我想实现操作M %*% t(M)
的并行化版本。我有一台拥有100个内核和共享内存/硬盘的机器,我希望尽可能高效地使用它。
因为我的矩阵很大(20000x15000),我想避免拆分矩阵并直接将块发送给worker,因为这是串行发生的。
相反,通过使用bigmemory
包,我只能将矩阵描述符和行索引发送给我的工作者。这是一个很小的数据,所以每个工作人员应该被简要介绍一下#34;太快了。然后,每个工作者将并行地附加在共享存储器中的矩阵。输出矩阵也是如此,每个工作者并行写入(当然每个都写入不同的行)
此外,下一个工作人员启动所需的持续时间越长,输入矩阵越大。我不明白这一点。如下所示,我并行执行的函数mult
得到一个整数向量为2(第一行/最后一行)和两个由bigmemory::describe
创建的对象;这些对象不依赖于输入矩阵的大小。为什么工人需要更长时间才能开始使用更大的矩阵?
htop
: matOuter
定义如下。执行:
mkmatrix <- function(n)
matrix(sample(x = 10, size = n^2, replace = TRUE), nrow = n, ncol = n)
clu <- parallel::makeCluster(100)
n <- 10000
# n <- 1000 # the workers start faster one after another with a smaller matrix
M <- mkmatrix(n)
p <- matOuter(M, parallel = clu)
stopCluster(clu)
功能定义:
matOuter <- function(M, parallel = 1){
if(is.numeric(parallel)){
# Make a new cluster
if(parallel == 1) return( M %*% t(M))
parallel = parallel::makeCluster(parallel)
on.exit(parallel::stopCluster(parallel))
} # else use the provided cluster.
# Get row index ranges for each worker to tackle
x <- ceiling(seq(1, nrow(M)+1, length.out = min(nrow(M), length(parallel)) + 1))
idxFrom <- x[-length(x)]
idxTo <- x[-1] - 1
idx <- cbind(idxFrom, idxTo)
bM <- bigmemory::as.big.matrix(M, type = "double") # bigalgebra::%*% needs type double (see bigalgebra:::check_matrix)
bMOut <- bigmemory::big.matrix(nrow(M), nrow(M), dimnames = rownames(M))
mult <- function(row, descIn, descOut){
library(bigalgebra)
A <- bigmemory::attach.big.matrix(descIn)
O <- bigmemory::attach.big.matrix(descOut)
O[row[1]:row[2], ] <- t(bigmemory::as.matrix(A %*% t(A[row[1]:row[2],,drop = FALSE])))
return(NULL)
}
dM <- bigmemory::describe(bM)
dO <- bigmemory::describe(bMOut)
# Serial version for debugging.
#apply(idx, 1, mult, descIn = dM, descOut = dO)
parallel::parApply(parallel, idx, 1, descIn = dM, descOut = dO, FUN = mult)
return(bigmemory::as.matrix(bMOut))
}
R version 3.4.3 (2017-11-30)
Platform: x86_64-pc-linux-gnu (64-bit)
Running under: CentOS Linux 7 (Core)
Matrix products: default
BLAS: /opt/Bio/R/3.4.3/lib64/R/lib/libRblas.so
LAPACK: /opt/Bio/R/3.4.3/lib64/R/lib/libRlapack.so
locale:
[1] LC_CTYPE=en_US.UTF-8 LC_NUMERIC=C
[3] LC_TIME=en_US.UTF-8 LC_COLLATE=en_US.UTF-8
[5] LC_MONETARY=en_US.UTF-8 LC_MESSAGES=en_US.UTF-8
[7] LC_PAPER=en_US.UTF-8 LC_NAME=C
[9] LC_ADDRESS=C LC_TELEPHONE=C
[11] LC_MEASUREMENT=en_US.UTF-8 LC_IDENTIFICATION=C
attached base packages:
[1] stats graphics grDevices utils datasets methods base
other attached packages:
[1] prAtm_0.0
loaded via a namespace (and not attached):
[1] Rcpp_0.12.17 roxygen2_6.0.1 digest_0.6.15
[4] withr_2.1.1 commonmark_1.4 R6_2.2.2
[7] magrittr_1.5 bigmemory.sri_0.1.3 rlang_0.1.6
[10] stringi_1.1.6 testthat_2.0.0 xml2_1.2.0
[13] bigmemory_4.5.33 devtools_1.13.4 tools_3.4.3
[16] stringr_1.2.0 parallel_3.4.3 yaml_2.1.16
[19] compiler_3.4.3 memoise_1.1.0