Help with Intel compilation

[Barry Moore]

Hello All,

I will preface this by stating I am not a CP2K user. I support CP2K users.

I have compiled CP2K a few times using the toolchain style with all of the

plugins (4.1, 5.1 w/ Intel MKL 2017.1.132). I have found that if I use Intel

MKL 2017.3.196 neither version 4.1 nor 5.1 passes regression tests. Anyway, recently a

student showed me some scaling plots and they are atrocious. I need to

reconcile this because the major users at our center are all using CP2K and I

am basically throwing away computer time on one of our busiest resources.

So, I built an arch file starting from Linux-x86-64-intel-mic.psmp (`h2p.psmp`

attached). I ignored ACC & MIC stuff and tried to simplify the MKL section.

 I installed libxc and libint by modifying the toolchain scripts

(scripts attached). Finally, I attach the regression log

(`cp2k-regtest-5.1.log`). Any help is appreciated.

Thanks,

Barry

[Alfio Lazzaro]

Hello Barry,

The arch you are using (Linux-x86-64-intel-mic.psmp) is pretty old and messy (it refers to Intel KNC with offload mode). Sorry about that, I will update it for the next release.

Now, we do test CP2K with Intel compiler in our dashboard: https://dashboard.cp2k.org/

For instance, this test

http://cp2k-www.epcc.ed.ac.uk/phi/popt/regtest-0

has an arch file (see on top of the log). We test ICC version 17.0.4, which is pretty outdated now. 

It can be that there are some bugs in the newer ICC version (we have seen this in the past). I would suggest you use the arch file from the dashboard, set -O0 mode, and try to run the regtests. Intel people told us to use "-fp-model precise" can be useful as well... 

Alfio

[Barry Moore]

Alfio,

Thanks a lot for that tip, I am working on compiling/regtesting now. Fingers crossed.

I do think cleaning up the arch directory would be really helpful for administrators. 

Best,

Barry

[Barry Moore]

For reference, this didn't work. Same failures.

[Barry Moore]

Alfio,

I borrowed a Linux-x86-64-intel-mic.psmp which I know to be working somewhere else (we are regtesting that version now) for cp2k 4.1. It is definitely a lot better than before, but I still get some failures. I attach the arch file and error_summary. Any comments?

- Barry

[Alfio Lazzaro]

Oh yes, the situation is definitely better... 

Concerning the wrong results, I noticed that you are using 12 ranks in your regtest (is this correct?). Some of the tests have references that depend on the number of ranks (usually we set it to 2 or 4 ranks), therefore I assume that if your run with 2 ranks the problem will go away. Could you test it? Then, there are only *primary* 4 tests that are failing. At this point, I don't know where the problem is, but likely it is an effect of the 12 ranks...

Alfio

[Barry Moore]

Alfio,

I will try that. I also would like to see if I can build a working cp2k 5.1 install with Intel compilers, as well as the toolchain style builds w/ 4.1 and 5.1.

By the way, the trick to getting this to run fast was removing libint. A lot of other QM packages are using that code with good speedups, but the recommendations/defaults in CP2K don't seem to yield good performance at our center.

- Barry

[Barry Moore]

Alfio,

With 2 ranks, there is less failures 8 (down from 26) but more wrong results 16 (up from 8).

- Barry