Parallelism

Concurrency

Replica parallelism can be configured with the following fields in the predictor configuration:

  • processes_per_replica (default: 1): Each replica runs a web server with processes_per_replica processes. For APIs running with multiple CPUs per replica, using 1-3 processes per unit of CPU generally leads to optimal throughput. For example, if cpu is 2, a value between 2 and 6 processes_per_replica is reasonable. The optimal number will vary based on the workload's characteristics and the CPU compute request for the API.

  • threads_per_process (default: 1): Each process uses a thread pool of size threads_per_process to process requests. For applications that are not CPU intensive such as high I/O (e.g. downloading files), GPU-based inference, or Inferentia-based inference, increasing the number of threads per process can increase throughput. For CPU-bound applications such as running your model inference on a CPU, using 1 thread per process is recommended to avoid unnecessary context switching. Some applications are not thread-safe, and therefore must be run with 1 thread per process.

processes_per_replica * threads_per_process represents the total number of requests that your replica can work on concurrently. For example, if processes_per_replica is 2 and threads_per_process is 2, and the replica was hit with 5 concurrent requests, 4 would immediately begin to be processed, and 1 would be waiting for a thread to become available. If the replica were hit with 3 concurrent requests, all three would begin processing immediately.

Server-side batching

Server-side batching is the process of aggregating multiple real-time requests into a single batch inference, which increases throughput at the expense of latency. Inference is triggered when either a maximum number of requests have been received, or when a certain amount of time has passed since receiving the first request, whichever comes first. Once a threshold is reached, inference is run on the received requests and responses are returned individually back to the clients. This process is transparent to the clients.

The TensorFlow Predictor allows for the use of the following 2 fields in the server_side_batching configuration:

  • max_batch_size: The maximum number of requests to aggregate before running inference. This is an instrument for controlling throughput. The maximum size can be achieved if batch_interval is long enough to collect max_batch_size requests.

  • batch_interval: The maximum amount of time to spend waiting for additional requests before running inference on the batch of requests. If fewer than max_batch_size requests are received after waiting the full batch_interval, then inference will run on the requests that have been received. This is an instrument for controlling latency.

In order to use server-side batching, the model's graph must be built such that batches can be accepted as input/output. The following is an example of how the input x and the output y of the graph could be shaped to be compatible with server-side batching:

batch_size = None
sample_shape = [340, 240, 3] # i.e. RGB image
output_shape = [1000] # i.e. image labels
‚Äč
with graph.as_default():
# ...
x = tf.placeholder(tf.float32, shape=[batch_size] + sample_shape, name="input")
y = tf.placeholder(tf.float32, shape=[batch_size] + output_shape, name="output")
# ...

Optimization

When optimizing for both throughput and latency, you will likely want keep the max_batch_size to a relatively small value. Even though a higher max_batch_size with a low batch_interval (when there are many requests coming in) can offer a significantly higher throughput, the overall latency could be quite large. The reason is that for a request to get back a response, it has to wait until the entire batch is processed, which means that the added latency due to the batch_interval can pale in comparison. For instance, let's assume that a single prediction takes 50ms, and that when the batch size is set to 128, the processing time for a batch is 1280ms (i.e. 10ms per sample). So while the throughput is now 5 times higher, it takes 1280ms + batch_interval to get back a response (instead of 50ms). This is the trade-off with server-side batching.

When optimizing for maximum throughput, a good rule of thumb is to follow these steps:

  1. Determine the maximum throughput of one API replica when server_side_batching is not enabled (same as if max_batch_size were set to 1). This can be done with a load test (make sure to set max_replicas to 1 to disable autoscaling).

  2. Determine the highest batch_interval with which you are still comfortable for your application. Keep in mind that the batch interval is not the only component of the overall latency - the inference on the batch and the pre/post processing also have to occur.

  3. Multiply the maximum throughput from step 1 by the batch_interval from step 2. The result is a number which you can assign to max_batch_size.

  4. Run the load test again. If the inference fails with that batch size (e.g. due to running out of GPU or RAM memory), then reduce max_batch_size to a level that works (reduce batch_interval by the same factor).

  5. Use the load test to determine the peak throughput of the API replica. Multiply the observed throughput by the batch_interval to calculate the average batch size. If the average batch size coincides with max_batch_size, then it might mean that the throughput could still be further increased by increasing max_batch_size. If it's lower, then it means that batch_interval is triggering the inference before max_batch_size requests have been aggregated. If modifying both max_batch_size and batch_interval doesn't improve the throughput, then the service may be bottlenecked by something else (e.g. CPU, network IO, processes_per_replica, threads_per_process, etc).

An example of server-side batching for the TensorFlow Predictor that has been benchmarked is found in ResNet50 in TensorFlow.