scheduleTaskList: Minimize Expression Start Time

In clarkfitzg/codedoctor: Transform Serial R Code into Parallel R Code

Description

Implementation of "list scheduling". This is a greedy algorithm that assigns each expression to the earliest possible processor.

Usage

scheduleTaskList(graph, platform = Platform(), data = list(),
  orderFun = orderBottomLevel, timeDefault = 1e-05,
  sizeDefault = as.numeric(utils::object.size(1L)), overhead = 8e-06,
  bandwidth = 1.5e+09)

Arguments

graph	TaskGraph as returned from inferGraph
orderFun	function that takes in a graph and returns a permutation of 1:length(graph@code) that respects the topological ordering of the graph.
timeDefault	numeric time in seconds to execute a single expression. Expression times default to this value, with a warning, if we can't find time from graph.
sizeDefault	numeric default size of objects to transfer in bytes
overhead	numeric seconds to send any object
bandwidth	numeric speed that the network can transfer an object between processors in bytes per second. We don't take network contention into account. This will have to be extended to account for multiple machines.
nWorkers	integer maximum number of processors

Details

This function is experimental and unstable. If you're trying to actually speed up your code through parallelism then consider using the default method in schedule for data parallelism. This function rewrites code to use task parallelism. Task parallelism means two or more processors run different R expressions simultaneously.

Value

schedule object of class TaskSchedule

References

Algorithm 10 in Task Scheduling for Parallel Systems, Sinnen (2007)

Examples

code <- parse(text = "a <- 100
     b <- 200
     c <- a + b")

g <- inferGraph(code)
s <- scheduleTaskList(g)
plot(s)

clarkfitzg/codedoctor documentation built on Nov. 18, 2020, 4:34 p.m.