Julia Parallel Computing Notes (4)

Seven, shared array

The penultimate section of this chapter introduces one of Julia's features: SharedArray.

Julia's shared array is a shared multi-dimensional array provided by the SharedArrays package. It can be cross-process. At this time, the SharedArrays package must be explicitly loaded to each process that needs it, or it can be cross-coroutine. At this time, just load the SharedArrays package on the main process. Can @everywherebe easily added to the package SharedArrays all processes:

julia> @everywhere using SharedArrays

But it seems that there is no way to @spawnatload it to the specified process with such commands, so be honest and practical @everywhere.

Looking back @distributed, it can also do coroutine-level parallelism on the main process, so we can use shared arrays to implement it at the coroutine level @distributed. Make a demonstration:

julia> using SharedArrays

julia> a = SharedArray{Float64}(1,10)
1×10 SharedArray{Float64,2}:
 0.0  0.0  0.0  0.0  0.0  0.0  0.0  0.0  0.0  0.0

julia> @distributed for i = 1:10
                       a[i]=i
                    end
Task (done) @0x0000000006671710

julia> a
1×10 SharedArray{Float64,2}:
 1.0  2.0  3.0  4.0  5.0  6.0  7.0  8.0  9.0  10.0

It can be seen that multiple coroutines operate the same array in parallel. Let's look at an example of cross-process operation:

julia> @everywhere using SharedArrays

julia> a = SharedArray{Float64}(1,10)
1×10 SharedArray{Float64,2}:
 0.0  0.0  0.0  0.0  0.0  0.0  0.0  0.0  0.0  0.0

julia> @fetchfrom 3 (a[6]=2;a[1]=a[6]+1)
3.0

It proves that the shared array can be operated as a normal array in each process, there is no difference.

When declaring a shared array, you can specify two key-value parameters initand pidsinitialize them and map them to the specified process.

Tips: Key-value parameters are also called keyword parameters. Different from general parameters, keywords and values ​​must be explicitly stated. For example: Matlab drawing parameters' 'FontSize','值'cannot be abbreviated as '值'. The key-value parameters of the Julia shared array initmust be written init=函数in the pidssame way. The advantage of key-value parameters is that they can be arranged out of order, but they are generally required to be arranged behind ordinary parameters.

for example:

julia> S = SharedArray{Int,2}((3,4),init = A->(for i=1:3;A[i,1]=100;end),pids=[2,4])
3×4 SharedArray{Int64,2}:
 100  0  0  0
 100  0  0  0
 100  0  0  0

Pay attention to the wording of vectors when passing parameters. For example, pids=[2,4]if you write pids=[2 4]achievements, you will get an error because the keyword does not accept multidimensional arrays. In Julia, `[2 4] will be considered as multi-line writing. For details, please refer to the chapter on multi-dimensional arrays.

The efficiency of shared arrays is the same in any process, so parallel operation of large-scale data has extremely high efficiency.

Finally, I specifically point out two important contents:

• The auxiliary initialization function demonstrated in the original work SharedArrays.localindices()reported an error in Julia 1.1.0 for unknown reasons.
• All the above methods are tested on a single host, that is, the so-called single-machine distributed platform or virtual cluster, which has not been tested on a distributed platform composed of multiple hosts (here, "cluster" refers to a distributed multi-host) test. Julia provides --machinefilecommands for clusters and a unified style manager for virtual clusters and clusters ClusterManager, including Distributed.LocalManagerand Distributed.SSHManager. However, it is a pity that it was mentioned in the original work. The next section will sort out the relevant content in the official documents and may supplement your own test results.