Can switching from GNOME to KDE improve performance of tensorflow-gpu?

alanmei · 2020-04-15 02:26:05

Given my parameters of deep learning the same, I found that the performance of my model improved about 100% compared with last time. I cannot explain it. My system changed in the following aspects:
before: GNOME3+python3.8.2+python-tensorflow-cuda2.1.0+intel-mkl----4.25min
now: KDE+python3.8.2+python-tensorflow-cuda2.2.0rc2+intel-parallel-studio-xe----2.25min and KDE+intelpython3.7.7+tensorflow-gpu2.1.0+intel-parallel-studio-xe---3.33min (but didn't install python-numpy-mkl)
Can anyone explain why the training time decreased that much?

Last edited by alanmei (2020-04-15 02:37:17)

Trilby · 2020-04-15 02:34:42

Explain what? That you used different computational packages with different configurations and optimizations and got different times? There's not a whole lot to explain there (unless 'Duh' is a suitable explanation). I doubt GNOME vs KDE has any detectable effect though.

That said, you are the only one equipped to really explain the differences. Systematically profile swapping out one component at a time (rather than changing several variables at once) to identify which variables have the biggest impact on computation time. My money will remain on the DE choice being the absolute least relevant variable.

As an aside, I find it ironic that you'd be working on this kind of task yet describe the change in performance as almost 100% faster. That's either meaningless, or your improved time would be closed to 0.0 seconds. Your data show the second attempt was almost 50% faster than the first.

Last edited by Trilby (2020-04-15 02:38:58)

alanmei · 2020-04-15 02:59:17

Since the inputs of the model are the same in the comparison, we can see that the total time become almost half before, which means the data it processes is twice as much as before. Therefore, it's a 100% performance uplift.
Second, I'm new to arch linux and have little knowledge of Desktop environment. I post it here just to know whether desktop environment has effect on GPU performance if anyone have similar experience.

Trilby · 2020-04-15 03:09:41

Yes, performance as in your edited post did increase almost 100%. But no, DE will not have any effect on GPU performance. There could be trivial effects from one being more or less bloated than the other and using it's share of CPU or memory, but this again would have no detectable difference in what you are doing. But the other changes in your configuration could be expected to have substantial impact which is why I'm surprised that you are asking about the DE.

Cutting through a 2x4 with a chainsaw while on the beach would be a lot faster than cutting through a 2x4 with a nail file in your garage. This wouldn't lead me to suspect that bringing sand into the garage might help cut 2x4s faster.

alanmei · 2020-04-15 03:26:56

Thanks for you explanation, I'll check it from other aspects.

Arch Linux

#1 2020-04-15 02:26:05

Can switching from GNOME to KDE improve performance of tensorflow-gpu?

#2 2020-04-15 02:34:42

Re: Can switching from GNOME to KDE improve performance of tensorflow-gpu?

#3 2020-04-15 02:59:17

Re: Can switching from GNOME to KDE improve performance of tensorflow-gpu?

#4 2020-04-15 03:09:41

Re: Can switching from GNOME to KDE improve performance of tensorflow-gpu?

#5 2020-04-15 03:26:56

Re: Can switching from GNOME to KDE improve performance of tensorflow-gpu?

Board footer