This section details some of the configuration options and some of the environment variables that can be used to tune for optimal performance when your system already has a version of nGraph installed with one or more of our supported Working with Backends.
Regardless of the framework, after the Build and Test step, a good place to start usually involves making the libraries available to the framework. On Linux* systems built on Intel® Architecture, that command tends to looks something like:
export NGRAPH_CPP_BUILD_PATH=path/to/ngraph_dist/ export LD_LIBRARY_PATH=path/to/ngraph_dist/lib/
If you’re working with the Python API, the following command may be useful:
python3 -c "import ngraph as ng; print('nGraph version: ',ng.__version__)";
FMV stands for Function Multi-Versioning, and it can also provide a number of generic ways to patch or bring architecture-based optimizations to the Operating System that is handling your ML environment. See the GCC wiki for details.
If your nGraph build is a Neural Network configured on Clear Linux* OS for Intel® Architecture, and it includes at least one older CPU, the following article may be helpful.
Before tweaking various environment variables, be aware that how the computation
gets executed depends on the data layout that the model is using.
NCHW are common layouts in Deep Learning models. Your ultimate
runtime can vary greatly – even when all other factors are exactly the same –
when this detail is overlooked.
For CPU (and most cuDNN) backends, the preferred layout is currently
N – Number of images per batch
C – Channel of the image (expressed as a number like 3 for RGB and 1 for grayscale)
H – Height of the image
W – Width of the image
Intel® MKL-DNN is automatically enabled as part of an nGraph default build; you do not need to add it separately or as an additional component to be able to use these configuration settings.
KMP_BLOCKTIMESets the time, in milliseconds, that a thread should wait after completing the execution of a parallel region, before sleeping.
KMP_AFFINITYEnables the runtime library to bind threads to physical processing units. A useful article that explains more about how to use this option for various CPU backends is here: https://web.archive.org/web/20190401182248/https://www.nas.nasa.gov/hecc/support/kb/Using-Intel-OpenMP-Thread-Affinity-for-Pinning_285.html
true) or disables (
false) the printing of OpenMP* runtime library environment variables during program execution.
OMP_NUM_THREADSSpecifies the number of threads to use.
The list below includes recommendations on data layout, parameters, and application configuration to achieve best performance running DNN workloads on Intel® Xeon® (CPU processor) systems.
The number of threads set by
OMP_NUM_THREADS ought not exceed the number of
physical cores. The threads should be pinned to their respective physical cores
and activated as follows:
Buffer pointers should be aligned on 64-byte boundaries. NUMA policy should be
configured for local memory allocation (
When running inference, or training for forward-propagation and weight updates, for best performance:
the number of input channels should be 1, 3, or a multiple of SIMD-width (8 for AVX2 systems, 16 for AVX512 systems).
the number of output channels should be a multiple of SIMD-width (8 for AVX2 systems, 16 for AVX512 systems).
When training backward propagation, the number of input and output channels should be a multiple of SIMD-width (8 for AVX2 systems, 16 for AVX512 systems),
padding should not exceed \(0.5x\) where \(x\) is the kernel size.
kernel width should be less than 14.
The best resource for this configuration option is the Intel® OpenMP* docs
at the following link: Intel OpenMP documentation.
defaults to the number of logical cores. To check the number of cores on your
system, you can run the following on the command-line to see the details
of your CPU:
Some frameworks, like TensorFlow*, use these settings to improve performance; however, they are often not sufficient for optimal performance. Framework-based adjustments cannot access the underlying NUMA configuration in multi-socket Intel® Xeon® processor-based platforms, which is a key requirement for many kinds of inference-engine computations. See the next section on NUMA performance to learn more about this performance feature available to systems utilizing nGraph.
NUMA stands for Non-Uniform Memory Access. It indicates how each CPU can access memory attached to each socket.
Without the “knowledge” of CPU socket and NUMA configuration, a simple thread affinity (as in the case of thread pool) does not lead to optimal performance. In fact, it can sometimes prohibitively decrease throughput; a core from socket 0 might have to continually access cache lines from the memory bank of socket 1, increasing bandwidth demands on the Intel® Ultra-Path Interconnect (Intel® UPI). This situation is exacerbated with larger number of sockets found in 4, 8, and 16-socket systems. We believe that users need to be aware of system level optimizations in addition to framework specific configuration parameters to achieve the best performance for NN workloads on CPU platforms. The nGraph Compiler stack runs on transformers handled by Intel® Architecture (IA), and thus can make more efficient use of the underlying hardware.