Hi, I have built a Docker image to run psi4 using Ubuntu 16.04, it works fine but it is pretty large (1.5GB).
See my Dockerfile https://github.com/paesanilab/docker-images/blob/master/psi4conda/Dockerfile
build-essential otherwise the conda gcc package wouldn’t install.Thanks!
Hi, I haven’t played around with Docker myself, so I don’t have a smaller image. However the conda packages are changing quite a bit since psi4conda v1.1 with all the new conda build tools. I don’t have a psi4conda installer ready. But would you want to try docker from a Miniconda + conda create -n p4env psi4 ... line that I can give you?
sure! I had a first attempt here but had some issues installing gcc from Anaconda, but I should be able to fix it. https://github.com/paesanilab/docker-images/blob/master/psi4/Dockerfile
Can you please check that the conda line is fine (or send yours)?
So you don’t want to be using plain -c psi4 (aka -c psi4/label/main) channel b/c that’s circa the last 1.1 release from May 2017. The next step beyond that is -c psi4/label/dev which is reputedly nightly build from master. That downloads full gcc 5.2 compilers (I’ll skip why unless you’re interested). There’s a big gap in dev subchannel when psi and conda weren’t in sync on build tools. The most modern psi is actually in -c psi4/label/agg and that should bring along only gcc 7.2 libraries, not compilers, so that may save a little space and linux fundamentals. I’m about to start migrating packages from agg to dev. But what you’re after is something like conda create -n p4env psi4 -c psi4/label/agg -c psi4 . Because it’s just been me on that channel, not even the core devs, no instructions are battle hardened, so unless it’s obvious, ping me when something goes wrong.
Lori, can we build with OpenBLAS and GCC. Should save 600MB+
@dgasmith if this is something I can do in the Dockerfile, if you send me details I can try it out, thanks!
@loriab, I tested the agg branch, good news is that it installs and works fine without installing any other package.
However the output container is still big, it is 1.58 GB.
The base image miniconda3 is just 409 MB, then conda installs:
The following packages will be downloaded:
package | build
---------------------------|-----------------
mkl_random-1.0.1 | py36h629b387_0 373 KB
libefp-1.5b2 | h470d631_0 4.2 MB psi4/label/agg
psi4-1.2a1.dev999+7086126 | py36h4680546_0 24.8 MB psi4/label/agg
numpy-1.14.2 | py36hdbf6ddf_1 4.1 MB
mkl-2018.0.2 | 1 205.2 MB
gau2grid-0.1 | 3 735 KB psi4/label/agg
certifi-2018.4.16 | py36_0 142 KB
py-1.5.3 | py36_0 135 KB
libxc-3.0.0 | h5eb71ee_3 814 KB psi4/label/agg
pytest-3.5.1 | py36_0 292 KB
more-itertools-4.1.0 | py36_0 76 KB
libgfortran-ng-7.2.0 | hdf63c60_3 1.2 MB
mkl_fft-1.0.1 | py36h3010b51_0 140 KB
gdma-2.2.6 | hf4d0741_5 763 KB psi4/label/agg
libint-1.2.1 | h73b9bb0_3 23.9 MB psi4/label/agg
intel-openmp-2018.0.0 | 8 620 KB
attrs-17.4.0 | py36_0 41 KB
openssl-1.0.2o | h20670df_0 3.4 MB
pluggy-0.6.0 | py36hb689045_0 23 KB
ca-certificates-2018.03.07 | 0 124 KB
------------------------------------------------------------
Total: 271.0 MB
The following NEW packages will be INSTALLED:
attrs: 17.4.0-py36_0
gau2grid: 0.1-3 psi4/label/agg
gdma: 2.2.6-hf4d0741_5 psi4/label/agg
intel-openmp: 2018.0.0-8
libefp: 1.5b2-h470d631_0 psi4/label/agg
libgfortran-ng: 7.2.0-hdf63c60_3
libint: 1.2.1-h73b9bb0_3 psi4/label/agg
libxc: 3.0.0-h5eb71ee_3 psi4/label/agg
mkl: 2018.0.2-1
mkl_fft: 1.0.1-py36h3010b51_0
mkl_random: 1.0.1-py36h629b387_0
more-itertools: 4.1.0-py36_0
numpy: 1.14.2-py36hdbf6ddf_1
pluggy: 0.6.0-py36hb689045_0
psi4: 1.2a1.dev999+7086126-py36h4680546_0 psi4/label/agg
py: 1.5.3-py36_0
pytest: 3.5.1-py36_0
The following packages will be UPDATED:
ca-certificates: 2017.08.26-h1d4fec5_0 --> 2018.03.07-0
certifi: 2018.1.18-py36_0 --> 2018.4.16-py36_0
openssl: 1.0.2n-hb7f436b_0 --> 1.0.2o-h20670df_0
however once they are uncompressed they are a lot larger, the /opt/conda/lib folder is 1.2GB, mostly due to mkl (the total of the 38 *mkl* files is 771 MB).
I also created an auto-build on Dockerhub: https://hub.docker.com/r/paesanilab/psi4/
@zonca The process would be to pull down a special conda base and compile the code in-place. This is a bit tricky as you need to pull the same BLAS and OpenMP as the other Python packages used. I think @loriab may have a few magical lines to do this laying around.
To use openblas, you’d have to grab an openblas numpy too (at least mkl is getting multiple uses now) and rebuild any addons with openblas, too. With docker you can build software up, then delete files before forming the final package, right? I bet one could cut out a good swath of mkl and keep only the needed files. On the surface, psi4’s just using libmkl_rt.so. So thread, core, and lp64 can go. Here’s the list for mkl-2018.0.2-1. Do you want to give pruning a try or the openblas route? I’d just rather see psi4+mkl used for production runs.
lib/libmkl_ao_worker.so
lib/libmkl_avx.so
lib/libmkl_avx2.so
lib/libmkl_avx512.so
lib/libmkl_avx512_mic.so
lib/libmkl_blacs_intelmpi_ilp64.so
lib/libmkl_blacs_intelmpi_lp64.so
lib/libmkl_blacs_openmpi_ilp64.so
lib/libmkl_blacs_openmpi_lp64.so
lib/libmkl_blacs_sgimpt_ilp64.so
lib/libmkl_blacs_sgimpt_lp64.so
lib/libmkl_cdft_core.so
lib/libmkl_core.so
lib/libmkl_def.so
lib/libmkl_gf_ilp64.so
lib/libmkl_gf_lp64.so
lib/libmkl_gnu_thread.so
lib/libmkl_intel_ilp64.so
lib/libmkl_intel_lp64.so
lib/libmkl_intel_thread.so
lib/libmkl_mc.so
lib/libmkl_mc3.so
lib/libmkl_pgi_thread.so
lib/libmkl_rt.so
lib/libmkl_scalapack_ilp64.so
lib/libmkl_scalapack_lp64.so
lib/libmkl_sequential.so
lib/libmkl_tbb_thread.so
lib/libmkl_vml_avx.so
lib/libmkl_vml_avx2.so
lib/libmkl_vml_avx512.so
lib/libmkl_vml_avx512_mic.so
lib/libmkl_vml_cmpt.so
lib/libmkl_vml_def.so
lib/libmkl_vml_mc.so
lib/libmkl_vml_mc2.so
lib/libmkl_vml_mc3.so
lib/mkl_msg.cat
Thanks,
I tried pruning MKL independently for lp64, thread and core, but all give an error similar to this:
Intel MKL FATAL ERROR: Cannot load libmkl_core.so.
After some attempts, I got down to 169M and my test case is still running:
65M libmkl_core.so
9.4M libmkl_intel_ilp64.so
11M libmkl_intel_lp64.so
37M libmkl_intel_thread.so
43M libmkl_mc3.so
5.8M libmkl_rt.so
112K mkl_msg.cat
169M total
pretty good gain, the image is now 934 MB, would you have any suggestion to go lower?
ideally 600-700MB would be fine
Can you give a conda list we may be able to eject a number of addons that you may not use.
thanks, here is conda list:
asn1crypto 0.24.0 py36_0
attrs 17.4.0 py36_0
ca-certificates 2018.03.07 0
certifi 2018.4.16 py36_0
cffi 1.11.4 py36h9745a5d_0
chardet 3.0.4 py36h0f667ec_1
conda 4.4.10 py36_0
conda-env 2.6.0 h36134e3_1
cryptography 2.1.4 py36hd09be54_0
gau2grid 0.1 3 psi4/label/agg
gdma 2.2.6 hf4d0741_5 psi4/label/agg
idna 2.6 py36h82fb2a8_1
intel-openmp 2018.0.0 8
libedit 3.1 heed3624_0
libefp 1.5b2 h470d631_0 psi4/label/agg
libffi 3.2.1 hd88cf55_4
libgcc-ng 7.2.0 h7cc24e2_2
libgfortran-ng 7.2.0 hdf63c60_3
libint 1.2.1 h73b9bb0_3 psi4/label/agg
libstdcxx-ng 7.2.0 h7a57d05_2
libxc 3.0.0 h5eb71ee_3 psi4/label/agg
mkl 2018.0.2 1
mkl_fft 1.0.1 py36h3010b51_0
mkl_random 1.0.1 py36h629b387_0
more-itertools 4.1.0 py36_0
ncurses 6.0 h9df7e31_2
numpy 1.14.2 py36hdbf6ddf_1
openssl 1.0.2o h20670df_0
pip 9.0.1 py36h6c6f9ce_4
pluggy 0.6.0 py36hb689045_0
psi4 1.2a1.dev999+7086126 py36h4680546_0 psi4/label/agg
py 1.5.3 py36_0
pycosat 0.6.3 py36h0a5515d_0
pycparser 2.18 py36hf9f622e_1
pyopenssl 17.5.0 py36h20ba746_0
pysocks 1.6.7 py36hd97a5b1_1
pytest 3.5.1 py36_0
python 3.6.4 hc3d631a_1
readline 7.0 ha6073c6_4
requests 2.18.4 py36he2e5f8d_1
ruamel_yaml 0.15.35 py36h14c3975_1
setuptools 38.4.0 py36_0
six 1.11.0 py36h372c433_1
sqlite 3.22.0 h1bed415_0
tk 8.6.7 hc745277_3
urllib3 1.22 py36hbe7ace6_0
wheel 0.30.0 py36hfd4bba0_1
xz 5.2.3 h55aa19d_2
yaml 0.1.7 had09818_2
zlib 1.2.11 ha838bed_2
Nothing from psi will use this one. Should be droppable unless numpy or something needs it.
If you’re planning on running dft, you’ll need to add dftd3 and gcp (psi4 channel)
I’m concerned about 1.2a1.dev999+7086126, as that’s about 200 commits ago. Looks like this line is registering only psi4/label/agg. It needs to be the equiv of [psi4/label/agg, psi4, defaults] in order of greatest to least priority. For packages that don’t change (like dftd3), we keep them on the main channel (hmm, just a label, I guess wouldn’t hurt to duplicate in future). I think the conda solver couldn’t find everything it needed in just [psi4/label/agg, defaults] (which is what you’re line’s accessing) so it fell back on an old psi4 package.
I tried:
python /psi4_test.py
Intel MKL FATAL ERROR: Cannot load libmkl_intel_lp64.so.
no we don’t need dft.
I fixed the installation, now the image is 1GB and here is conda list:
asn1crypto 0.24.0 py36_0
attrs 17.4.0 py36_0
ca-certificates 2018.03.07 0
certifi 2018.4.16 py36_0
cffi 1.11.4 py36h9745a5d_0
chardet 3.0.4 py36h0f667ec_1
chemps2 1.8.7 ha40901b_0 psi4/label/agg
conda 4.4.10 py36_0
conda-env 2.6.0 h36134e3_1
cryptography 2.1.4 py36hd09be54_0
decorator 4.3.0 py36_0
deepdiff 3.3.0 py_0 psi4
dkh 1.2 h173d85e_2 psi4/label/agg
gau2grid 1.0.1 h14c3975_0 psi4/label/agg
gdma 2.2.6 hf4d0741_5 psi4/label/agg
hdf5 1.10.1 h9caa474_1
idna 2.6 py36h82fb2a8_1
intel-openmp 2018.0.0 8
jsonpickle 0.9.6 py_0 psi4
libedit 3.1 heed3624_0
libefp 1.5b2 h470d631_0 psi4/label/agg
libffi 3.2.1 hd88cf55_4
libgcc-ng 7.2.0 h7cc24e2_2
libgfortran-ng 7.2.0 hdf63c60_3
libint 1.2.1 h73b9bb0_3 psi4/label/agg
libstdcxx-ng 7.2.0 h7a57d05_2
libxc 4.0.2 h14c3975_0 psi4/label/agg
mkl 2018.0.2 1
mkl_fft 1.0.1 py36h3010b51_0
mkl_random 1.0.1 py36h629b387_0
more-itertools 4.1.0 py36_0
ncurses 6.0 h9df7e31_2
networkx 2.1 py36_0
numpy 1.14.2 py36hdbf6ddf_1
openssl 1.0.2o h20670df_0
pcmsolver 1.2.0rc2 py36h8733eb9_0 psi4/label/agg
pip 9.0.1 py36h6c6f9ce_4
pluggy 0.6.0 py36hb689045_0
psi4 1.2a1.dev1255+a9aa2db py36hd70b514_0 psi4/label/agg
py 1.5.3 py36_0
pycosat 0.6.3 py36h0a5515d_0
pycparser 2.18 py36hf9f622e_1
pyopenssl 17.5.0 py36h20ba746_0
pysocks 1.6.7 py36hd97a5b1_1
pytest 3.5.1 py36_0
python 3.6.4 hc3d631a_1
readline 7.0 ha6073c6_4
requests 2.18.4 py36he2e5f8d_1
ruamel_yaml 0.15.35 py36h14c3975_1
setuptools 38.4.0 py36_0
six 1.11.0 py36h372c433_1
sqlite 3.22.0 h1bed415_0
tk 8.6.7 hc745277_3
urllib3 1.22 py36hbe7ace6_0
wheel 0.30.0 py36hfd4bba0_1
xz 5.2.3 h55aa19d_2
yaml 0.1.7 had09818_2
zlib 1.2.11 ha838bed_2
it’s ilp64 I think you can drop. psi needs lp64.
Among the psi-supplied pkgs, looks like only libint and psi4 itself break 5 MB, and those are vital, so any savings will have to come from other pkgs (some of which may be deps of psi optional deps).
All the lib*-ng won’t save much
-rw-rw-r--. 1 psilocaluser psilocaluser 6.1M Apr 6 18:01 /home/psilocaluser/toolchainconda/pkgs/libgcc-ng-7.2.0-hdf63c60_3.tar.bz2
-rw-rw-r--. 1 psilocaluser psilocaluser 1.3M Apr 6 18:34 /home/psilocaluser/toolchainconda/pkgs/libgfortran-ng-7.2.0-hdf63c60_3.tar.bz2
-rw-rw-r--. 1 psilocaluser psilocaluser 2.6M Apr 6 18:01 /home/psilocaluser/toolchainconda/pkgs/libstdcxx-ng-7.2.0-hdf63c60_3.tar.bz2
even expanded they’re not too big, so no use just cutting out the fortran addons. (everything in lib >5mb below)
-rwxrwxr-x. 1 psilocaluser psilocaluser 5.6M May 1 16:47 /home/psilocaluser/toolchainconda/envs/p4dev36/lib/libchemps2.so.3
-rwxrwxr-x. 4 psilocaluser psilocaluser 5.6M Mar 8 13:41 /home/psilocaluser/toolchainconda/envs/p4dev36/lib/libgfortran.so.4.0.0
-rwxrwxr-x. 5 psilocaluser psilocaluser 5.8M Mar 22 12:20 /home/psilocaluser/toolchainconda/envs/p4dev36/lib/libmkl_rt.so
-rwxrwxr-x. 2 psilocaluser psilocaluser 6.1M Apr 25 15:38 /home/psilocaluser/toolchainconda/envs/p4dev36/lib/libpcm.so.1
-rwxrwxr-x. 5 psilocaluser psilocaluser 6.2M Mar 22 12:20 /home/psilocaluser/toolchainconda/envs/p4dev36/lib/libmkl_vml_cmpt.so
-rwxrwxr-x. 5 psilocaluser psilocaluser 6.2M Mar 22 12:20 /home/psilocaluser/toolchainconda/envs/p4dev36/lib/libmkl_vml_def.so
-rwxrwxr-x. 6 psilocaluser psilocaluser 7.0M Mar 8 13:41 /home/psilocaluser/toolchainconda/envs/p4dev36/lib/libasan.so.4.0.0
-rwxrwxr-x. 5 psilocaluser psilocaluser 7.2M Mar 22 12:20 /home/psilocaluser/toolchainconda/envs/p4dev36/lib/libmkl_scalapack_ilp64.so
-rwxrwxr-x. 5 psilocaluser psilocaluser 7.2M Mar 22 12:20 /home/psilocaluser/toolchainconda/envs/p4dev36/lib/libmkl_scalapack_lp64.so
-rw-rw-r--. 1 psilocaluser psilocaluser 7.7M May 2 00:47 /home/psilocaluser/toolchainconda/envs/p4dev36/lib/libhdf5.a
-rwxrwxr-x. 6 psilocaluser psilocaluser 7.8M Mar 8 13:42 /home/psilocaluser/toolchainconda/envs/p4dev36/lib/libstdc++.so.6.0.24
-rwxrwxr-x. 5 psilocaluser psilocaluser 9.3M Mar 22 12:20 /home/psilocaluser/toolchainconda/envs/p4dev36/lib/libmkl_gf_ilp64.so
-rwxrwxr-x. 5 psilocaluser psilocaluser 9.4M Mar 22 12:20 /home/psilocaluser/toolchainconda/envs/p4dev36/lib/libmkl_intel_ilp64.so
-rwxrwxr-x. 5 psilocaluser psilocaluser 11M Mar 22 12:20 /home/psilocaluser/toolchainconda/envs/p4dev36/lib/libmkl_gf_lp64.so
-rwxrwxr-x. 5 psilocaluser psilocaluser 11M Mar 22 12:20 /home/psilocaluser/toolchainconda/envs/p4dev36/lib/libmkl_intel_lp64.so
-rwxrwxr-x. 5 psilocaluser psilocaluser 11M Mar 22 12:20 /home/psilocaluser/toolchainconda/envs/p4dev36/lib/libmkl_vml_mc2.so
-rwxrwxr-x. 5 psilocaluser psilocaluser 11M Mar 22 12:20 /home/psilocaluser/toolchainconda/envs/p4dev36/lib/libmkl_vml_mc.so
-rwxrwxr-x. 5 psilocaluser psilocaluser 11M Mar 22 12:20 /home/psilocaluser/toolchainconda/envs/p4dev36/lib/libmkl_vml_mc3.so
-rwxrwxr-x. 5 psilocaluser psilocaluser 12M Mar 22 12:20 /home/psilocaluser/toolchainconda/envs/p4dev36/lib/libmkl_vml_avx512.so
-rwxrwxr-x. 5 psilocaluser psilocaluser 13M Mar 22 12:20 /home/psilocaluser/toolchainconda/envs/p4dev36/lib/libmkl_vml_avx2.so
-rwxrwxr-x. 5 psilocaluser psilocaluser 13M Mar 22 12:20 /home/psilocaluser/toolchainconda/envs/p4dev36/lib/libmkl_vml_avx.so
-rwxrwxr-x. 5 psilocaluser psilocaluser 14M Mar 22 12:20 /home/psilocaluser/toolchainconda/envs/p4dev36/lib/libmkl_vml_avx512_mic.so
-rwxrwxr-x. 3 psilocaluser psilocaluser 19M Jul 6 2017 /home/psilocaluser/toolchainconda/envs/p4dev36/lib/libint.so
-rwxrwxr-x. 5 psilocaluser psilocaluser 19M Mar 22 12:20 /home/psilocaluser/toolchainconda/envs/p4dev36/lib/libmkl_sequential.so
-rwxrwxr-x. 1 psilocaluser psilocaluser 22M Apr 19 15:19 /home/psilocaluser/toolchainconda/envs/p4dev36/lib/libpython3.6m.a
-rwxrwxr-x. 5 psilocaluser psilocaluser 23M Mar 22 12:20 /home/psilocaluser/toolchainconda/envs/p4dev36/lib/libmkl_gnu_thread.so
-rwxrwxr-x. 5 psilocaluser psilocaluser 23M Mar 22 12:20 /home/psilocaluser/toolchainconda/envs/p4dev36/lib/libmkl_tbb_thread.so
-rwxrwxr-x. 5 psilocaluser psilocaluser 36M Mar 22 12:20 /home/psilocaluser/toolchainconda/envs/p4dev36/lib/libmkl_pgi_thread.so
-rwxrwxr-x. 5 psilocaluser psilocaluser 36M Mar 22 12:20 /home/psilocaluser/toolchainconda/envs/p4dev36/lib/libmkl_def.so
-rwxrwxr-x. 5 psilocaluser psilocaluser 37M Mar 22 12:20 /home/psilocaluser/toolchainconda/envs/p4dev36/lib/libmkl_intel_thread.so
-rwxrwxr-x. 5 psilocaluser psilocaluser 42M Mar 22 12:20 /home/psilocaluser/toolchainconda/envs/p4dev36/lib/libmkl_mc.so
-rwxrwxr-x. 5 psilocaluser psilocaluser 43M Mar 22 12:20 /home/psilocaluser/toolchainconda/envs/p4dev36/lib/libmkl_mc3.so
-rwxrwxr-x. 5 psilocaluser psilocaluser 46M Mar 22 12:20 /home/psilocaluser/toolchainconda/envs/p4dev36/lib/libmkl_avx.so
-rwxrwxr-x. 5 psilocaluser psilocaluser 51M Mar 22 12:20 /home/psilocaluser/toolchainconda/envs/p4dev36/lib/libmkl_ao_worker.so
-rwxrwxr-x. 5 psilocaluser psilocaluser 57M Mar 22 12:20 /home/psilocaluser/toolchainconda/envs/p4dev36/lib/libmkl_avx2.so
-rwxrwxr-x. 3 psilocaluser psilocaluser 62M Jul 6 2017 /home/psilocaluser/toolchainconda/envs/p4dev36/lib/libderiv.so
-rwxrwxr-x. 5 psilocaluser psilocaluser 65M Mar 22 12:20 /home/psilocaluser/toolchainconda/envs/p4dev36/lib/libmkl_core.so
-rwxrwxr-x. 5 psilocaluser psilocaluser 68M Mar 22 12:20 /home/psilocaluser/toolchainconda/envs/p4dev36/lib/libmkl_avx512.so
-rwxrwxr-x. 5 psilocaluser psilocaluser 77M Mar 22 12:20 /home/psilocaluser/toolchainconda/envs/p4dev36/lib/libmkl_avx512_mic.so
Ah sorry, I dropped ilp64 already after I posted the list
looks like the only substantial savings left will be mkl --> openblas. do you need libefp or chemps2?
We won’t need libefp or chemps2 (the latter is for dmrg, right?). We need dft, mp2 and coupled cluster energies.
For dft it would be nice to have access to wb97m-v (requires vv10). My understanding is that this should be available if our build is based on the latest development branch.
Correct on chemps2=dmrg and that wb97m-v is in master. Though I’d think you’d want dftd3 also to have -d3 available.
I looked into doing an openblas conda pkg, and it looks straightforward (except for actually selecting it once built). Haven’t attempted yet, as I’ve been busy with the v1.2rc1 release, but I’ll do so soon.