Advanced Topics

This section covers a few advanced topics that are mentioned in the API documentation.

Note

For typical use cases, the default DALI configuration performs well out of the box, and you do not need to review this section.

Thread Affinity

This functionality allows you to pin DALI threads to the specified CPU. Thread affinity avoids the overhead of worker threads jumping from core to core and improves performance with CPU-heavy workloads. You can set the DALI CPU thread affinity by using the DALI_AFFINITY_MASK environment variable, which is a comma-separated list of CPU IDs that will be assigned to corresponding DALI threads. The number of DALI threads is set during the pipeline construction by the num_threads argument and set_affinity enables thread affinity for the CPU worker threads.

Note

For performance reasons, the hybrid nvidia.dali.ops.ImageDecoder() operator, which is nvJPEG based, creates threads on its own, and these threads are always affined.

In DALI_AFFINITY_MASK, if the number of threads is higher than the number of CPU IDs, the following process is applied:

  1. The threads are assigned to the CPU IDs until all of the CPU IDs from DALI_AFFINITY_MASK are assigned.

  2. For the remaining threads, the CPU IDs from nvmlDeviceGetCpuAffinity will be used.

An example:

num_threads=5
DALI_AFFINITY_MASK=3,5,6,10

This example sets thread 0 to CPU 3, thread 1 to CPU 5, thread 2 to CPU 6, thread 3 to CPU 10, and thread 4 to the CPU ID that is returned by nvmlDeviceGetCpuAffinity.

Memory Consumption

DALI uses the following memory types:

  • Host

  • Host-page-locked

  • GPU

Allocating and freeing GPU and host page-locked (or pinned) memory require device synchronization. As a result, when possible, DALI avoids reallocating these kinds of memory. The buffers that are allocated with these storage types will only grow when the existing buffer is too small to accommodate the requested shape. This strategy reduces the number of total memory management operations and increases the processing speed when the memory requirements become stable and no more allocations are required.

By contrast, ordinary host memory is relatively inexpensive to allocate and free. To reduce host memory consumption, the buffers might shrink when the new requested size is smaller than the fraction of the old size. This is called shrink threshold. It can be adjusted to a value between 0 (never shrink) and 1 (always shrink). The default is 0.9. The value can be controlled by the DALI_HOST_BUFFER_SHRINK_THRESHOLD environmental variable or be set in Python by calling the nvidia.dali.backend.SetHostBufferShrinkThreshold function.

During processing, DALI works on batches of samples. For GPU and some CPU operators, each batch is stored as contiguous memory and is processed at once, which reduces the number of necessary allocations. For some CPU operators that cannot calculate their output size ahead of time, the batch is stored as a vector of separately allocated samples.

For example, if your batch consists of nine 480p images and one 4K image in random order, the contiguous allocation can accommodate all possible combinations of these batches. On the other hand, the CPU batch that is stored as separate buffers needs to keep a 4K allocation for every sample after several iterations. The GPU buffers that keep the operator outputs can grow are as large as the largest possible batch, whereas the non-contiguous CPU buffers can reach the size of the largest sample in the data set multiplied by the number of samples in the batch.

The host and the GPU buffers have a configurable growth factor. If the factor is greater than 1, and to potentially avoid subsequent reallocations. This functionality is disabled by default, and the growth factor is set to 1. The growth factors can be controlled with the DALI_HOST_BUFFER_GROWTH_FACTOR and DALI_DEVICE_BUFFER_GROWTH_FACTOR environmental variables and with the nvidia.dali.backend.SetHostBufferGrowthFactor and nvidia.dali.backend.SetDeviceBufferGrowthFactor Python API functions. For convenience, the DALI_BUFFER_GROWTH_FACTOR environment variable and the nvidia.dali.backend.SetBufferGrowthFactor Python function can be used to set the same growth factor for the host and the GPU buffers.

Operator Buffer Presizing

When you can precisely forecast the memory consumption during a DALI run, this functionality helps you fine tune the processing pipeline. One of the benefits is that the overhead of some reallocations can be avoided.

DALI uses intermediate buffers to pass data between operators in the processing graph. The capacity of this buffers is increased to accommodate new data, but is never reduced. Sometimes, however, even this limited number of allocations might still affect DALI performance. If you know how much memory each operator buffer requires, you can add a hint to preallocate the buffers before the pipeline is first run.

The following parameters are available:

  • The bytes_per_sample pipeline argument, which accepts one value that is used globally across all operators for all buffers.

  • The bytes_per_sample_hint per operator argument, which accepts one value or a list of values.

When one value is provided, it is used for all output buffers for an operator. When a list is provided, each operator output buffer is presized to the corresponding size. To determine the amount of memory output that each operator needs, complete the following tasks:

  1. Create the pipeline by setting enable_memory_stats to True.

  2. Query the pipeline for the operator’s output memory statistics by calling the nvidia.dali.pipeline.Pipeline.executor_statistics() method on the pipeline.

The max_real_memory_size value represents the biggest tensor in the batch for the outputs that allocate memory per sample and not for the entire batch at the time or the average tensor size when the allocation is contiguous. This value should be provided to bytes_per_sample_hint.

Prefetching Queue Depth

The DALI pipeline allows the buffering of one or more batches of data, which is important when the processing time varies from batch to batch. The default prefetch depth is 2. You can change this value by using the prefetch_queue_depth pipeline argument. If the variation is not hidden by the default prefetch depth value, we recommend that you prefetch more data ahead of time.

Note

Increasing queue depth also increases memory consumption.

Running DALI pipeline

DALI pipeline can be run in one of the following ways:

The first API, nvidia.dali.pipeline.Pipeline.run() method completes the following tasks:

  1. Launches the DALI pipeline.

  2. Executes the prefetch iterations if necessary.

  3. Waits until the first batch is ready.

  4. Returns the resulting buffers.

Buffers are marked as in-use until the next call to nvidia.dali.pipeline.Pipeline.run(). This process can be wasteful because the data is usually copied to the DL framework’s native storage objects and DALI pipeline outputs could be returned to DALI for reuse.

The second API, which consists of nvidia.dali.pipeline.Pipeline.schedule_run(), nvidia.dali.pipeline.Pipeline.share_outputs(), and nvidia.dali.pipeline.Pipeline.release_outputs() allows you to explicitly manage the lifetime of the output buffers. The nvidia.dali.pipeline.Pipeline.schedule_run() method instructs DALI to prepare the next batch of data, and, if necessary, to prefetch. If the execution mode is set to asynchronous, this call returns immediately, without waiting for the results. This way, another task can be simultaneously executed. The data batch can be requested from DALI by calling nvidia.dali.pipeline.Pipeline.share_outputs(), which returns the result buffer. If the data batch is not yet ready, DALI will wait for it. The data is ready as soon as the nvidia.dali.pipeline.Pipeline.share_outputs()`() is complete. When the DALI buffers are no longer needed, because data was copied or has already been consumed, call nvidia.dali.pipeline.Pipeline.release_outputs() to return the DALI buffers for reuse in subsequent iterations.

Built-in iterators use the second API to provide convenient wrappers for immediate use in Deep Learning Frameworks. The data is returned in the framework’s native buffers. The iterator’s implementation copies the data internally from DALI buffers and recycles the data by calling nvidia.dali.pipeline.Pipeline.release_outputs().

We recommend that you do not mix the APIs. The APIs follow a different logic for the output buffer lifetime management, and the details of the process are subject to change without notice. Mixing the APIs might result in undefined behavior, such as a deadlock or an attempt to access an invalid buffer.

Sharding

Sharding allows DALI to partition the dataset into nonoverlapping pieces on which each DALI pipeline instance can work. This functionality addresses the issue of having a global and a shared state that allows the distribution of training samples among the ranks. After each epoch, by default, the DALI pipeline advances to the next shard to increase the entropy of the data that is seen by this pipeline. You can alter this behavior by setting the stick_to_shard reader parameter.

This mode of operation, however, leads to problems when the dataset size is not divisible by the number of pipelines used or when the shard size is not divisible by the batch size. To address this issue, and adjust the behavior, you can use the pad_last_batch reader parameter.

This parameter asks the reader to duplicate the last sample in the last batch of a shard, which prevents DALI from reading data from the next shard when the batch doesn’t divide its size. The parameter also ensures that all pipelines return the same number of batches, when one batch is divisible by the batch size but other batches are bigger by one sample. This process pads every shard to the same size, which is a multiple of the batch size.

DALI is used in the Deep Learning Frameworks through dedicated iterators, and these iterators need to be aware of this padding and other reader properties.

Here are the iterator options:

  • fill_last_batch – Determines whether the last batch should be full, regardless of whether the shard size is divisible by the batch size.

  • reader_name - Allows you to provide the name of the reader that drives the iterator and provides the necessary parameters.

    Note

    We recommend that you use this option. With this option, the next two options are excluded and cannot be used.

    This option is more flexible and accurate and takes into account that shard size for a pipeline can differ between epochs when the shards are rotated.

  • size: Displays the size of the shard for an iterator or, if there is more than one shard, the sum of all shard sizes for all wrapped pipelines.

  • last_batch_padded: Determines whether the tail of the data consists of data from the next shard (False) or is duplicated dummy data (True).
    It is applicable when the shard size is not a multiple of the batch size,

Here is the formula to calculate the shard size for a shard ID:

floor((id + 1) * dataset_size / num_shards) - floor(id * dataset_size / num_shards)

When the pipeline advances through the epochs and the reader moves to the next shard, the formula needs to be extended to reflect this change:

floor(((id + epoch_num) % num_shards + 1) * dataset_size / num_shards) - floor(((id + epoch_num) % num_shards) * dataset_size / num_shards)

When the second formula is used, providing a size value once at the beginning of the training works only when the stick_to_shard reader option is enabled and prevents DALI from rotating shards. When this occurs, use the first formula.

To address these challenges, use the reader_name parameter and allow the iterator handle the details.

C++ API

Note

This feature is not officially supported and may change without notice

The C++ API allows you to use DALI as a library from native applications. Refer to the PipelineTest family of tests for more information about how to use this API.