Environment Variables¶
This page lists all environment variables that can modify dtFFT behavior at runtime, providing users with granular control over logging, performance measurement, and data transposition strategies.
Most of these variables override settings specified in the dtfft_config_t structure, allowing users to adjust configurations without modifying code.
DTFFT_ENABLE_LOG¶
Enables logging within the dtFFT library. By default, the library runs silently with no output. When enabled, it provides detailed insights into internal processes, aiding analysis and debugging.
Purpose¶
Logging enables monitoring of key operations, including:
Execution Times During Autotune: Logs timing data for autotuning stages.
Accepted Values¶
Type: Integer
Supported Values::
0(disabled)1(enabled)
Default:
0
DTFFT_MEASURE_WARMUP_ITERS¶
Defines the number of warmup iterations performed when effort exceeds DTFFT_ESTIMATE. Warmup iterations ensure stable performance measurements in parallel environments.
Purpose¶
Warmup iterations stabilize system performance by preloading caches, establishing communication channels, and mitigating initial overhead. This is crucial for accurate benchmarking, especially in distributed setups, preventing skewed results from cold starts.
Accepted Values¶
Type: Non-negative integer
Recommended Range: 2–10 (values > 0 are advised for reliable results)
Default:
2
DTFFT_MEASURE_ITERS¶
Specifies the number of measurement iterations for transposition and data exchange when effort exceeds DTFFT_ESTIMATE. Multiple iterations enhance measurement reliability.
Purpose¶
Single measurements may be inconsistent due to system noise or cache effects. By repeating transpositions, dtFFT averages performance data, ensuring robust selection of optimal backends or kernel during autotuning.
Accepted Values¶
Type: Positive integer
Recommended Range: 5–20 (values > 1 balance accuracy and runtime)
Default:
5
DTFFT_PLATFORM¶
Specifies the execution platform for dtFFT plans.
This environment variable allows users to override the platform set via the dtfft_config_t structure,
taking precedence over API configuration.
Purpose¶
The DTFFT_PLATFORM variable provides a flexible way to control whether dtFFT executes on the host (CPU) or a CUDA-enabled GPU
without modifying code or API calls. It ensures that runtime platform selection aligns with user preferences or system capabilities,
prioritizing environment settings over programmatic defaults.
Accepted Values¶
Type: String
Supported Values:
host: Execute on the host (CPU).cuda: Execute on a CUDA device (GPU).
Default:
host
Note
Case-insensitive (e.g.,
HOSTis equivalent tohost).Only applicable in builds with CUDA support (
DTFFT_WITH_CUDAdefined). In non-CUDA builds, it is ignored, and execution defaults to the host.If an unsupported value is provided, it is silently ignored, and the default (
host) is used.
DTFFT_BACKEND¶
Specifies the backend used by dtFFT for data transposition and communication when executing plans.
This environment variable allows users to override the backend selected through the dtfft_config_t structure,
taking precedence over API configuration.
Purpose¶
The DTFFT_BACKEND variable enables users to select a specific backend for optimizing data movement and computation in dtFFT plans.
Different backends offer varying performance characteristics depending on the system configuration, workload, and MPI implementation,
allowing fine-tuned control over execution without modifying code.
Accepted Values¶
Type: String
Supported Values:
mpi_dt: Backend using MPI datatypes.mpi_p2p: MPI peer-to-peer backend.mpi_a2a: MPI backend usingMPI_Alltoallv.mpi_p2p_pipe: Pipelined MPI peer-to-peer backend with overlapping data transfers and unpacking.mpi_p2p_sched: MPI peer-to-peer backend with scheduled communication.mpi_p2p_fused: Fused MPI peer-to-peer backend that integrates packing into the communication pipeline.mpi_p2p_compressed: Extension of Fused MPI P2P backend that uses data compression before communication.mpi_rma: MPI RMA backend that uses MPI_Rget for data transfers.mpi_rma_pipe: Pipelined MPI RMA backend with overlapping data transfers and unpacking.mpi_rma_fused: Fused MPI RMA backend that integrates packing into the communication pipeline.mpi_rma_compressed: Extension of Fused MPI RMA backend that uses data compression before communication.nccl: NCCL backend.nccl_pipe: Pipelined NCCL backend with overlapping data transfers and unpacking.nccl_compressed: NCCL backend that performs compression before data exchange and decompression after.cufftmp: cuFFTMp backend.cufftmp_pipe: cuFFTMp backend that uses additional buffer to avoid extra copy and gain performance.adaptive: Adaptive backend that selects best backend for each transpose/reshape operation during plan creation.
Default: When built with CUDA Support:
ncclif NCCL is available in the library build; otherwise,mpi_p2p. When built without CUDA Support:mpi_dt.
Note
Case-insensitive (e.g.,
MPI_DTis equivalent tompi_dt).If an unsupported value is provided, it is silently ignored, and the default backend (
ncclormpi_p2p, depending on build) is used.Availability of some backends (e.g.,
nccl,cufftmp) depends on additional library support (e.g., NCCL, cuFFTMp) during compilation.
DTFFT_RESHAPE_BACKEND¶
Specifies the backend used by dtFFT specifically for data reshaping operations (converting between pencils and bricks).
This environment variable allows users to override the reshape backend selected through the dtfft_config_t structure,
taking precedence over API configuration.
Purpose¶
The DTFFT_RESHAPE_BACKEND variable enables users to independently control the backend used for reshape operations,
which may have different performance characteristics than regular FFT transpositions.
This allows fine-tuning of reshape performance without affecting the main FFT backend.
Accepted Values¶
Type: String
Supported Values: Same as
DTFFT_BACKEND(see above)Default: Same as
DTFFT_BACKEND
Note
GPU Compatibility Requirement: When working on GPU (platform is
cuda), the reshape backend must be compatible with the main backend’s communication library. Specifically:If
DTFFT_BACKENDuses NCCL (e.g.,ncclornccl_pipe), thenDTFFT_RESHAPE_BACKENDmust also use NCCL.If
DTFFT_BACKENDuses cuFFTMp (e.g.,cufftmporcufftmp_pipe), thenDTFFT_RESHAPE_BACKENDmust also use cuFFTMp.MPI backends can be mixed with each other on GPU, but cannot be mixed with NCCL or cuFFTMp backends.
On CPU (platform is
host), any backend combination is allowed.
DTFFT_NCCL_BUFFER_REGISTER¶
Specifies whether to enable buffer registration for NCCL operations. When enabled, NCCL buffers are registered, which can improve performance for certain workloads.
Purpose¶
Buffer registration can reduce the overhead of memory operations in NCCL by pre-registering memory regions. This is particularly useful for workloads with repeated communication patterns. However, in some cases, disabling registration may be beneficial, depending on the specific system configuration or workload characteristics.
Accepted Values¶
Type: Integer
Accepted Values:
0: Disable NCCL buffer registration.1: Enable NCCL buffer registration.
Default:
1
DTFFT_ENABLE_Z_SLAB¶
Specifies whether to enable Z-slab optimization for dtFFT plans.
When enabled, Z-slab optimization reduces network data transfers by employing a two-dimensional FFT algorithm.
Purpose¶
Z-slab optimization is designed to improve performance for plans decomposed as NX × NY × NZ / P.
Disabling it may resolve issues like DTFFT_ERROR_VKFFT_R2R_2D_PLAN or improve performance if the underlying 2D FFT implementation is suboptimal.
Accepted Values¶
Type: Integer
Accepted Values:
0: Disable Z-slab optimization.1: Enable Z-slab optimization.
Default:
1
DTFFT_ENABLE_Y_SLAB¶
Specifies whether to enable Y-slab optimization for dtFFT plans.
When enabled, Y-slab optimization reduces network data transfers by employing a two-dimensional FFT algorithm.
Purpose¶
Y-slab optimization is designed to improve performance for plans decomposed as NX × NY / P × NZ.
Disabling it may resolve issues like DTFFT_ERROR_VKFFT_R2R_2D_PLAN or improve performance if the underlying 2D FFT implementation is suboptimal.
Accepted Values¶
Type: Integer
Accepted Values:
0: Disable Y-slab optimization.1: Enable Y-slab optimization.
Default:
0
DTFFT_ENABLE_MPI_DT¶
Specifies whether to enable MPI datatype backend when effort is DTFFT_PATIENT.
When enabled, the MPI datatype backend is tested during autotuning.
Purpose¶
The MPI datatype backend is a simple and robust method for data transposition using MPI derived datatypes. However, it may not be the most efficient option for large-scale systems or specific data layouts.
Accepted Values¶
Type: Integer
Accepted Values:
0: Disable MPI datatype backend.1: Enable MPI datatype backend.
Default:
1
DTFFT_ENABLE_MPI¶
Specifies whether to enable MPI-based backends for dtFFT when effort is DTFFT_PATIENT.
When enabled, MPI backends (e.g., MPI P2P) are tested during autotuning.
Purpose¶
The following applies only to CUDA builds: MPI backends are useful for distributed GPU systems but may cause GPU memory leaks in certain OpenMPI versions. Disabling this option can prevent such issues.
Accepted Values¶
Type: Integer
Accepted Values:
0: Disable MPI-based backends.1: Enable MPI-based backends.
Default:
0
DTFFT_ENABLE_NCCL¶
Specifies whether to enable NCCL backends when effort is DTFFT_PATIENT.
When enabled, NCCL backends are tested during autotuning.
Purpose¶
NCCL backends are optimized for GPU-to-GPU communication and can significantly improve performance in multi-GPU systems.
Accepted Values¶
Type: Integer
Accepted Values:
0: Disable NCCL backends.1: Enable NCCL backends.
Default:
1
Note
Only applicable in builds with CUDA support (
DTFFT_WITH_CUDAdefined) and when the execution platform is set tocuda(via DTFFT_PLATFORM ordtfft_config_t).
DTFFT_ENABLE_NVSHMEM¶
Specifies whether to enable NVSHMEM backends when effort is DTFFT_PATIENT.
When enabled, NVSHMEM backends are tested during autotuning.
Purpose¶
NVSHMEM backends provide efficient communication for GPU clusters, leveraging shared memory capabilities.
Accepted Values¶
Type: Integer
Accepted Values:
0: Disable NVSHMEM backends.1: Enable NVSHMEM backends.
Default:
1
Note
Only applicable in builds with CUDA support (
DTFFT_WITH_CUDAdefined) and when the execution platform is set tocuda(via DTFFT_PLATFORM ordtfft_config_t).
DTFFT_ENABLE_PIPE¶
Specifies whether to enable pipelined backends when effort is DTFFT_PATIENT.
When enabled, pipelined backends (e.g., overlapping data copy and unpack) are tested during autotuning.
Purpose¶
Pipelined backends improve performance by overlapping communication and computation, but they require additional internal buffers.
Accepted Values¶
Type: Integer
Accepted Values:
0: Disable pipelined backends.1: Enable pipelined backends.
Default:
1
DTFFT_ENABLE_RMA¶
Specifies whether to enable RMA backends when effort is DTFFT_PATIENT.
When enabled, RMA backends are tested during autotuning.
Purpose¶
RMA backends provide efficient communication for distributed systems using Remote Memory Access.
Accepted Values¶
Type: Integer
Accepted Values:
0: Disable RMA backends.1: Enable RMA backends.
Default:
1
DTFFT_ENABLE_FUSED¶
Specifies whether to enable fused backends when effort is DTFFT_PATIENT.
When enabled, fused backends are tested during autotuning.
Purpose¶
Fused backends improve the pipeline algorithm by integrating packing operations into the communication pipeline for enhanced performance. They utilize a round-robin communication schedule to optimize all-to-all data exchanges.
Accepted Values¶
Type: Integer
Accepted Values:
0: Disable fused backends.1: Enable fused backends.
Default:
1
DTFFT_ENABLE_COMPRESSED¶
Controls whether compressed backends are enabled during autotuning when effort is DTFFT_PATIENT or DTFFT_EXHAUSTIVE.
Purpose¶
Compressed backends use data compression techniques to reduce the amount of data transferred during transpositions, potentially improving performance for certain workloads. However, compression introduces additional computational overhead and may not always provide benefits.
Only fixed-rate compression can be used during autotuning, since it provides predictable performance characteristics and does not require data-dependent decisions at runtime.
To enable compressed backends during autotuning, set this option to 1, set compression type to DTFFT_COMPRESSION_FIXED_RATE and provide desired compression rate.
Accepted Values¶
Type: Integer
Accepted Values:
0: Disable compressed backends during autotuning (default).1: Enable compressed backends during autotuning.
Default:
0
DTFFT_ENABLE_KERNEL_AUTOTUNE¶
Controls whether to enable kernel optimization when effort is below DTFFT_EXHAUSTIVE.
When enabled, dtFFT tries to optimize kernel launch parameters during plan creation.
Purpose¶
Kernel optimization is always enabled for DTFFT_EXHAUSTIVE effort level.
Setting this to 1 enables kernel optimization for lower effort levels (DTFFT_ESTIMATE, DTFFT_MEASURE, DTFFT_PATIENT).
This may increase plan creation time but can improve runtime performance.
Since kernel optimization is performed without data transfers, the time increase is usually minimal.
Accepted Values¶
Type: Integer
Accepted Values:
0: Disable kernel optimization for effort levels belowDTFFT_EXHAUSTIVE.1: Enable kernel optimization for all effort levels.
Default:
0
Note
Only applicable in builds with CUDA support (
DTFFT_WITH_CUDAdefined) and when the execution platform is set tocuda(via DTFFT_PLATFORM ordtfft_config_t).
DTFFT_ENABLE_FOURIER_RESHAPE¶
Controls whether dtFFT should reshape data from pencils to bricks and vice versa in Fourier space during execute() calls.
Purpose¶
By default, dtFFT keeps data in pencil layout throughout the FFT process to minimize data transpositions and maximize performance. When this option is enabled, dtFFT performs additional reshaping operations to ensure data is in brick layout in Fourier space. This can be useful if you need to perform operations on the data in Fourier space between forward and backward transforms.
However, enabling this feature requires additional data transpositions, which will reduce overall performance. Only enable this option if your application specifically requires brick layout in Fourier space.
Accepted Values¶
Type: Integer
Accepted Values:
0: Keep data in pencil layout in Fourier space (default, better performance).1: Reshape data to brick layout in Fourier space (reduced performance, but provides brick layout).
Default:
0
DTFFT_TRANSPOSE_MODE¶
Specifies at which stage the local transposition is performed during global exchange.
Purpose¶
By default, dtFFT performs local transposition before data exchange (packing) by executing single computationally intensive kernel. This is efficient for most cases, but in some scenarios, performing local transposition after data exchange (unpacking) may yield better performance. This variable allows users to select the preferred transpose mode based on their specific use case and system characteristics.
Accepted Values¶
Type: String
Supported Values:
pack: Perform local transposition before data exchange (default).unpack: Perform local transposition after data exchange.
Default:
pack
Note
This setting is used only if
effortis less thanDTFFT_EXHAUSTIVE. ForDTFFT_EXHAUSTIVE, the library autotunes the best transpose mode automatically.
DTFFT_ACCESS_MODE¶
Specifies the memory access pattern (write-aligned vs read-aligned) for local transposition in Generic backends.
Purpose¶
This variable controls the loop order during local data transposition. Write-aligned access (contiguous writes, scattered reads) is generally faster on CPUs due to better cache line utilization and avoiding false sharing in some cases. Read-aligned access (contiguous reads, scattered writes) might be beneficial on certain architectures or memory subsystems.
Accepted Values¶
Type: String
Supported Values:
write: Optimize for contiguous write access (default).read: Optimize for contiguous read access.
Default:
write
Note
This variable applies only to Host (CPU) execution.