Redirect stderr to file on network driver is very slow since VC++ 2015 / 14.0 - redirect

Question

We experience a strange performance issue when using Visual C++ 2013 (also known as Visual C++ 12.0) compared to more recent versions like Visual C++ 2015 (also known as Visual C++ 14.0):
The same C code is much slower when writing to stderr and redirecting this stream to a file on a network drive (Server is Linux Debian 8 / Samba).
C:\tmp> myprog.exe 2>X:\devel\fgl\tests\zz
When redirecting to a file on the local disk, or when redirecting to nul, the execution is fast.
We have also tested with VC++ 2017 / 15.0, same problem.
The reason seams to be the /MD option (link with MSVCRT.LIB), I could reproduce with a simple program:
#include <stdio.h>
int main(int argc, char **argv)
{
int n;
system("echo %TIME%");
for (n=0; n<200; n++) {
fprintf(stderr,"aaaaaaaaaaaaaaaaaaaaaaaaaaaa %d\n", n);
}
system("echo %TIME%");
return 0;
}
Then compile with and without /MD option:
C:\tmp>cl /O1 myprog.c
Microsoft (R) C/C++ Optimizing Compiler Version 19.11.25507.1 for x64
Copyright (C) Microsoft Corporation. All rights reserved.
myprog.c
Microsoft (R) Incremental Linker Version 14.11.25507.1
Copyright (C) Microsoft Corporation. All rights reserved.
/out:myprog.exe
myprog.obj
C:\tmp>myprog.exe 2>X:\devel\fgl\tests\zz
9:32:27.96
9:32:27.97
C:\tmp>cl /O1 /MD myprog.c
Microsoft (R) C/C++ Optimizing Compiler Version 19.11.25507.1 for x64
Copyright (C) Microsoft Corporation. All rights reserved.
myprog.c
Microsoft (R) Incremental Linker Version 14.11.25507.1
Copyright (C) Microsoft Corporation. All rights reserved.
/out:myprog.exe
myprog.obj
C:\tmp>myprog.exe 2>X:\devel\fgl\tests\zz
9:32:37.15
9:32:40.12
C:\tmp>myprog.exe 2>c:\tmp\zz
9:32:47.15
9:32:47.20
Is there any known issue related to stderr / network drives with VC++?
Notes:
No problem when writing to stdout
No problem when using /MDd option (MSVCRTD.lib)
No problem with Visual C++ 2013 (Visual C++ 12.0, cl version 18.00.40629)
Thanks for reading!

Answer 1

Er, we have been experiencing the same issue with our C++ code so I added my comment using C++, initially. Sorry about that.
For C, the same "remedy" would look like this:
char buf[4096];
setvbuf(stderr, _IOLBF/*or _IOFBF*/, buf, sizeof(buf));
Initial response:
Try this at the beginning of your program:
static char buf[16384];
cerr.rdbuf()->pubsetbuf(buf, sizeof(buf));
cerr.unsetf(std::ios_base::unitbuf);
You can decrease the buffer size (to 4096 or less) and still see the speed boost -- the thing is that cerr is completely unbuffered by default, and having unitbuf set on it causes even more slowdown by implicit flush operations injected after each output to the stream.