Shared GPU resources Pool via OpenCL for use inside (ADM?) Containerised Environments

Just about everything is virtual these days. Topics relating to virutalization go here.

Moderator: Lillian.W@AST

Post Reply
majika
Posts: 5
youtube meble na wymiar Warszawa
Joined: Fri Feb 03, 2023 2:46 am

Shared GPU resources Pool via OpenCL for use inside (ADM?) Containerised Environments

Post by majika »

Question/Idea/Avenues to Approach for a possible solution:

My question occured when configuring Portainer and noticed an option for GPU resource pool.

I have a LOT of Powerful (NVIDIA 1080's, 2070's, 3070's, 3080's 24G Titan /AMD 470's/480's , 580's XT6800's [Ranging from 4GB - 24GB capaity Cards running GDDR5X, GDDR6X, HBM2 Memory types) GPU's which are sitting around, doing a lot of nothing atm...

Is this technically feisable or just plain impossible to achieve a shared Network-Attached compute cluster? Even if there is a way to use the GPU(s)-MEM as a mini-cache store, or something? (as to not worry too much about latency.)

I do have past experiance running both mining pools/rigs and coding miners and developing some neat crypto algos (from back in the day) and thus can find my way around OpenCL API / ASM / CPP incl all the usual backend MVC WebApp Devel/REPL nodejs stuff too.

1st hurdle will be the Low-level kernel challanges. Next, There will be middleware challanges But the idea is appealing none the less..

I am considering if there either an already existing solution available already or if there *could be* a new method to *Somehow* use a Local old x12 gpu-card mining test rig (based on the B250 mining expert x19 pcie platform) to host the GPU's on a seperate machine (The mining rig) and figure out if just like in crypt0 mining there is the stratum://<url>:<port>/threads-n0+1/ protocol (like a proxy) to poll WUs worker units to enable collaborative mining efforts on large workloads in mining pools, the same can be done for Docker or making GPU Compute resorces available to Host containers through the local network (eventually, maybe even remotely over the WAN for consuming on remote infrastructure)

The first step would be to see if this is even possible to translate into a protocol to allocate/deligate compute resources to a Docker environment then see if it will work on ADM/Asustor NAS.???

I am just spit-balling here but seriously interested in Any suggestions/Idea's.
Cheers
Post Reply

Return to “Virtualization”