Inference Batcher¶

This docs explains on how batch prediction for any ML frameworks (TensorFlow, PyTorch, ...) without decreasing the performance.

This batcher is implemented in the KServe model agent sidecar, so the requests first hit the agent sidecar, when a batch prediction is triggered the request is then sent to the model server container for inference.

• We use webhook to inject the model agent container in the InferenceService pod to do the batching when batcher is enabled.

• We use go channels to transfer data between http request handler and batcher go routines.

• Currently we only implemented batching with KServe v1 HTTP protocol, gRPC is not supported yet.

• When the number of instances (For example, the number of pictures) reaches the maxBatchSize or the latency meets the maxLatency, a batch prediction will be triggered.

Example¶

We first create a pytorch predictor with a batcher. The maxLatency is set to a big value (500 milliseconds) to make us be able to observe the batching process.

apiVersion: serving.kserve.io/v1beta1
kind: InferenceService
metadata:
  name: "torchserve"
spec:
  predictor:
    minReplicas: 1
    timeout: 60
    batcher:
      maxBatchSize: 32
      maxLatency: 500
    model:
      modelFormat:
        name: pytorch
      storageUri: gs://kfserving-examples/models/torchserve/image_classifier/v1

apiVersion: serving.kserve.io/v1beta1
kind: InferenceService
metadata:
  name: "torchserve"
spec:
  predictor:
    minReplicas: 1
    timeout: 60
    batcher:
      maxBatchSize: 32
      maxLatency: 500
    pytorch:
      storageUri: gs://kfserving-examples/models/torchserve/image_classifier/v1

• maxBatchSize: the max batch size for triggering a prediction.

• maxLatency: the max latency for triggering a prediction (In milliseconds).

• timeout: timeout of calling predictor service (In seconds).

All of the bellowing fields have default values in the code. You can config them or not as you wish.

• maxBatchSize: 32.

• maxLatency: 500.

• timeout: 60.

kubectl create -f pytorch-batcher.yaml

We can now send requests to the pytorch model using hey. The first step is to determine the ingress IP and ports and set INGRESS_HOST and INGRESS_PORT

MODEL_NAME=mnist
INPUT_PATH=@./input.json
SERVICE_HOSTNAME=$(kubectl get inferenceservice torchserve -o jsonpath='{.status.url}' | cut -d "/" -f 3)

hey -z 10s -c 5 -m POST -host "${SERVICE_HOSTNAME}" -H "Content-Type: application/json" -D ./input.json "http://${INGRESS_HOST}:${INGRESS_PORT}/v1/models/$MODEL_NAME:predict"

The request will go to the model agent container first, the batcher in sidecar container batches the requests and send the inference request to the predictor container.

Summary:
  Total:    10.5361 secs
  Slowest:  0.5759 secs
  Fastest:  0.4983 secs
  Average:  0.5265 secs
  Requests/sec: 9.4912

  Total data:   24100 bytes
  Size/request: 241 bytes

Response time histogram:
  0.498 [1] |■
  0.506 [0] |
  0.514 [44]    |■■■■■■■■■■■■■■■■■■■■■■■■■■■■■■■■■■■■■■■■
  0.522 [21]    |■■■■■■■■■■■■■■■■■■■
  0.529 [4] |■■■■
  0.537 [5] |■■■■■
  0.545 [4] |■■■■
  0.553 [0] |
  0.560 [7] |■■■■■■
  0.568 [4] |■■■■
  0.576 [10]    |■■■■■■■■■


Latency distribution:
  10% in 0.5100 secs
  25% in 0.5118 secs
  50% in 0.5149 secs
  75% in 0.5406 secs
  90% in 0.5706 secs
  95% in 0.5733 secs
  99% in 0.5759 secs

Details (average, fastest, slowest):
  DNS+dialup:   0.0004 secs, 0.4983 secs, 0.5759 secs
  DNS-lookup:   0.0001 secs, 0.0000 secs, 0.0015 secs
  req write:    0.0002 secs, 0.0000 secs, 0.0076 secs
  resp wait:    0.5257 secs, 0.4981 secs, 0.5749 secs
  resp read:    0.0001 secs, 0.0000 secs, 0.0009 secs

Status code distribution:
  [200] 100 responses