The Distributed
package provides functionality for creating and controlling multiple Julia processes remotely, and for performing distributed and parallel computing. It uses network sockets or other supported interfaces to communicate between Julia processes, and relies on Julia's Serialization
stdlib package to transform Julia objects into a format that can be transferred between processes efficiently. It provides a full set of utilities to create and destroy new Julia processes and add them to a "cluster" (a collection of Julia processes connected together), as well as functions to perform Remote Procedure Calls (RPC) between the processes within a cluster. See the API
section for details.
This package ships as part of the Julia stdlib.
To use a newer version of this package, you need to build Julia from scratch. The build process is the same as any other build except that you need to change the commit used in stdlib/Distributed.version
.
It's also possible to load a development version of the package using the trick used in the Section named "Using the development version of Pkg.jl" in the Pkg.jl
repo, but the capabilities are limited as all other packages will depend on the stdlib version of the package and will not work with the modified package.
The public API of Distributed
consists of a variety of functions for various tasks; for creating and destroying processes within a cluster:
addprocs
- create one or more Julia processes and connect them to the clusterrmprocs
- shutdown and remove one or more Julia processes from the cluster
For controlling other processes via RPC:
remotecall
- call a function on another process and return aFuture
referencing the result of that callFuture
- an object that references the result of aremotecall
that hasn't yet completed - usefetch
to return the call's result, orwait
to just wait for the remote call to finishremotecall_fetch
- the same asfetch(remotecall(...))
remotecall_wait
- the same aswait(remotecall(...))
remote_do
- likeremotecall
, but does not provide a way to access the result of the call@spawnat
- likeremotecall
, but in macro form@spawn
- like@spawn
, but the target process is picked automatically@fetch
- macro equivalent offetch(@spawn expr)
@fetchfrom
- macro equivalent offetch(@spawnat p expr)
myid
- returns theInt
identifier of the process calling itnprocs
- returns the number of processes in the clusternworkers
- returns the number of worker processes in the clusterprocs
- returns the set of IDs for processes in the clusterworkers
- returns the set of IDs for worker processes in the clusterinterrupt
- interrupts the specified process
For communicating between processes in the style of a channel or stream:
RemoteChannel
- aChannel
-like object that can beput!
to ortake!
from any process
For controlling multiple processes at once:
WorkerPool
- a collection of processes than can be passed instead a process ID to certain APIsCachingPool
- likeWorkerPool
, but caches functions (including closures which capture large data) on each process@everywhere
- runs a block of code on all (or a subset of all) processes and waits for them all to completepmap
- performs amap
operation where each element may be computed on another process@distributed
- implements afor
-loop where each iteration may be computed on another process
Julia processes connected with Distributed
are all assigned a cluster-unique Int
identifier, starting from 1
. The first Julia process within a cluster is given ID 1
, while other processes added via addprocs
get incrementing IDs (2
, 3
, etc.). Functions and macros which communicate from one process to another usually take one or more identifiers to determine which process they target - for example, remotecall_fetch(myid, 2)
calls myid()
on process 2.
Note: Only process 1 (often called the "head", "primary", or "master") may add or remove processes, and manages the rest of the cluster. Other processes (called "workers" or "worker processes") may still call functions on each other and send and receive data, but addprocs
/rmprocs
on worker processes will fail with an error.