A different approach uses doRedis https://CRAN.R-project.org/package=doRedis (currently archived, but actively developed) for use with the foreach package, or RedisParam https://github.com/mtmorgan/RedisParam (not released) for use with Bioconductor's BiocParallel package.
These use a redis server https://redis.io/ to communicate -- the manager submits jobs / obtains results from the redis server, the workers retrieve jobs / submit results to the redis server. Manager and worker need to know the (http) address of the server, etc, but there are no other ports involved. Redis servers are easy to establish in a cloud environment, using e.g., existing AWS or docker images. The README for doRedis https://github.com/bwlewis/doRedis probably provides the easiest introduction. The (not mature) k8sredis Kubernetes / helm chart https://github.com/Bioconductor/k8sredis illustrates a complete system using RedisParam, deploying manager and workers locally or in the google cloud; the app could be modified to only start the workers in the cloud, exposing the redis server for access by a local 'manager'; this would be cool. Martin On 1/19/21, 1:50 AM, "R-help on behalf of Henrik Bengtsson" <r-help-boun...@r-project.org on behalf of henrik.bengts...@gmail.com> wrote: On Mon, Jan 18, 2021 at 9:42 PM Jiefei Wang <szwj...@gmail.com> wrote: > > Thanks for introducing this interesting package to me! it is great to know a new powerful tool, but it seems like this method does not work in my environment. ` parallelly::makeClusterPSOCK` will hang until timeout. > > I checked the verbose output and it looks like the parallelly package also depends on `parallel:::.slaveRSOCK` on the remote instance to build the connection. This explains why it failed for the local machine does not have a public IP and the remote does not know how to build the connection. It's correct that the worker does attempt to connect back to the parent R process that runs on your local machine. However, it does *not* do so by your local machines public IP address but it does it by connecting to a port on its own machine - a port that was set up by SSH. More specifically, when parallelly::makeClusterPSOCK() connects to the remote machine over SSH it also sets up a so-called reverse SSH tunnel with a certain port on your local machine and certain port of your remote machine. This is what happens: > cl <- parallelly::makeClusterPSOCK("machine1.example.org", verbose=TRUE) [local output] Workers: [n = 1] 'machine1.example.org' [local output] Base port: 11019 ... [local output] Starting worker #1 on 'machine1.example.org': '/usr/bin/ssh' -R 11068:localhost:11068 machine1.example.org "'Rscript' --default-packages=datasets,utils,grDevices,graphics,stats,methods -e 'workRSOCK <- tryCatch(parallel:::.slaveRSOCK, error=function(e) parallel:::.workRSOCK); workRSOCK()' MASTER=localhost PORT=11068 OUT=/dev/null TIMEOUT=2592000 XDR=FALSE" [local output] - Exit code of system() call: 0 [local output] Waiting for worker #1 on 'machine1.example.org' to connect back '/usr/bin/ssh' -R 11019:localhost:11019 machine1.example.org "'Rscript' --default-packages=datasets,utils,grDevices,graphics,stats,methods -e 'workRSOCK <- tryCatch(parallel:::.slaveRSOCK, error=function(e) parallel:::.workRSOCK); workRSOCK()' MASTER=localhost PORT=11019 OUT=/dev/null TIMEOUT=2592000 XDR=FALSE" All the magic is in that SSH option '-R 11068:localhost:11068' SSH options, which allow the parent R process on your local machine to communicate with the remote worker R process on its own port 11068, and vice versa, the worker R process will communicate with the parent R process as if it was running on MASTER=localhost PORT=11068. Basically, for all that the worker R process' knows, the parent R process runs on the same machine as itself. You haven't said what operating system you're running on your local machine, but if it's MS Windows, know that the 'ssh' client that comes with Windows 10 has some bugs in its reverse tunneling. See ?parallelly::makeClusterPSOCK for lots of details. You also haven't said what OS the cloud workers run, but I assume it's Linux. So, my guesses on your setup is, the above "should work" for you. For your troubleshooting, you can also set argument outfile=NULL. Then you'll also see output from the worker R process. There are additional troubleshooting suggestions in Section 'Failing to set up remote workers' of ?parallelly::makeClusterPSOCK that will help you figure out what the problem is. > > I see in README the package states it works with "remote clusters without knowing public IP". I think this might be where the confusion is, it may mean the remote machine does not have a public IP, but the server machine does. I'm in the opposite situation, the server does not have a public IP, but the remote does. I'm not sure if this package can handle my case, but it looks very powerful and I appreciate your help! Thanks. I've updated the text to "remote clusters without knowing [local] public IP". /Henrik > > Best, > Jiefei > > > > > > On Tue, Jan 19, 2021 at 1:22 AM Henrik Bengtsson <henrik.bengts...@gmail.com> wrote: >> >> If you have SSH access to the workers, then >> >> workers <- c("machine1.example.org", "machine2.example.org") >> cl <- parallelly::makeClusterPSOCK(workers) >> >> should do it. It does this without admin rights and port forwarding. >> See also the README in https://cran.r-project.org/package=parallelly. >> >> /Henrik >> >> On Mon, Jan 18, 2021 at 6:45 AM Jiefei Wang <szwj...@gmail.com> wrote: >> > >> > Hi all, >> > >> > I have a few cloud instances and I want to use them to do parallel >> > computing. I would like to create a socket cluster on my local machine to >> > control the remote instances. Here is my network setup: >> > >> > local machine -- NAT -- Internet -- cloud instances >> > >> > In the parallel package, the server needs to call `makeCluster()` and >> > listens to the connection from the workers. In my case, the server is the >> > local machine and the workers are the cloud instances. However, since the >> > local machine is hidden behind the NAT, it does not have a public address >> > and the worker cannot connect to it. Therefore, `makeCluster()` will never >> > be able to see the connection from the workers and hang forever. >> > >> > One solution for letting the external machine to access the device inside >> > the NAT is to use port forwarding. However, this would not work for my case >> > as the NAT is set by the network provider(not my home router) so I do not >> > have access to the router. As the cloud instances have public addresses, >> > I'll wonder if there is any way to build the cluster by letting the server >> > connect to the cloud? I have checked `?parallel::makeCluster` and >> > `?snow::makeSOCKcluster` but I found no result. The only promising solution >> > I can see now is to use TCP hole punching, but it is quite complicated and >> > may not work for every case. Since building a connection from local to the >> > remote is super easy, I would like to know if there exists any simple >> > solution. I have searched it on Google for a week but find no answer. I'll >> > appreciate it if you can provide me any suggestions! >> > >> > Best, >> > Jiefei >> > >> > [[alternative HTML version deleted]] >> > >> > ______________________________________________ >> > R-help@r-project.org mailing list -- To UNSUBSCRIBE and more, see >> > https://stat.ethz.ch/mailman/listinfo/r-help >> > PLEASE do read the posting guide http://www.R-project.org/posting-guide.html >> > and provide commented, minimal, self-contained, reproducible code. ______________________________________________ R-help@r-project.org mailing list -- To UNSUBSCRIBE and more, see https://stat.ethz.ch/mailman/listinfo/r-help PLEASE do read the posting guide http://www.R-project.org/posting-guide.html and provide commented, minimal, self-contained, reproducible code. ______________________________________________ R-help@r-project.org mailing list -- To UNSUBSCRIBE and more, see https://stat.ethz.ch/mailman/listinfo/r-help PLEASE do read the posting guide http://www.R-project.org/posting-guide.html and provide commented, minimal, self-contained, reproducible code.
