are on the classpath. Default value: 100 - hence by default, the CUDA backend will be loaded when both it and the CPU/native backend are on the classpath

See Also:

Constant Field Values

BACKEND_DYNAMIC_LOAD_CLASSPATH

public static final String BACKEND_DYNAMIC_LOAD_CLASSPATH

Applicability: always - but only if an ND4J backend cannot be found/loaded via standard ServiceLoader mechanisms
Description: Set this environment variable to a set fully qualified JAR files to attempt to load before failing on not loading a backend. JAR files should be semi-colon delimited; i.e., "/some/file.jar;/other/path.jar". This should rarely be required in practice - for example, only in dynamic class loading/dynamic classpath scenarios
For equivalent system property, see ND4JSystemProperties.DYNAMIC_LOAD_CLASSPATH_PROPERTY for the equivalent system property (that will take precidence if both are set)

See Also:

Constant Field Values

OMP_NUM_THREADS

public static final String OMP_NUM_THREADS

Applicability: nd4j-native backend
Description: Sets the number of OpenMP parallel threads for ND4J native operations (and also native BLAS libraries such as Intel MKL and OpenBLAS). By default, this will be set to the number of physical cores (i.e., excluding hyperthreading cores), which usually provides optimal performance. Setting this to a larger value than the number of physical cores (for example, equal to number of logical cores - i.e., setting to 16 on an 8-core + hypethreading processor) - can result in reduced performance
Note that if you have a significant number of parallel Java threads (for example, Spark or ParallelWrapper), or you want to keep some cores free for other programs - you may want to reduce this value.

See Also:

ND4J_SKIP_BLAS_THREADS, Constant Field Values

ND4J_SKIP_BLAS_THREADS

public static final String ND4J_SKIP_BLAS_THREADS

Applicability: nd4j-native backend
Description: Skips the setting of the OMP_NUM_THREADS property for ND4J ops. Note that this property will usually still take effect for native BLAS libraries (MKL, OpenBLAS) even if this property is set

See Also:

Constant Field Values

ND4J_FALLBACK

public static final String ND4J_FALLBACK

Applicability: nd4j-native backend
Description: Whether build-in BLAS matrix multiplication (GEMM) should be used instead of the native BLAS library such as MKL or OpenBLAS. This can have a noticable performance impact for these ops. Note that this is typically only useful as a workaround (or test) for bugs in these underlying native libraries, which are rare (but do occasionally occur on some platforms)

See Also:

Constant Field Values

DL4J_VOID_IP

public static final String DL4J_VOID_IP

Applicability: nd4j-parameter-server
Usage: A fallback for determining the local IP the parameter server, if other approaches fail to determine the local IP

See Also:

Constant Field Values

ND4J_CUDA_MAX_BLOCK_SIZE

public static final String ND4J_CUDA_MAX_BLOCK_SIZE

Applicability: nd4j-cuda-xx
Description:

See Also:

Constant Field Values

ND4J_CUDA_MIN_BLOCK_SIZE

public static final String ND4J_CUDA_MIN_BLOCK_SIZE

Applicability: nd4j-cuda-xx
Description:

See Also:

Constant Field Values

ND4J_CUDA_MAX_GRID_SIZE

public static final String ND4J_CUDA_MAX_GRID_SIZE

Applicability: nd4j-cuda-xx
Description:

See Also:

Constant Field Values

ND4J_CUDA_FORCE_SINGLE_GPU

public static final String ND4J_CUDA_FORCE_SINGLE_GPU

Applicability: nd4j-cuda-xx used on multi-GPU systems
Description: If set, only a single GPU will be used by ND4J, even if multiple GPUs are available in the system

See Also:

Constant Field Values

ND4J_CUDA_USE_PREALLOCATION

public static final String ND4J_CUDA_USE_PREALLOCATION

Applicability: nd4j-cuda-xx
Description:

See Also:

Constant Field Values

ND4J_CUDA_MAX_DEVICE_CACHE

public static final String ND4J_CUDA_MAX_DEVICE_CACHE

Applicability: nd4j-cuda-xx
Description:

See Also:

Constant Field Values

ND4J_CUDA_MAX_HOST_CACHE

public static final String ND4J_CUDA_MAX_HOST_CACHE

Applicability: nd4j-cuda-xx
Description:

See Also:

Constant Field Values

ND4J_CUDA_MAX_DEVICE_ALLOCATION

public static final String ND4J_CUDA_MAX_DEVICE_ALLOCATION

Applicability: nd4j-cuda-xx
Description:

See Also:

Constant Field Values