parallel:::newPSOCKnode(): background worker fails immediately if socket on master is not set up in time (BUG?)

I'm happy to look at a patch that does this.  I'd start with a small
interval and increase it by 50%, say, on each try wit a max retry time
limit. This isn't eliminating the problem,only reducing the
probability, but still worth it. I had considered doing something like
this but it didn't seem necessary at the time. You don't want to retry
indefinitely since the connection could be failing because the master
died, and then you want the workers to die as well.

Best,

luke


On Fri, 9 Mar 2018, Henrik Bengtsson wrote:

A solution is to have parallel:::.slaveRSOCK() attempt to connect
multiple times before failing, e.g.

   makeSOCKmaster <- function(master, port, timeout, useXDR, maxTries
= 10L, interval = 1.0) {
       port <- as.integer(port)
       for (i in seq_len(maxTries)) {
         con <- tryCatch({
           socketConnection(master, port = port, blocking = TRUE,
                                   open = "a+b", timeout = timeout)
         }, error = identity)
         if (inherits(con, "connection")) break
         Sys.sleep(interval)
       }
       if (inherits(con, "error")) stop(con)
       structure(list(con = con), class = if (useXDR)
           "SOCKnode"
       else "SOCK0node")
   }

One could set 'maxTries' and 'interval' via commandArgs() like what is
done for the other arguments.

I'm happy to submit an SVN patch if R core thinks this is an
acceptable solution.

/Henrik

On Thu, Mar 8, 2018 at 7:36 PM, Henrik Bengtsson
<henrik.bengtsson at gmail.com> wrote:

I just noticed that parallel:::.slaveRSOCK() passes 'timeout' to
socketConnection() as a character, i.e. there's a missing timeout <-
as.integer(timeout), cf. port <- as.integer(port) and useXDR <-
as.logical(value):

parallel:::.slaveRSOCK

function ()
{
    makeSOCKmaster <- function(master, port, timeout, useXDR) {
        port <- as.integer(port)
        con <- socketConnection(master, port = port, blocking = TRUE,
            open = "a+b", timeout = timeout)
        structure(list(con = con), class = if (useXDR)
            "SOCKnode"
        else "SOCK0node")
    }
    master <- "localhost"
    port <- NA_integer_
    outfile <- Sys.getenv("R_SNOW_OUTFILE")
    methods <- TRUE
    useXDR <- TRUE
    for (a in commandArgs(TRUE)) {
        pos <- regexpr("=", a)
        name <- substr(a, 1L, pos - 1L)
        value <- substr(a, pos + 1L, nchar(a))
        switch(name, MASTER = {
            master <- value
        }, PORT = {
            port <- value
        }, OUT = {
            outfile <- value
        }, TIMEOUT = {
            timeout <- value
        }, XDR = {
            useXDR <- as.logical(value)
        })
    }
    if (is.na(port))
        stop("PORT must be specified")
    sinkWorkerOutput(outfile)
    msg <- sprintf("starting worker pid=%d on %s at %s\n", Sys.getpid(),
        paste(master, port, sep = ":"), format(Sys.time(), "%H:%M:%OS3"))
    cat(msg)
    slaveLoop(makeSOCKmaster(master, port, timeout, useXDR))
}
<bytecode: 0x4bd4b58>
<environment: namespace:parallel>

Yet, fix that does *not* seem to change anything.

/Henrik

On Thu, Mar 8, 2018 at 7:05 PM, Henrik Bengtsson
<henrik.bengtsson at gmail.com> wrote:

BACKGROUND:

While troubleshooting random, occasionally occurring, errors from
parallel::makePSOCKcluster("localhost", port = 11000);

Error in socketConnection("localhost", port = port, server = TRUE,
blocking = TRUE, :
    cannot open the connection

I had another look at parallel:::newPSOCKnode(), which is used
internally to set up each background worker.  It is designed to, first
launch the background worker as:

  system('R --slave --no-restore -e "parallel:::.slaveRSOCK()" --args
MASTER=localhost PORT=11000 OUT=/dev/null TIMEOUT=2592000 XDR=TRUE',
wait = FALSE)

which immediately tries to connect to a socket on localhost:11000 with
timeout.  Immediately after the master launched the above (without
waiting), it will set up the connection waiting for the connect from
the background worker:

    con <- socketConnection("localhost", port = 11000, server = TRUE,
        blocking = TRUE, open = "a+b", timeout = timeout)


ISSUE:

If we emulate the above process, and remove the OUT=/dev/null such
that we can see the output produces by the worker, as:

setup <- function(delay = 0) {
  system('R --slave --no-restore -e "parallel:::.slaveRSOCK()" --args
MASTER=localhost PORT=11000 TIMEOUT=2592000 XDR=TRUE', wait = FALSE)
  Sys.sleep(delay)
  socketConnection("localhost", port = 11000, server = TRUE, blocking
= TRUE, open = "a+b", timeout = 20)
}

doing:

con <- setup(0)

starting worker pid=24983 on localhost:11000 at 18:44:30.087

will most likely work, but adding a delay:

con <- setup(5)

starting worker pid=25099 on localhost:11000 at 18:45:23.617
Warning in socketConnection(master, port = port, blocking = TRUE, open
= "a+b",  :
  localhost:11000 cannot be opened
Error in socketConnection(master, port = port, blocking = TRUE, open =
"a+b",  :
  cannot open the connection
Calls: <Anonymous> ... doTryCatch -> recvData -> makeSOCKmaster ->
socketConnection

will produce an *instant* error on the worker, and before master opens
the server socket.  Eventually, master will produce the originally
observed error:

Error in socketConnection("localhost", port = 11000, server = TRUE,
blocking = TRUE,  :
  cannot open the connection

In other words, if the master fails to setup socketConnection()
*before* the background workers attempts to connect, it all fails.
Such a delay may happen for instance when there is a large CPU load on
the test machine.

Is this the above bug?

/Henrik

PS. The background is that I, very occasionally, observe R CMD check
error on the above (on CRAN and elsewhere) when testing my future
package. The error always go away when retested. This far I've though
this is due to port clashes (since the port is random selected in
[11000:11999]) and accepted that it happens.  However, after
discovering the above, it could be due to the worker launching "too
soon".

sessionInfo()

R version 3.4.3 (2017-11-30)
Platform: x86_64-pc-linux-gnu (64-bit)
Running under: Ubuntu 16.04.4 LTS

Matrix products: default
BLAS: /usr/lib/atlas-base/atlas/libblas.so.3.0
LAPACK: /usr/lib/atlas-base/atlas/liblapack.so.3.0

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

loaded via a namespace (and not attached):
[1] compiler_3.4.3

--
Luke Tierney
Ralph E. Wareham Professor of Mathematical Sciences
University of Iowa                  Phone:             319-335-3386
Department of Statistics and        Fax:               319-335-3017
   Actuarial Science
241 Schaeffer Hall                  email:   luke-tierney at uiowa.edu
Iowa City, IA 52242                 WWW:  http://www.stat.uiowa.edu