BlockHaloArrays

The BlockHaloArray type is an array-like type (does not extend AbstractArray) that is designed for shared-memory multi-threaded workloads. Typical shared-memory parallelization is done via multi-threaded loops (with @threads, @tturbo, etc.), i.e.:

Threads.@threads for j in jlo:jhi
    for i in ilo:ihi
        A[i,j] = ...
    end
end

However, this type of parallelization tends not to scale well, especially for memory-bandwidth limited workloads. This is especially true for stencil operations, where functions access memory at addresses such as A[i,j], A[i+1,j], A[i-1,j+1]. The aim of the BlockHaloArray type do mini-domain decomposition into separate thread-specific chunks, or blocks, of memory for each thread to operate on without data race conditions or memory bandwidth contention. Because the context is shared memory however, certain operations are more efficient compared to the typical MPI domain decomposition method.Each thread will be able to operate on it's own block of memory and maintain NUMA-aware locality and communication is done via copies, rather than MPI library calls. Future releases will include an MPI-aware BlockHaloArray that will facilitate the MPI+Threads hybrid parallelization.

An example of a blocked-style stencil operation looks like the following. Note that it can still take advantage of the single-threaded @turbo macro to vectorize the nested-loop.

function stencil!(A::BlockHaloArray, blockid)
    ilo, ihi, jlo, jhi = A.loop_limits[blockid]
    data = @views A.blocks[blockid]
    @turbo for j in jlo:jhi
        for i in ilo:ihi
            data[i, j] = i * j + 0.25(data[i-2, j] + data[i-1, j] + data[i+2, j] + data[i+1, j] +
                                      data[i, j] +
                                      data[i, j-2] + data[i, j-1] + data[i, j+2] + data[i, j+1])
        end
    end
end

Then the stencil!() function is split among threads via:

@sync for block_id in 1:nthreads()
    @tspawnat block_id stencil!(A, block_id)
end

BlockHaloArray types should be used in a task-based parallel workload, which has better scaling than multi-threaded loops. The only synchronization required is during the exchange of halo data.

Benchmark results are coming soon...

Exports

Types

BlockHaloArray: A blocked array-like type (does not extent AbstractArray) to be used in a shared-memory type system. This partitions an Array into N blocks that are to be operated on by threads.

Constructors

BlockHaloArray(dims::NTuple{N,Int}, halodims::NTuple{N2,Int}, nhalo::Integer, nblocks=nthreads(); T=Float64, use_numa=true)
BlockHaloArray(dims::NTuple{N,Int}, nhalo::Integer, nblocks=nthreads(); T=Float64, use_numa=true)
BlockHaloArray(A::AbstractArray{T,N}, nhalo::Integer, nblocks=nthreads())
BlockHaloArray(A::AbstractArray{T,N}, halodims::NTuple{N2,Integer}, nhalo::Integer, nblocks=nthreads())

Example

using .Threads
using ThreadPools
using BlockHaloArrays

dims = (30, 20) # a 30 x 20 matrix
nhalo = 2 # number of halo entries in each dimension
nblocks = nthreads() # nblocks must be ≤ nthreads() or a warning will be issued

A = BlockHaloArray(dims, nhalo, nblocks; T=Float64) # create with empty data
B = BlockHaloArray(rand(dims...), nhalo, nblocks) # create from an existing array

# Create a 3D array, but only do halo exchange on the 2nd and 3rd dimensions
newdims = (4, 100, 200)
haloax = (2, 3) # these must be monotonically increasing and the outer-most dims, i.e. can't be (1, 2), or (1, 3)
C = BlockHaloArray(newdims, haloax, nhalo, nblocks; T=Float64)

# Fill each block with it's corresponding threadid value
function init(A)
    dom = domainview(A, threadid())
    fill!(dom, threadid())
end

# Let each thread initialize it's own block
@sync for tid in 1:nblocks(A)
    @tspawnat tid init(A1)
end

# Let each block sync its halo region with its neighbors
updatehalo!(A)

Anew = flatten(A) # Anew is a 2D Array -- no longer a BlockHaloArray

MPIBlockHaloArray: (to be implemented) An MPI-aware version of a BlockHaloArray. This facilitates halo communication between nodes via MPI.

Functions

flatten(A): Return a flattened Array of A. This is a copy.
repartition(A, nblocks): Repartition the BlockHaloArray into a different block layout
updatehalo!(A, include_periodic_bc=false): Sync all block halo regions. This uses a @sync loop with @tspawnat to sync each block.
updateblockhalo!(A, block_id, include_periodic_bc=false): Sync the halo regions of a single block. No @sync or @spawn calls are used.
domainview(A, blockid): Return a view of a the domain (no halo regions) of block blockid
onboundary(A, blockid): Is the current block blockid on a boundary? This is used help apply boundary conditions in a user code.

Additional overloaded methods include

eltype(A::AbstractBlockHaloArray) = eltype(first(A.blocks))

size(A::AbstractBlockHaloArray) = A.globaldims
size(A::AbstractBlockHaloArray, dim) = size.(A.blocks, dim)

axes(A::AbstractBlockHaloArray) = axes.(A.blocks)
axes(A::AbstractBlockHaloArray, dim) = axes.(A.blocks, dim)

first(A::AbstractBlockHaloArray) = first(A.blocks)
firstindex(A::AbstractBlockHaloArray) = firstindex(A.blocks)

last(A::AbstractBlockHaloArray) = last(A.blocks)
lastindex(A::AbstractBlockHaloArray) = lastindex(A.blocks)

eachindex(A::AbstractBlockHaloArray) = eachindex(A.blocks)

nblocks(A::AbstractBlockHaloArray) = length(A.blocks)

smillerc / blockhaloarrays.jl Goto Github PK