I believe I now see the light vis-?-vis iterators when combined with foreach()
calls in R. I have now been able to reduce computational workload to minutes
instead of hours. I want to verify that the way I am using them is
"safe". By safe I mean does the iterator traverse elements in the same
way as I have below in my toy example to illustrate what I mean.
In the first "traditional" example, I have only one index variable for
the loop and so I know that the same list in r1 and r2 are always being grabbed.
That is, in iteration 1 it is guaranteed to use r1[[1]] + r2[[1]].
In the example that uses the iterators, is this also guaranteed even though I
now have two iterator objects? That is, will the index for element i always be
the same as the index for element j when using this across many different cores?
It seems to be true and in all my test cases so far I am seeing it to be true.
But, that could be just luck, so I wonder if there is a condition under which
that would NOT be true.
Thank you
Harold
library(foreach)
library(doParallel)
cl <- makeCluster(2)
registerDoParallel(cl)
### Create random data
r1 <- vector("list", 20)
for(i in 1:20){
r1[[i]] <- rnorm(10)
}
### Create random data
r2 <- vector("list", 20)
for(i in 1:20){
r2[[i]] <- rnorm(10)
}
### Use a for loop traditionally
result1 <- vector("list", 20)
for(i in 1:20){
result1[[i]] <- r1[[i]] + r2[[i]]
}
### Use iterators
itx1 <- iter(r1)
itx2 <- iter(r2)
result2 <- foreach(i = itx1, j = itx2) %dopar% {
i + j
}
all.equal(result1, result2)
> On Dec 9, 2016, at 9:15 AM, Doran, Harold <HDoran at air.org> wrote: > > I believe I now see the light vis-?-vis iterators when combined with foreach() calls in R. I have now been able to reduce computational workload to minutes instead of hours. I want to verify that the way I am using them is "safe". By safe I mean does the iterator traverse elements in the same way as I have below in my toy example to illustrate what I mean. > > In the first "traditional" example, I have only one index variable for the loop and so I know that the same list in r1 and r2 are always being grabbed. That is, in iteration 1 it is guaranteed to use r1[[1]] + r2[[1]]. > > In the example that uses the iterators, is this also guaranteed even though I now have two iterator objects? That is, will the index for element i always be the same as the index for element j when using this across many different cores? > > It seems to be true and in all my test cases so far I am seeing it to be true. But, that could be just luck, so I wonder if there is a condition under which that would NOT be true. > > Thank you > Harold > > > library(foreach) > library(doParallel) > cl <- makeCluster(2) > registerDoParallel(cl) > > ### Create random data > r1 <- vector("list", 20) > for(i in 1:20){ > r1[[i]] <- rnorm(10) > } > > ### Create random data > r2 <- vector("list", 20) > for(i in 1:20){ > r2[[i]] <- rnorm(10) > } > > ### Use a for loop traditionally > result1 <- vector("list", 20) > for(i in 1:20){ > result1[[i]] <- r1[[i]] + r2[[i]] > } > > ### Use iterators > itx1 <- iter(r1) > itx2 <- iter(r2) > > result2 <- foreach(i = itx1, j = itx2) %dopar% { > i + j > } > > all.equal(result1, result2)I wasn't sure how this would or should behave. I'm not an experienced user, merely a reader of help pages. Neither the help page, not the vignette references on the help page answered my questions in this case. I expected that call would behave analogously to the behavior of mapply when given iterators of unequal length. (The shorter of the objects is recycled to reach the length of the longer object.) That expectation was not realized. It appears that the length of first object of the objects determines computation length, but that missing values will not be recycled for the shorter iterator. An error is not reported, but rather numeric(0) is returned. So in one sense the %dopar% version is "safer" at least to the extent of not failing with an error that would have occurred when using a for-loop. This was my test case: r1 <- vector("list", 10) for(i in 1:20){ r1[[i]] <-20:29+i*10 # random numbers are not good for determining sequences of operations } r2 <- vector("list", 20) for(i in 1:10){ r2[[i]] <- 1:10 +i } itx1 <- iter(r1) itx2 <- iter(r2) result2 <- foreach(i = itx1, j = itx2) %dopar% { i + j } result2 -- David Winsemius Alameda, CA, USA
That is a helpful, and important, caveat. So, perhaps I should amend my original question to ask something like is it safe *when* length(r1) == length(r2) -----Original Message----- From: David Winsemius [mailto:dwinsemius at comcast.net] Sent: Friday, December 09, 2016 1:27 PM To: Doran, Harold <HDoran at air.org> Cc: r-help at r-project.org Subject: Re: [R] "Safe" use of iterator (package iterators)> On Dec 9, 2016, at 9:15 AM, Doran, Harold <HDoran at air.org> wrote: > > I believe I now see the light vis-?-vis iterators when combined with foreach() calls in R. I have now been able to reduce computational workload to minutes instead of hours. I want to verify that the way I am using them is "safe". By safe I mean does the iterator traverse elements in the same way as I have below in my toy example to illustrate what I mean. > > In the first "traditional" example, I have only one index variable for the loop and so I know that the same list in r1 and r2 are always being grabbed. That is, in iteration 1 it is guaranteed to use r1[[1]] + r2[[1]]. > > In the example that uses the iterators, is this also guaranteed even though I now have two iterator objects? That is, will the index for element i always be the same as the index for element j when using this across many different cores? > > It seems to be true and in all my test cases so far I am seeing it to be true. But, that could be just luck, so I wonder if there is a condition under which that would NOT be true. > > Thank you > Harold > > > library(foreach) > library(doParallel) > cl <- makeCluster(2) > registerDoParallel(cl) > > ### Create random data > r1 <- vector("list", 20) > for(i in 1:20){ > r1[[i]] <- rnorm(10) > } > > ### Create random data > r2 <- vector("list", 20) > for(i in 1:20){ > r2[[i]] <- rnorm(10) > } > > ### Use a for loop traditionally > result1 <- vector("list", 20) > for(i in 1:20){ > result1[[i]] <- r1[[i]] + r2[[i]] > } > > ### Use iterators > itx1 <- iter(r1) > itx2 <- iter(r2) > > result2 <- foreach(i = itx1, j = itx2) %dopar% { > i + j > } > > all.equal(result1, result2)I wasn't sure how this would or should behave. I'm not an experienced user, merely a reader of help pages. Neither the help page, not the vignette references on the help page answered my questions in this case. I expected that call would behave analogously to the behavior of mapply when given iterators of unequal length. (The shorter of the objects is recycled to reach the length of the longer object.) That expectation was not realized. It appears that the length of first object of the objects determines computation length, but that missing values will not be recycled for the shorter iterator. An error is not reported, but rather numeric(0) is returned. So in one sense the %dopar% version is "safer" at least to the extent of not failing with an error that would have occurred when using a for-loop. This was my test case: r1 <- vector("list", 10) for(i in 1:20){ r1[[i]] <-20:29+i*10 # random numbers are not good for determining sequences of operations } r2 <- vector("list", 20) for(i in 1:10){ r2[[i]] <- 1:10 +i } itx1 <- iter(r1) itx2 <- iter(r2) result2 <- foreach(i = itx1, j = itx2) %dopar% { i + j } result2 -- David Winsemius Alameda, CA, USA