Building Yocto on the Xeon Phi?

[Ulf_S_]

Building a yocto filesystem can be quite time consuming, and is a highly parallel activity.

Anyone that as tried building a Yocto filesystem running the build on the Xeon Phi?

[Frances_R_Intel]

People doing a standard install of the MPSS (including those who will be adding in the optional software, such as OFED) will not notice much difference between the pre-Yocto and post-Yocto versions other than that some files have been moved to new locations and that some command line options have changed. I know that I was able to update systems with a pre-release version with no problems.

Similarly, application developers will find little has changed. For them the most noticeable change will probably be some library locations.

The people who will notice the biggest change will be those who customize their system (for example modifying and/or rebuilding kernel modules) AND who choose to use the Yocto tooling to do so. Notice the AND. If you need to recompile the mic.ko kernel driver because you are using an unsupported version of Linux on your host processor, you are not forced to go through a full Yocto build. I say this so that those (like me) who are inexperienced with Yocto, won't panic.

That said, I am sure some people will want to take full advantage of the move to Yocto, especially those maintaining multiple systems or clusters using modified or recompiled versions of the MPSS. When the new version of the MPSS is released, I will be interested in seeing what people's experiences are.

[Ulf_S_]

Thanks, Maybe I wasn't too clear.

I am wondering if it is possible to bitbake an image for another Yocto target (ARM, PowerPC, Embedded x86 etc.)

ON the Xeon Phi and get a speedup over bitbaking on my Core i7 780X Machine.

[James_C_Intel2]

Ulf S. wrote:

Thanks, Maybe I wasn't too clear.

I am wondering if it is possible to bitbake an image for another Yocto target (ARM, PowerPC, Embedded x86 etc.)

ON the Xeon Phi and get a speedup over bitbaking on my Core i7 780X Machine.

That seems very unlikely. Last time I looked compiling an operating system was not a highly parallel vector floating point application :-)

It also seems very unlikely that the cross compilers to run on the Xeon Phi and produce code for those machines exist.

[Ulf_S_]

Certainly not Vector floating point, bit definitely parallell.

If you can run a parallell make and a c compiler on the Xeon Phi, then it is a good start. The cross compilers will then be built by Yocto

[James_C_Intel2]

But, remember that integer serial code on the Xeon Phi runs much slower than on the Xeon (maybe up to 10x slower), and all your files are further away. This still seems a very long shot to me.

For this task I would spend the money on a faster Xeon, a solid state drive for the build machine ( http://www.intel.com/content/www/us/en/solid-state-drives/solid-state-drives-530-series.html ), more memory (or all of the above), before I started trying to use a Xeon Phi.

My very strong guess is that those will be much easier routes to faster OS build performance. (Just because you can do something it doesn't mean you should, or that it's the best use of your time or money. The simple and obvious ways to do things may be simple and obvious because they are the best!)