Python - Using A Xeon Phi With Multiprocessing?

The Old Myron Siren · February 7, 2021

Hopefully this is in the right category, if not my apologies!

I've been toying around with multiprocessing in Python the last few days and want to see just how far I can push some it. Currently I have a VM running Ubuntu Server that is for a Discord bot I use for testing purposes. I've broken the bot up into multiple processes so that I can process more commands at the same time (although it's unnecessary and not needed at all for its use, I just wanna have fun). In my pursuit of trying to run as many bots as I can, I remembered vaguely about the old Xeon Phi co processors for days past. I have zero understanding of how to work with something like that though. I know I need to make sure my motherboard supports co processers to begin with but beyond that I'm not sure. Would I have to modify my current Python files and implement other libraries to be able to utilize the Phi?

Grabhanem · February 7, 2021

I'd guess it's not worth it - those coprocessors don't just add cores to your system, they run a stripped-down Linux installation of their own and need to have programs specifically written for them. Think of it less as adding a second CPU to your system and more as adding another computer to your network that you can remotely access via command line.

Another concern is heat - most Phi cards are passively cooled, so you'd need to duct tape a fan to them somehow to keep them from overheating, as they're designed for wind-tunnel-like server airflow paths.

Rugg · February 8, 2021

This is vastly above my knowledge. However, I did some searching, and I found that your project should be possible. Check out some code samples from intel. This raises a better question, would this even the right solution for you? Running something cloud based seems like a more scalable option than building your own HPC server.

The Old Myron Siren · February 8, 2021

General question: Would I be better off just building a new system using the actual socketable CPUs rather than a co processor?

5 hours ago, Grabhanem said:

I'd guess it's not worth it - those coprocessors don't just add cores to your system, they run a stripped-down Linux installation of their own and need to have programs specifically written for them. Think of it less as adding a second CPU to your system and more as adding another computer to your network that you can remotely access via command line.

Another concern is heat - most Phi cards are passively cooled, so you'd need to duct tape a fan to them somehow to keep them from overheating, as they're designed for wind-tunnel-like server airflow paths.

I'm fine with having to SSH into them and treating them like their own computers. I have a rack that I could build a system into and just let it sit. I figured they would be different since its not like I'm adding an extra socket to a board. My main goal here is to get as many cores as I can while spending the least amount of money possible (aren't we all?) As far as heat goes, I have plenty of fans I don't mind cramming into a chassis or even mounting them to the card to keep em cool.

4 hours ago, Rugg said:

This is vastly above my knowledge. However, I did some searching, and I found that your project should be possible. Check out some code samples from intel. This raises a better question, would this even the right solution for you? Running something cloud based seems like a more scalable option than building your own HPC server.

I'm sure its not the best solution for sure, however I'd wanna be running the machine myself and go along for the ride with it for the sake of it. I might end up having to do something cloud based if that's what it comes to though. But if possible, I'd like to build the rig myself and kinda let it rip.

1 hour ago, gabrielcarvfer said:

If you manage to get your hands on a bootable Phi and a compatible board (Linus even showed one of them and tried to play games on it), you could run your stuff as is. For the accelerator cards, you usually need to ssh into it or run specially crafted programs that will do that for you. Too much trouble if you don't have one laying around.

The specially crafted program is the bit that I'm worried about. If the Phi Co Processor is treated like a secondary system then could I install python packages to it and almost treat it like a VM? Or do I need to rerwrite my current Python files to be able to access/utilize the Phi?

The Old Myron Siren · February 8, 2021

40 minutes ago, gabrielcarvfer said:

I had the opportunity, but ended up not using the Phi available in the lab, so I'm not completely sure.

I remember Intel had an SDK for compiled programs.

Looking at the Intel mailing lists I've found this, for earlier Phis:

https://github.com/intel/pyMIC

Stuff that supposedly can get automatically accelerated use the Intel MKL (for BLAS, LAPACK, FFTs). Their own python distribution has it baked in.

https://colfaxresearch.com/isc16-intel-python/

I might be better off just building a cluster of raspberry pi 4s and using something like ZMQ to have them all communicate together. I'd have to add a third layer of management to distribute the commands but I think I could get away with having 6-8 worker bots per Pi.