8.4 kB

Raw Blame History

Joint Inference

Joint Inference

Joint Inference

Motivation

Inference on the edge can get a shorter latency and a higher throughput, and inference on the cloud can get better inference precision.
The collaborative inference technology detects hard samples on the edge and sends them to the cloud for inference.
In this way, simple samples inference on the edge ensures latency and throughput, while hard samples inference on the cloud improves the overall precision.

Goals

Joint inference improves the inference precision without significantly reducing the time and throughput.

Proposal

We propose using Kubernetes Custom Resource Definitions (CRDs) to describe
the joint inference specification/status and a controller to synchronize these updates between edge and cloud.

Use Cases

User can create a joint inference service with providing a training script,
specifying the aggregation algorithm, configuring training hyper parameters,
configuring training datasets.
Users can get the joint inference status, including the counts of inference at the edge/cloud.

Design Details

CRD API Group and Version

The JointInferenceService CRD will be namespace-scoped.
The tables below summarize the group, kind and API version details for the CRD.

JointInferenceService

Field	Description
Group	sedna.io
APIVersion	v1alpha1
Kind	JointInferenceService

Joint inference CRD

see crd source

Joint inference type definition

see go source

Validation

Open API v3 Schema based validation can be used to guard against bad requests.
Invalid values for fields ( example string value for a boolean field etc) can be validated using this.

Here is a list of validations we need to support :

The dataset specified in the crd should exist in k8s.
The model specified in the crd should exist in k8s.
The edgenode name specified in the crd should exist in k8s.

joint inference sample

see sample source

Controller Design

The joint inference controller starts three separate goroutines called upstream, downstream and joint-inferencecontroller. These are not separate controllers as such but named here for clarity.

joint inference: watch the updates of joint-inference-task crds, and create the workers to complete the task.
downstream: synchronize the joint-inference updates from the cloud to the edge node.
upstream: synchronize the joint-inference updates from the edge to the cloud node.

Joint Inference Controller

The joint-inference controller watches for the updates of joint-inference tasks and the corresponding pods against the K8S API server.
Updates are categorized below along with the possible actions:

Update Type	Action
New Joint-inference-service Created	Create the cloud/edge worker
Joint-inference-service Deleted	NA. These workers will be deleted by GM.
The corresponding pod created/running/completed/failed	Update the status of joint-inference task.

Downstream Controller

The downstream controller watches for joint-inference updates against the K8S API server.
Updates are categorized below along with the possible actions that the downstream controller can take:

Update Type	Action
New Joint-inference-service Created	Sends the task information to LCs.
Joint-inference-service Deleted	The controller sends the delete event to LCs.

Upstream Controller

The upstream controller watches for joint-inference-task updates from the edge node and applies these updates against the API server in the cloud.
Updates are categorized below along with the possible actions that the upstream controller can take:

Update Type	Action
Joint-inference-service Reported State Updated	The controller appends the reported status of the Joint-inference-service in the cloud.

Details of api between GM(cloud) and LC(edge)

GM(downstream controller) syncs the task info to LC:

// POST <namespace>/sedna/downstream/jointinferenceservices/<name>/insert
// body same to the task crd of k8s api, omitted here.

LC uploads the task status which reported by the worker to GM(upstream controller):

// POST <namespace>/sedna/upstream/jointinferenceservices/<name>/status

// JoinInferenceServiceStatus defines status that send to GlobalManager
type JoinInferenceServiceStatus struct {
	Phase  string  `json:"phase"`
	Status string  `json:"status"`
	Output *Output `json:"output"`
}

// Output defines task output information
type Output struct {
	Models   []Model   `json:"models"`
	TaskInfo *TaskInfo `json:"taskInfo"`
}

// Model defines the model information
type Model struct {
	Format string `json:"format"`
	URL    string `json:"url"`
}

// TaskInfo defines the task information
type TaskInfo struct {
	InferenceNumber   int     `json:"inferenceNumber"`
	HardExampleNumber int     `json:"hardExampleNumber"`
	UploadCloudRatio  float64 `json:"uploadCloudRatio"`
	StartTime         string  `json:"startTime"`
	CurrentTime       string  `json:"currentTime"`
}

Details of api between Worker(edge) and LC(edge)

Worker sends inference info to LC in same edge node:

// POST /sedna/workers/<worker-name>/info

{
   "name": "worker-name",
   "namespace": "default",
   "ownerName": "jointinferenceservice-name",
   "ownerKind": "jointinferenceservice",
   "kind": "inference",
   "status": "completed/failed/running",
   "taskInfo": {
       "inferenceNumber": 1000,
       "hardExampleNumber": 100,
       "uploadCloudRatio": 0.1,
       "startTime": "2020-11-03T08:39:22.517Z",
       "updateTime": "2020-11-03T08:50:22.517Z"
   }
}

Flow of Joint Inference

The flow of joint inference service creation:

8.4 kB Raw Blame History