Tensorflow on LXPLUS

Using the LCG installed stuff

source /cvmfs/sft.cern.ch/lcg/views/LCG_88/x86_64-slc6-gcc49-opt/setup.sh

I intall tensorflow into the local HOME folder.

pip install --user tensorflow

Unfortunately it doesn't run due to a GLIBC conflict

http://lightofdawn.org/wiki/wiki.cgi/-wiki/NewAppsOnOldGlibc

readelf -V .local/lib/python2.7/site-packages/tensorflow/python/_pywrap_tensorflow_internal.so
readelf -s .local/lib/python2.7/site-packages/tensorflow/python/_pywrap_tensorflow_internal.so

The table .gnu.version_r starts at file offset 0x88a374
The symbol GLIBC_2.14 is located at offset 0x0040 (+ 0x88a374 = 0x88a3b4)
The symbol GLIBC_2.17 is located at offset 0x00a0 (+ 0x88a374 = 0x88a414)
The symbol GLIBC_2.15 is located at offset 0x00c0 (+ 0x88a374 = 0x88a434)
The symbol GLIBC_2.16 is located at offset 0x00e0 (+ 0x88a374 = 0x88a454)

Editing with vim

vim .local/lib/python2.7/site-packages/tensorflow/python/_pywrap_tensorflow_internal.so

Switch to the hex editor mode and go to the corresponding line:

:%!xxd
/88a3b0

Find the flags 2-bytes (start at offset 88a3b8) change 0000 -> 0200.
Repeat for all the libraries. Then save with:

:%!xxd -r
:wq

Check that readelf -v says Flags: WEAK for the four libraries
Check that python complains about the missing memcpy on import tensorflow

Here is the stubs file for the functions that tensorflow wants:

#include <time.h>
#include <string.h>
#include <stdlib.h>
void* memcpy(void *dest, const void *src, size_t n) { return memmove(dest, src, n); }
char *secure_getenv(const char *name) { return getenv(name); }

#include <sys/poll.h>
int __poll_chk (struct pollfd *fds, nfds_t nfds, int timeout, __SIZE_TYPE__ fdslen)
{
    if (fdslen / sizeof (*fds) < nfds)
        __chk_fail ();
    return __poll (fds, nfds, timeout);
}

Compile with:

gcc -s -shared -o mylibc.so -fPIC -fno-builtin mylibc.c

And run python (or jupyter with)

LD_PRELOAD=$PWD/mylibc.so python

KKostya/TF.md