Skip to content

Memory leak with tons of closed connections

6 messages · Martin Maechler, Gergely Daróczi, Gábor Csárdi

Original
Gergely Daróczi
# 
Dear All,

I'm developing an R application running inside of a Java daemon on
multiple threads, and interacting with the parent daemon via stdin and
stdout.

Everything works perfectly fine except for having some memory leaks
somewhere. Simplified version of the R app:

    while (TRUE) {
        con <- file('stdin', open = 'r', blocking = TRUE)
        line <- scan(con, what = character(0), nlines = 1, quiet = TRUE)
        close(con)
    }

This loop uses more and more RAM as time passes (see more on this
below), not sure why, and I have no idea currently on how to debug
this further. Can someone please try to reproduce it and give me some
hints on what is the problem?

Sample bash script to trigger an R process with such memory leak:

    Rscript --vanilla -e "while(TRUE)cat(runif(1),'\n')" | Rscript
--vanilla -e "cat(Sys.getpid(),'\n');while(TRUE){con<-file('stdin',open='r',blocking=TRUE);line<-scan(con,what=character(0),nlines=1,quiet=TRUE);close(con);rm(con);gc()}"

Maybe you have to escape '\n' depending on your shell.

Thanks for reading this and any hints would be highly appreciated!

Best,
Gergely

PS1 see the image posted at
http://stackoverflow.com/questions/40522584/memory-leak-with-closed-connections
on memory usage over time
PS2 the issue doesn't seem to be due to writing more data in the first
R app compared to what the second R app can handle, as I tried the
same with adding a Sys.sleep(0.01) in the first app and that's not an
issue at all in the real application
PS3 I also tried using stdin() instead of file('stdin'), but that did
not work well for the stream running on multiple threads started by
the same parent Java daemon
PS4 I've tried this on Linux using R 3.2.3 and 3.3.2
Martin Maechler
# 
> Dear All,
    > I'm developing an R application running inside of a Java daemon on
    > multiple threads, and interacting with the parent daemon via stdin and
    > stdout.

    > Everything works perfectly fine except for having some memory leaks
    > somewhere. Simplified version of the R app:

    > while (TRUE) {
    > con <- file('stdin', open = 'r', blocking = TRUE)
    > line <- scan(con, what = character(0), nlines = 1, quiet = TRUE)
    > close(con)
    > }

    > This loop uses more and more RAM as time passes (see more on this
    > below), not sure why, and I have no idea currently on how to debug
    > this further. Can someone please try to reproduce it and give me some
    > hints on what is the problem?

    > Sample bash script to trigger an R process with such memory leak:

    > Rscript --vanilla -e "while(TRUE)cat(runif(1),'\n')" | Rscript
    > --vanilla -e "cat(Sys.getpid(),'\n');while(TRUE){con<-file('stdin',open='r',blocking=TRUE);line<-scan(con,what=character(0),nlines=1,quiet=TRUE);close(con);rm(con);gc()}"

    > Maybe you have to escape '\n' depending on your shell.

    > Thanks for reading this and any hints would be highly appreciated!

I have no hints, sorry... but give some more "data":

I've changed the above to *print* the gc() result every 1000th
iteration, and after 100'000 iterations, there is still no
memory increase from the point of view of R itself.

However, monitoring the process (via 'htop', e.g.) shows about
1 MB per second increase in memory foot print of the process.

One could argue that the error is with the OS / pipe / bash
rather than with R itself... but I'm not expert enough to do
argue  here at all.

Here's my version of your sample bash script and its output:

$  Rscript --vanilla -e "while(TRUE)cat(runif(1),'\n')" | Rscript --vanilla -e "cat(Sys.getpid(),'\n');i <- 0; while(TRUE){con<-file('stdin',open='r',blocking=TRUE);line<-scan(con,what=character(0),nlines=1,quiet=TRUE);close(con);rm(con);a <- gc(); i <- i+1; if(i %% 1000 == 1) {cat('i=',i,'\\n'); print(a)} }"

11059 
i= 1 
         used (Mb) gc trigger  (Mb) max used (Mb)
Ncells  83216  4.5   10000000 534.1   213529 11.5
Vcells 172923  1.4   16777216 128.0   562476  4.3
i= 1001 
         used (Mb) gc trigger  (Mb) max used (Mb)
Ncells  83255  4.5   10000000 534.1   213529 11.5
Vcells 172958  1.4   16777216 128.0   562476  4.3
.......
...............................................
...............................................
...............................................
i= 80001 
         used (Mb) gc trigger  (Mb) max used (Mb)
Ncells  83255  4.5   10000000 534.1   213529 11.5
Vcells 172958  1.4   16777216 128.0   562476  4.3
i= 81001 
         used (Mb) gc trigger  (Mb) max used (Mb)
Ncells  83255  4.5   10000000 534.1   213529 11.5
Vcells 172959  1.4   16777216 128.0   562476  4.3
i= 82001 
         used (Mb) gc trigger  (Mb) max used (Mb)
Ncells  83255  4.5   10000000 534.1   213529 11.5
Vcells 172959  1.4   16777216 128.0   562476  4.3
i= 83001 
         used (Mb) gc trigger  (Mb) max used (Mb)
Ncells  83255  4.5   10000000 534.1   213529 11.5
Vcells 172958  1.4   16777216 128.0   562476  4.3
i= 84001 
         used (Mb) gc trigger  (Mb) max used (Mb)
Ncells  83255  4.5   10000000 534.1   213529 11.5
Vcells 172958  1.4   16777216 128.0   562476  4.3


    > Best,
    > Gergely

    > PS1 see the image posted at
    > http://stackoverflow.com/questions/40522584/memory-leak-with-closed-connections
    > on memory usage over time
    > PS2 the issue doesn't seem to be due to writing more data in the first
    > R app compared to what the second R app can handle, as I tried the
    > same with adding a Sys.sleep(0.01) in the first app and that's not an
    > issue at all in the real application
    > PS3 I also tried using stdin() instead of file('stdin'), but that did
    > not work well for the stream running on multiple threads started by
    > the same parent Java daemon
    > PS4 I've tried this on Linux using R 3.2.3 and 3.3.2

For me, it's Linux, too (Fedora 24), using  'R 3.3.2 patched'..

Martin
Gergely Daróczi
# 
On Fri, Nov 11, 2016 at 12:08 PM, Martin Maechler
<maechler at stat.math.ethz.ch> wrote:
Thank you very much, this was very useful!

I tried to do some more research on this, as Gabor Csardi also
suspected that the memory grow might be due to the writer being faster
than the reader, so data is simply accumulating in the input buffer of
the reader. I double checked this via:

    Rscript --vanilla -e
"i<-1;while(TRUE){cat(runif(1),'\n');i<-i+1;if(i==1e6){Sys.sleep(15);i<-1}}"
| Rscript --vanilla -e
"cat(Sys.getpid(),'\n');i<-0;while(TRUE){con<-file('stdin',open='r',blocking=TRUE);line<-scan(con,what=character(0),nlines=1,quiet=TRUE);close(con);rm(con);a<-gc();i<-i+1;if(i%%1e3==1){cat('i=',i,'\\n');print(a)}}"scan(con,what=character(0),nlines=1,quiet=TRUE);close(con);rm(con);gc()}"

So the writer generates a good number of lines, but sleeps for 15
seconds after a while so that the reader can catch up. Monitoring the
memory footprint of the process (by the way gc reported no memory
increase in the reader, just like in Martin's output) shows that the
memory grows when the writer sends data, and it's constant when the
writer is sleeping, but it never decreases: http://imgur.com/r7T02pK

Maybe it's more like an OS-specific question based on this, you are
absolutely right, but I was not able to reproduce the same memory
issue in plain bash via:

    while :;do echo '1';done | bash -c "while :;do read;done"

But I'm not sure if this does exactly the same as the original R
script, so this is rather just a guess.

On the other hand, I tried to modify the original minimal R script in
other ways as well to see which part might result in the strange
memory growth, and it seems that opening the connection once but
keeping the rest of the script (so still generating and reading tons
of lines without any sleep), did not show any memory leak:

    Rscript --vanilla -e "while(TRUE)cat(runif(1),'\n')" | Rscript
--vanilla -e "cat(Sys.getpid(),'\n');con<-file('stdin',open='r',blocking=TRUE);while(TRUE){line<-scan(con,what=character(0),nlines=1,quiet=TRUE);};close(con)"

Based on this, I think I can (should) modify my R application to open
stdin only once and read from that connection in the infinite loop,
but I'm still interested in understanding what's causing the extra
memory usage when opening and closing many connections (if my above
findings are correct).

Thank you very much again, and I'm still looking for any suggestion or
advice on how to debug this further.
Gábor Csárdi
# 
On Fri, Nov 11, 2016 at 12:46 PM, Gergely Dar?czi
<daroczig at rapporter.net> wrote:
[...]
Yes, R does not know about it, it does not manage this memory (any
more), but the R process requested this memory from the OS, and never
gave it back, which is basically the definition of a memory leak. No?

I think the leak is because 'stdin' is special and R opens it with fdopen():
https://github.com/wch/r-source/blob/f8cdadb769561970cc42776f563043ea5e12fe05/src/main/connections.c#L561-L579

and then it does not close it:
https://github.com/wch/r-source/blob/f8cdadb769561970cc42776f563043ea5e12fe05/src/main/connections.c#L636

I understand that R cannot fclose the FILE*, because that would also
close the file descriptor, but anyway, this causes a memory leak. I
think.

It seems that you cannot close the FILE* without closing the
descriptor, so maybe a workaround would be to keep one FILE* open,
instead of calling fdopen() to create new ones every time. Another
possible workaround is to use dup(), but I don't know enough about the
details to be sure.

Gabor

[...]
2 days later
Gábor Csárdi
# 
Using dup() before fdopen() (and calling fclose() on the connection
when it is closed) indeed fixes the memory leak.

FYI,
Gabor

Index: src/main/connections.c
===================================================================
--- src/main/connections.c (revision 71653)
+++ src/main/connections.c (working copy)
@@ -576,7 +576,7 @@
     fp = R_fopen(name, con->mode);
     } else {  /* use file("stdin") to refer to the file and not the console */
 #ifdef HAVE_FDOPEN
- fp = fdopen(0, con->mode);
+        fp = fdopen(dup(0), con->mode);
 #else
  warning(_("cannot open file '%s': %s"), name,
  "fdopen is not supported on this platform");
@@ -633,8 +633,7 @@
 static void file_close(Rconnection con)
 {
     Rfileconn this = con->private;
-    if(con->isopen && strcmp(con->description, "stdin"))
- con->status = fclose(this->fp);
+    con->status = fclose(this->fp);
     con->isopen = FALSE;
 #ifdef Win32
     if(this->anon_file) unlink(this->name);
On Fri, Nov 11, 2016 at 1:12 PM, G?bor Cs?rdi <csardi.gabor at gmail.com> wrote:
Martin Maechler
# 
Thank you, G?bor!
Yes I can confirm that this fixes the memory leak.

I'm testing ('make check-all') currently and then (probably) will
commit the patch.... R-devel only for the time being.

Martin