AIR builds its training data pipeline on Ray Datasets <datasets>, which is a scalable, framework-agnostic data loading and preprocessing library. Datasets enables AIR to seamlessly load data for local and distributed training with Train.
This page describes how to setup and configure these datasets in Train under different scenarios and scales.
The following figure illustrates a simple Ray AIR training job that (1) loads parquet data from S3, (2) applies a simple user-defined function to preprocess batches of data, and (3) runs an AIR Trainer with the given dataset and preprocessor.
Let's walk through the stages of what happens when Trainer.fit() is called.
Preprocessing: First, AIR will fit the preprocessor (e.g., compute statistics) on the "train" dataset, and then transform all given datasets with the fitted preprocessor. This is done by calling prep.fit_transform() on the train dataset passed to the Trainer, followed by prep.transform() on remaining datasets.
Training: Then, AIR passes the preprocessed dataset to Train workers (Ray actors) launched by the Trainer. Each worker calls get_dataset_shard to get a handle to its assigned data shard, and then calls one of iter_batches, iter_torch_batches, or iter_tf_batches to loop over the data.
The following is a simple example of how to configure ingest for a dummy TorchTrainer. Below, we are passing a small tensor dataset to the Trainer via the datasets argument. In the Trainer's train_loop_per_worker, we access the preprocessed dataset using get_dataset_shard().
Bulk Ingest
By default, AIR loads all datasets into the Ray object store at the start of training. This provides the best performance if the cluster can fit the datasets entirely in memory, or if the preprocessing step is expensive to run more than once.
doc_code/air_ingest.py
You should use bulk ingest when:
- you have enough memory to fit data blocks in cluster object store;
- your preprocessing step is expensive per each epoch; and
- you want best performance when both or either the above conditions are met.
Streaming Ingest (experimental)
In streaming ingest mode, get_dataset_shard returns a DatasetPipeline pipeline that can be used to read data in a streaming way. To enable streaming ingest, set use_stream_api=True in the dataset config.
By default, this will tell AIR to load windows of 1GiB of data into memory at a time. Performance can be increased with larger window sizes, which can be adjusted using the stream_window_size config. A reasonable stream window size is something like 20% of available object store memory.
doc_code/air_ingest.py
Use streaming ingest when:
- you have large datasets that don't fit into memory;
- you want to process small chunks or blocks per window;
- you can use small windows with small data blocks minimizing or avoiding memory starvation or OOM errors; and
- your preprocessing step is not a bottleneck or not an expensive operation since it's re-executed on each pass over the data.
Shuffling or data randomization is important for training high-quality models. By default, AIR will randomize the order the data files (blocks) are read from. AIR also offers options for further randomizing data records within each file:
Local Shuffling
Local shuffling is the recommended approach for randomizing data order. To use local shuffle, simply specify a non-zero local_shuffle_buffer_size as an argument to iter_batches(). The iterator will then use a local buffer of the given size to randomize record order. The larger the buffer size, the more randomization will be applied, but it will also use more memory.
See ds.iter_batches() <ray.data.Dataset.iter_batches> for more details.
doc_code/air_ingest.py
You should use local shuffling when:
- a small in-memory buffer provides enough randomization; or
- you want the highest possible ingest performance; or
- your model is not overly sensitive to shuffle quality
Global Shuffling (slower)
Global shuffling provides more uniformly random (decorrelated) samples and is carried out via a distributed map-reduce operation. This higher quality shuffle can often lead to more precision gain per training step, but it is also an expensive distributed operation and will decrease the ingest throughput. As long as the shuffled ingest throughput matches or exceeds the model training (forward pass, backward pass, gradient sync) throughput, this higher-quality shuffle shouldn't slow down the overall training.
If global shuffling is causing the ingest throughput to become the training bottleneck, local shuffling may be a better option.
doc_code/air_ingest.py
You should use global shuffling when:
- you suspect high-quality shuffles may significantly improve model quality; and
- absolute ingest performance is less of a concern
You can use the DatasetConfig object to configure how Datasets are preprocessed and split across training workers. Each DataParallelTrainer has a default _dataset_config class field. It is a mapping from dataset names to DatasetConfig objects, and implements the default behavior described in the overview <ingest_basics>:
# The default DataParallelTrainer dataset config, which is inherited
# by sub-classes such as TorchTrainer, HorovodTrainer, etc.
_dataset_config = {
# Fit preprocessors on the train dataset only. Split the dataset
# across workers if scaling_config["num_workers"] > 1.
"train": DatasetConfig(fit=True, split=True),
# For all other datasets, use the defaults (don't fit, don't split).
# The datasets will be transformed by the fitted preprocessor.
"*": DatasetConfig(),
}These configs can be overriden via the dataset_config constructor argument. Here are some examples of configuring Dataset ingest options and what they do:
Example: Split All Datasets
This example shows overriding the split config for the "valid" and "test" datasets. This means that both the valid and test datasets here will be .split() across the training workers.
doc_code/air_ingest.py
Example: Disable Transform on Aux Dataset
This example shows overriding the transform config for the "side" dataset. This means that the original dataset will be returned by .get_dataset_shard("side").
doc_code/air_ingest.py
Datasets uses Ray tasks to execute data processing operations. These tasks use CPU resources in the cluster during execution, which may compete with resources needed for Training.
Unreserved CPUs
By default, Dataset tasks use cluster CPU resources for execution. This can sometimes conflict with Trainer resource requests. For example, if Trainers allocate all CPU resources in the cluster, then no Datasets tasks can run.
./doc_code/air_ingest.py
Unreserved CPUs work well when:
- you are running only one Trainer and the cluster has enough CPUs; or
- your Trainers are configured to use GPUs and not CPUs
Using Reserved CPUs (experimental)
The _max_cpu_fraction_per_node option can be used to exclude CPUs from placement group scheduling. In the below example, setting this parameter to 0.8 enables Tune trials to run smoothly without risk of deadlock by reserving 20% of node CPUs for Dataset execution.
./doc_code/air_ingest.py
You should use reserved CPUs when:
- you are running multiple concurrent CPU Trainers using Tune; or
- you want to ensure predictable Datasets performance
Warning
_max_cpu_fraction_per_node is experimental and not currently recommended for use with autoscaling clusters (scale-up will not trigger properly).
Data ingest problems can be challenging to debug when combined in a full training pipeline. To isolate data ingest issues from other possible training problems, we provide the ray.air.util.check_ingest.DummyTrainer utility class that can be used to debug ingest problems. Let's walk through using DummyTrainer to understand and resolve an ingest misconfiguration.
First, let's create a synthetic in-memory dataset and setup a simple preprocessor pipeline. For this example, we'll run it on a 3-node cluster with m5.4xlarge nodes. In practice we might want to use a single machine to keep data local, but we'll use a cluster for illustrative purposes.
doc_code/air_ingest.py
Next, we instantiate and fit a DummyTrainer with a single training worker and no GPUs. You can customize these parameters to simulate your use training use cases (e.g., 16 trainers each with GPUs enabled).
doc_code/air_ingest.py
Let's walk through the output. First, the job starts and executes preprocessing. You can see that the preprocessing runs in 6.8s below. The dataset stats for the preprocessing is also printed:
Starting dataset preprocessing
Preprocessed datasets in 6.874227493000035 seconds
Preprocessor Chain(preprocessors=(BatchMapper(fn=<lambda>), BatchMapper(fn=<lambda>)))
Preprocessor transform stats:
Stage 1 read->map_batches: 100/100 blocks executed in 4.57s
* Remote wall time: 120.68ms min, 522.36ms max, 251.53ms mean, 25.15s total
* Remote cpu time: 116.55ms min, 278.08ms max, 216.38ms mean, 21.64s total
* Output num rows: 500 min, 500 max, 500 mean, 50000 total
* Output size bytes: 102400128 min, 102400128 max, 102400128 mean, 10240012800 total
* Tasks per node: 16 min, 48 max, 33 mean; 3 nodes used
Stage 2 map_batches: 100/100 blocks executed in 2.22s
* Remote wall time: 89.07ms min, 302.71ms max, 175.12ms mean, 17.51s total
* Remote cpu time: 89.22ms min, 207.53ms max, 137.5ms mean, 13.75s total
* Output num rows: 500 min, 500 max, 500 mean, 50000 total
* Output size bytes: 102400128 min, 102400128 max, 102400128 mean, 10240012800 total
* Tasks per node: 30 min, 37 max, 33 mean; 3 nodes usedWhen the train job finishes running, it will print out some more statistics.
P50/P95/Max batch delay (s) 1.101227020500005 1.120024863100042 1.9424749629999951
Num epochs read 1
Num batches read 100
Num bytes read 9765.64 MiB
Mean throughput 116.59 MiB/sLet's break it down:
- Batch delay: Time the trainer spents waiting for the next data batch to be fetched. Ideally this value is as close to zero as possible. If it is too high, Ray may be spending too much time loading data from disk.
- Num epochs read: The number of times the trainer read the dataset during the run.
- Num batches read: The number of batches read.
- Num bytes read: The number of bytes read.
- Mean throughput: The average read throughput.
Finally, we can query memory statistics (this can be run in the middle of a job) to get an idea of how this workload used the object store.
ray memory --stats-onlyAs you can see, this run used 18GiB of object store memory, which was 32% of the total memory available on the cluster. No disk spilling was reported:
--- Aggregate object store stats across all nodes ---
Plasma memory usage 18554 MiB, 242 objects, 32.98% full, 0.17% needed
Objects consumed by Ray tasks: 38965 MiB.So why was the data ingest only 116MiB/s above? That's sufficient for many models, but one would expect faster if the trainer was doing nothing except read the data. Based on the stats above, there was no object spilling, but there was a high batch delay.
We can guess that perhaps AIR was spending too much time loading blocks from other machines, since we were using a multi-node cluster. We can test this by setting prefetch_blocks=10 to prefetch blocks more aggressively and rerunning training.
P50/P95/Max batch delay (s) 0.0006792084998323844 0.0009853049503362856 0.12657493300002898
Num epochs read 47
Num batches read 4700
Num bytes read 458984.95 MiB
Mean throughput 15136.18 MiB/sThat's much better! Now we can see that our DummyTrainer is ingesting data at a rate of 15000MiB/s, and was able to read through many more epochs of training. This high throughput means that all data was able to be fit into memory on a single node.
Once you're happy with the ingest performance of with DummyTrainer with synthetic data, the next steps are to switch to adapting it for your real workload scenario. This involves:
- Scaling the DummyTrainer: Change the scaling config of the DummyTrainer and cluster configuration to reflect your target workload.
- Switching the Dataset: Change the dataset from synthetic tensor data to reading your real dataset.
- Switching the Trainer: Swap the DummyTrainer with your real trainer.
Switching these components one by one allows performance problems to be easily isolated and reproduced.
Memory availability: To maximize ingest performance, consider using machines with sufficient memory to fit the dataset entirely in memory. This avoids the need for disk spilling, streamed ingest, or fetching data across the network. As a rule of thumb, a Ray cluster with fewer but bigger nodes will outperform a Ray cluster with more smaller nodes due to better memory locality.
Autoscaling: We generally recommend first trying out AIR training with a fixed size cluster. This makes it easier to understand and debug issues. Autoscaling can be enabled after you are happy with performance to autoscale experiment sweeps with Tune, etc. We also recommend starting with autoscaling with a single node type. Autoscaling with hetereogeneous clusters can optimize costs, but may complicate performance debugging.
Partitioning: By default, Datasets will create up to 200 blocks per Dataset, or less if there are fewer base files for the Dataset. If you run into out-of-memory errors during preprocessing, consider increasing the number of blocks to reduce their size. To increase the max number of partitions, set the parallelism option when calling ray.data.read_*(). To change the number of partitions at runtime, use ds.repartition(N). As a rule of thumb, blocks should be not more than 1-2GiB each.
When you pass Datasets to a Tuner, Datasets are executed independently per-trial. This could potentially duplicate data reads in the cluster. To share Dataset blocks between trials, call ds = ds.fully_executed() prior to passing the Dataset to the Tuner. This ensures that the initial read operation will not be repeated per trial.