ps

Documentation

Index

• Constants
• func Start(logger *zap.Logger, port int, schedulerUrl string, standaloneJobs bool)
• type ParameterServer
  • func (ps *ParameterServer) GetHandler() http.Handler
  • func (ps *ParameterServer) Serve(port int)
• type TaskType

Constants

const (
	KubeMlNamespace = "kubeml"
	KubeMlContainer = "diegostock12/kubeml"
)

Functions

func Start

func Start(logger *zap.Logger, port int, schedulerUrl string, standaloneJobs bool)

Start Starts a New parameter server which will execute the tasks 1) start the new functions 2) receive the notifications from the PS API about functions that have finished processing which will trigger the execution retrieval of gradients and the update of the model 3) Start the API to get the requests from the functions

Types

type ParameterServer

type ParameterServer struct {
	// contains filtered or unexported fields
}

ParameterServer is the main parameter server that spins out the worker jobs that serve training jobs. The parameter server saves the index to communicate with each of the jobs. The jobs send requests to the Scheduler when they finish an epoch, and the scheduler responds to the main PS api, which then makes the response reach the correct train job.

func (*ParameterServer) GetHandler

func (ps *ParameterServer) GetHandler() http.Handler

GetHandler Returns the handler for calls from the functions

func (*ParameterServer) Serve

func (ps *ParameterServer) Serve(port int)

Start the API at the given port All of the parameter server threads share the same API, and they communicate through channels

type TaskType

type TaskType string

const (
	TrainTask     TaskType = "train"
	InferenceTask TaskType = "infer"
)