Pre-compiling images for different CPUs on GCloud

[Pavel]

I am deploying a CPU-intensive application on Google Cloud Compute with Julia code running inside a Docker container. GCloud instances have a few different CPU architectures depending on the zone. Ideally, I would like to pre-compile Julia images (i.e. different docker container images) for all of those architectures.

Currently the following runs on container launch:

    include(joinpath(JULIA_HOME, Base.DATAROOTDIR, "julia", "build_sysimg.jl"))
    build_sysimg(joinpath(dirname(Libdl.dlpath("libjulia")),"sys"), "native", "/home/juser/jimg.jl", force=true)

so that the "native" setting pre-compiles for the CPU that these commands are running on. That takes time and having that done at the container image build time would save cloud runtime.

Would it be possible to pre-compile for different CPUs on my development machine or do I need to run the pre-compilation on those exact CPUs? If the former is possible, what build_sysimg arguments would help? GCloud virtual machine architectures of interest include Sandy Bridge, Ivy Bridge, and Haswell CPU types.

Thanks for reading,
Pavel

[Milan Bouchet-Valat]

You should be able to build images for older architectures on a machine
with a newer one, i.e. build images for Sandy Bridge and Ivy Bridge on
a Haswell machine. I don't think it would work in the other direction,
since some instructions wouldn't be supported and the code wouldn't
run.

So in your case you could replace "native" with "sandybridge",
"ivybridge", "haswell", or any value from the list printed by
llc -mattr=help.


Regards

[Pavel]

I tried this approach using a local machine with
product: Intel(R) Core(TM) i7-4710MQ CPU @ 2.50GHz
which in my understanding belongs to the "haswell" type.

setting
    build_sysimg(joinpath(dirname(Libdl.dlpath("libjulia")),"sys"), "haswell", "/home/juser/jimg.jl", force=true)

resulted in
    'haswell' is not a recognized processor for this target (ignoring processor)

On the other hand, setting "native" instead, after deploying on GCloud haswell-zone says that the image was precompiled for a different architecture.

[Milan Bouchet-Valat]

Le vendredi 05 février 2016 à 18:24 -0800, Pavel a écrit :
> I tried this approach using a local machine with
> product: Intel(R) Core(TM) i7-4710MQ CPU @ 2.50GHz
> which in my understanding belongs to the "haswell" type.
>
> setting
>     build_sysimg(joinpath(dirname(Libdl.dlpath("libjulia")),"sys"),
> "haswell", "/home/juser/jimg.jl", force=true)
>
> resulted in 
>     'haswell' is not a recognized processor for this target (ignoring
> processor)

Have you looked at the output of llc -mattr=help (or usr/bin/llc
-mattr=help from the Julia install/build directory)? If this is with
LLVM 3.3, then it doesn't look like it supported "haswell" as a target.
You should be able to specify "core-avx2" instead.

> On the other hand, setting "native" instead, after deploying on
> GCloud haswell-zone says that the image was precompiled for a
> different architecture.

By definition, "native" isn't intended to be portable, so it doesn't
sound unexpected.


Regards

[Erik Schnetter]

You can run `llc --version` to determine the host's CPU type. You
should probably use the same llc that comes with the LLVM that Julia
is using.

The exact command is

`llc --version | awk '/Host CPU:/ { print $3; }')`

-erik

--
Erik Schnetter <schn...@gmail.com>
http://www.perimeterinstitute.ca/personal/eschnetter/

[Pavel]

The "core-avx2" target worked (LLVL 3.4), and it was in the output of `llc --version`. I was able to precompile locally and deploy on GCloud haswell. The run time of my test suite was effectively the same with that and with the "native" option precompilation. Thank you Milan and Erik for your input.

[Erik Schnetter]

An even better command is:
```
julia -e 'println(Sys.cpu_name)'
```

-erik