Automatic Offload Controls

The table below lists the environment variables for Automatic Offload and the functions that cause similar results. See the Intel MKL Developer Reference for detailed descriptions of the functions. To control the division of work between the host CPU and Intel Xeon Phi coprocessors, the environment variables use a fractional measure ranging from zero to one.

Environment Variable	Support Function	Description	Value
`MKL_MIC_ENABLE`	`mkl_mic_enable`	Enables Automatic Offload (AO).	1
`OFFLOAD_DEVICES`	None	`OFFLOAD_DEVICES` is a common setting for Intel MKL and Intel® Compilers. It specifies a list of coprocessors to be used for any offload, including Intel MKL AO. In particular, this setting may help you to configure the environment for an MPI application to run Intel MKL in the AO mode. If this variable is not set, all the coprocessors available on the system are used for AO. You can set this environment variable if AO is enabled by the environment setting or function call. Setting this variable to an empty value is equivalent to completely disabling AO regardless of the value of `MKL_MIC_ENABLE`. After setting this environment variable, Intel MKL support functions and environment variables refer to the specified coprocessors by their indexes in the list, starting with zero. For more information, refer to the Intel® Compiler User and Reference Guides.	A comma-separated list of integers, each ranging from 0 to the largest number of an Intel Xeon Phi coprocessor on the system, with the maximum of 31. Values out of this range are ignored. Moreover, if the list contains any non-integer data, the list is ignored completely as if the environment variable were not set at all. For example, if your system has 4 Intel Xeon Phi coprocessors and the value of the list is 1,3, Intel MKL uses only coprocessors 1 and 3 for AO, and Intel MKL support functions and environment variables refer to these coprocessors as coprocessors 0 and 1.
`OFFLOAD_ENABLE_ORSL`	None	Enables the mode in which Intel MKL and Intel Compilers synchronize their accesses to coprocessors. Set this variable if your application uses both Compiler Assisted and AO but does not implement its own synchronization.	1
`MKL_HOST_WORKDIVISION`	`mkl_mic_set_workdivision`	Specifies the fraction of work for the host CPU to do.	A floating-point number ranging from 0.0 to 1.0. For example, the value could be 0.2 or 0.33. Intel MKL ignores negative values and treats values greater than 1 as 1.0.
`MKL_MIC_WORKDIVISION`	`mkl_mic_set_workdivision`	Specifies the fraction of work to do on all the Intel Xeon Phi coprocessors on the system.	See `MKL_HOST_WORKDIVISION`
`MKL_MIC_<number>_WORKDIVISION`	`mkl_mic_set_workdivision`	Specifies the fraction of work to do on a specific Intel Xeon Phi coprocessor. Here `<number>` is an integer ranging from 0 to the largest number of an Intel Xeon Phi coprocessor on the system, with the maximum of 31. For example, if the system has two Intel Xeon Phi coprocessors, `<number>` can be 0 or 1.	See `MKL_HOST_WORKDIVISION`
`MKL_MIC_MAX_MEMORY`	`mkl_mic_set_max_memory`	Specifies the maximum coprocessor memory reserved for AO computations on all of the Intel Xeon Phi coprocessors on the system. Each process that performs AO computations uses additional coprocessor memory specified by the environment variable.	Memory size in Kilobytes (`K`), megabytes (`M`), gigabytes (`G`), or terabytes (`T`). For example, `MKL_MIC_MAX_MEMORY = 4096M` limits the coprocessor memory reserved for AO computations to 4096 megabytes or 4 gigabytes. Setting `MKL_MIC_MAX_MEMORY = 4G` specifies the same amount of memory in gigabytes.
`MKL_MIC_<number>_MAX_MEMORY`	`mkl_mic_set_max_memory`	Specifies the maximum coprocessor memory reserved for AO computations on a specific Intel Xeon Phi coprocessor on the system. Here `<number>` is an integer ranging from 0 to the largest number of an Intel Xeon Phi coprocessor on the system, with the maximum of 31. For example, if the system has two Intel Xeon Phi coprocessors, `<number>` can be 0 or 1.	Memory size in Kilobytes (`K`), megabytes (`M`), gigabytes (`G`), or terabytes (`T`). For example, `MKL_MIC_MAX_MEMORY = 4096M` limits the coprocessor memory reserved for AO computations to 4096 megabytes or 4 gigabytes. Setting `MKL_MIC_MAX_MEMORY = 4G` specifies the same amount of memory in gigabytes.
`MKL_MIC_REGISTER_MEMORY`	`mkl_mic_register_memory`	Enables/disables the mkl_malloc function running in AO mode to register allocated memory. If AO is disabled, this setting has no effect. Setting this environment variable to 1 may improve performance if the same memory region allocated by mkl_malloc is passed multiple times to Intel MKL functions enabled for AO (for a list of AO enabled functions, see Intel MKL Release Notes).	Desired behavior of mkl_malloc: 0 - not register allocated memory 1 - register allocated memory
`MKL_MIC_RESOURCE_LIMIT`	`mkl_mic_set_resource_limit`	Specifies how much of the computational resources of Intel Xeon Phi coprocessors can be used by the calling process. Use this environment variable if you need to share Intel Xeon Phi coprocessor cores automatically across multiple processes that call Intel MKL in the AO mode. For example, this might be useful in MPI applications. Actual reservation is made during a call to an Intel MKL AO function.	A floating-point number ranging from 0.0 to 1.0. Special values: 0.0 - do not share Intel Xeon Phi coprocessors across multiple processes. Default. MKL_MPI_PPN - enable special fully automated mode for MPI applications. In this mode Intel MKL tries to read the number of MPI processes per node (`ppn`) from the environment variables passed to the process by MPI. If this attempt is successful, Intel MKL sets the actual resource limit to 1.0/`ppn`. Note For Intel® MPI Library, Open MPI, and IBM Platform MPI, Intel MKL automatically detects `ppn`. For other MPI implementations, use the MKL_MPI_PPN environment variable to set `ppn`. Examples: Two 61-core Intel Xeon Phi coprocessors are available on the system. Three processes simultaneously set `MKL_MIC_RESOURCE_LIMIT=0.34` and then call dpotrf in AO mode. As a result, one process receives 40 cores from coprocessor 1, another process receives 40 cores from coprocessor 2, and the remaining process receives cores from each of the two coprocessors for a total of 40 cores. Two 58-core Intel Xeon Phi coprocessors are available on the system. The user makes these settings `MKL_MIC_1_WORKDIVISION=0.0` `MKL_MIC_PPN=4` `MKL_MIC_RESOURCE_LIMIT=MKL_MPI_PPN` and then runs an MPI application with Intel MKL AO calls. As a result, each MPI process receives 14 cores on Intel Xeon Phi coprocessor 0 because the user excluded coprocessor 1 from computations by setting zero workdivision for it.
`MIC_OMP_NUM_THREADS`	`mkl_mic_set_device_num_threads`	Specifies the maximum number of OpenMP* threads to use for AO computations on all the Intel Xeon Phi coprocessors on the system.	An integer greater than 0.
`MIC_<number>_OMP_NUM_THREADS`	`mkl_mic_set_device_num_threads`	Specifies the maximum number of OpenMP threads to use for AO computations on a specific Intel Xeon Phi coprocessor on the system. Here `<number>` is an integer ranging from 0 to the largest number of an Intel Xeon Phi coprocessor on the system, with the maximum of 31. For example, if the system has two Intel Xeon Phi coprocessors, `<number>` can be 0 or 1.	An integer greater than 0.
`OFFLOAD_REPORT`	`mkl_mic_set_offload_report`	`OFFLOAD_REPORT` is a common setting for Intel MKL and Intel® Compilers. It specifies the profiling report level for any offload, including Intel MKL AO. For more information, refer to the Intel® Compiler User and Reference Guides. Note that the `mkl_mic_set_offload_report` function enables you to turn profile reporting on/off at run time but does not change the reporting level.	An integer ranging from 0 to 2: 0 - No reporting, default. 1 - The report includes: The name of the function called in the AO mode. Effective work division. The value of -1 indicates that the hint, that is, the work division specified by the mkl_mic_set_workdivison function or the appropriate MKL_*_WORKDIVISION environment variable was ignored in this function call. The time spent on the host CPU during the call. The time spent on each available Intel Xeon Phi coprocessor during the call. 2 - In addition to the above information, the report includes: The amounts of data transferred to and from each available coprocessor during the call.
`LD_LIBRARY_PATH`	None	Specifies the search path for host-side dynamic libraries.	Must contain the path to host-side Intel MIC Platform Software Stack libraries used by Intel MKL. The default path is /opt/intel/mic/coi/host-linux-release/lib.
`MIC_LD_LIBRARY_PATH`	None	Specifies the search path for coprocessor-side dynamic libraries.	Must contain: The path to coprocessor-side Intel MIC Platform Software Stack libraries used by Intel MKL. The default path is /opt/intel/mic/coi/device-linux-release/lib. The path to Intel MKL coprocessor-side libraries. The default path is `<mkl directory>`/lib/mic.
`MKL_MIC_THRESHOLDS_?GEMM`	None	Specifies matrix size thresholds for ?GEMM computations in the AO mode.	Three comma-separated integers: `M`, `N`, `K`. If this environment variable is set, any call to a ?GEMM function with problem sizes `M`₁, `N`₁, and `K`₁ tries to offload computations only if `M`₁>`M`, `N`₁>`N`, and `K`₁>`K`. This setting is only a hint, and Intel MKL may decide to not offload computations depending on the problem size and environment. Example: To set the thresholds to `M`=2000, `N`=1000, `K`=500 for DGEMM, set `MKL_MIC_THRESHOLDS_DGEMM=2000,1000,500`.

Note

Settings specified by the functions take precedence over the settings specified by the respective environment variables.
Intel MKL interprets the values of MKL_HOST_WORKDIVISION, MKL_MIC_WORKDIVISION, and MKL_MIC_<number>_WORKDIVISION as guidance toward dividing work between coprocessors, but the library may choose a different work division if necessary.
For LAPACK routines, setting the fraction of work to any value other than 0.0 enables the specified processor for AO mode. However Intel MKL LAPACK does not use the value specified to divide the workload. For example, setting the fraction to 0.5 has the same effect as setting the fraction to 1.0.

Automatic Offload Controls

Note

Note

See Also