#deep-learning #server #backpressure #batched-prediction

batched-fn

Middleware for serving deep learning models with batched prediction

17 releases

0.2.5 Mar 10, 2024
0.2.4 Mar 14, 2022
0.2.2 Dec 17, 2020
0.2.1 Sep 15, 2020
0.1.7 Mar 31, 2020

#108 in Machine learning

Apache-2.0

29KB
282 lines

batched-fn

Rust server plugin for deploying deep learning models with batched prediction.

Build License Crates Docs


Deep learning models are usually implemented to make efficient use of a GPU by batching inputs together in "mini-batches". However, applications serving these models often receive requests one-by-one. So using a conventional single or multi-threaded server approach will under-utilize the GPU and lead to latency that increases linearly with the volume of requests.

batched-fn is a drop-in solution for deep learning webservers that queues individual requests and provides them as a batch to your model. It can be added to any application with minimal refactoring simply by inserting the batched_fn macro into the function that runs requests through the model.

Features

  • 🚀 Easy to use: drop the batched_fn! macro into existing code.
  • 🔥 Lightweight and fast: queue system implemented on top of the blazingly fast flume crate.
  • 🙌 Easy to tune: simply adjust max_delay and max_batch_size.
  • 🛑 Back pressure mechanism included: just set channel_cap and handle Error::Full by returning a 503 from your webserver.

Examples

Suppose you have a model API that look like this:

// `Batch` could be anything that implements the `batched_fn::Batch` trait.
type Batch<T> = Vec<T>;

#[derive(Debug)]
struct Input {
    // ...
}

#[derive(Debug)]
struct Output {
    // ...
}

struct Model {
    // ...
}

impl Model {
    fn predict(&self, batch: Batch<Input>) -> Batch<Output> {
        // ...
    }

    fn load() -> Self {
        // ...
    }
}

Without batched-fn a webserver route would need to call Model::predict on each individual input, resulting in a bottleneck from under-utilizing the GPU:

use once_cell::sync::Lazy;
static MODEL: Lazy<Model> = Lazy::new(Model::load);

fn predict_for_http_request(input: Input) -> Output {
    let mut batched_input = Batch::with_capacity(1);
    batched_input.push(input);
    MODEL.predict(batched_input).pop().unwrap()
}

But by dropping the batched_fn macro into your code you automatically get batched inference behind the scenes without changing the one-to-one relationship between inputs and outputs:

async fn predict_for_http_request(input: Input) -> Output {
    let batch_predict = batched_fn! {
        handler = |batch: Batch<Input>, model: &Model| -> Batch<Output> {
            model.predict(batch)
        };
        config = {
            max_batch_size: 16,
            max_delay: 50,
        };
        context = {
            model: Model::load(),
        };
    };
    batch_predict(input).await.unwrap()
}

❗️ Note that the predict_for_http_request function now has to be async.

Here we set the max_batch_size to 16 and max_delay to 50 milliseconds. This means the batched function will wait at most 50 milliseconds after receiving a single input to fill a batch of 16. If 15 more inputs are not received within 50 milliseconds then the partial batch will be ran as-is.

Tuning max batch size and max delay

The optimal batch size and delay will depend on the specifics of your use case, such as how big of a batch you can fit in memory (typically on the order of 8, 16, 32, or 64 for a deep learning model) and how long of a delay you can afford. In general you want to set max_batch_size as high as you can, assuming the total processing time for N examples is minimized with a batch size of N, and keep max_delay small relative to the time it takes for your handler function to process a batch.

Implementation details

When the batched_fn macro is invoked it spawns a new thread where the handler will be ran. Within that thread, every object specified in the context is initialized and then passed by reference to the handler each time it is run.

The object returned by the macro is just a closure that sends a single input and a callback through an asyncronous channel to the handler thread. When the handler finishes running a batch it invokes the callback corresponding to each input with the corresponding output, which triggers the closure to wake up and return the output.

Dependencies

~1–1.4MB
~24K SLoC