5

I am getting started with installation of Tensorflow on my RHEL 6.5 box. But it turns out that Tensorflow needs glibc >= 2.17 and the default glibc on rhel 6.5 is 2.12.

I was wondering if anybody could help me with minimum/recommended system specifications for tensorflow?

Saheb
  • 1,666
  • 3
  • 18
  • 24
  • I am not flagging this as a duplicate because the OP is not requesting a fix for glibc but instead trying to get TensorFlow working. I might flag it as too broad or off topic because it is not specific enough for StackOverflow. – Guy Coder Dec 30 '15 at 13:54
  • Have you seen [Tensor Flow installation on RHEL 7](http://stackoverflow.com/q/33735833/1243762) or [Error while importing Tensorflow in python2.7 in Red Hat release 6.6. 'GLIBC_2.17 not found'](http://stackoverflow.com/q/33731366/1243762) – Guy Coder Dec 30 '15 at 13:56

2 Answers2

4

The TensorFlow requirements are listed here, but these do not recommend a particular operating system or glibc version.

The best-supported operating systems are Ubuntu 14.04 64-bit, and Mac OS X 10.10 (Yosemite) and later. The current limiting factor is the set of supported operating systems for Bazel, which we use to make the binary packages. You may be able to install Bazel from source, and then install TensorFlow from source, to get around these issues. Many users find it easier to install TensorFlow in a Docker container to avoid this problem.

mrry
  • 125,488
  • 26
  • 399
  • 400
  • Hi @mrry, Thanks for your input. We are running on rhel 6.5 with Kernel v2.6 and even Docker won't work here. Do you think CentOS7/RHEL7 should ideally do it? It comes with Kernel >= 3.1 and glibc >= 2.17 by default. – Saheb Dec 31 '15 at 12:16
  • Yes, we have heard reports of people installing the binary packages on RHEL7 and CentOS 7. Good luck! – mrry Dec 31 '15 at 18:11
  • Though Mac is great, I would go against Mac cause the newer Mac's dont have a cuda supported GPU. I would always do it on a Ubuntu system – Ashik Vetrivelu Feb 23 '17 at 03:57
  • What amounts of memory are usually needed when running tensorflow? Want to start playing with it and got an old desktop PC for the purpose but it only has 2GB RAM and no suitable GPU. Or if I get a good GPU enough memory, would I need extra RAM in the PC? – akostadinov Dec 23 '17 at 09:11
0

Although your system is old and you can not install tensorflow from pip directly, you can install it from source. I installed tensorflow from source successfully. My system is Redhat enterprises 6.0, gcc version is 6.1.0, bazel version is 0.1.4. First you need to install a new gcc version, then you need to compile bazel, which is a tool to compile tensorflow. DO NOT use the newest version of bazel !!!! The 0.1.4 is fine. You can follow this page to find some information.https://github.com/tensorflow/tensorflow/issues/110

Nils Cao
  • 1,409
  • 2
  • 15
  • 23