search

hpcaitech / FastFold

Optimizing AlphaFold Training and Inference on GPU Clusters
Apache License 2.0
556 stars 84 forks source link

Would it be possible to run this locally in computers/laptops with nvidia cuda drivers+docker? #173

Open juankaamez opened 1 year ago

juankaamez commented 1 year ago

Would it be possible to use this to run nodes of a cluster of computers which would be normal computers/laptops to create a beowulf-like infraestructure approach? How would in that case be the database to fold proteins be distributed, meaning, would it be needed to download the database in each node? Would it be enough to download it in a distributed-database-like form?

Gy-Lu commented 1 year ago

Hi, thanks for your issue. And the answer to your question is yes. It is possible to run it on normal computers/laptops with Nvidia GPU. However, if you want to distribute saving your database, it should be organized carefully and the preprocessing code should be modified. Now FastFold has not offered such a feature yet.

juankaamez commented 1 year ago

Thanks a lot for your answer Gy-Lu! I was wondering, if that were an option to be added (it might really be interesting to be able to switch from supercomputing or beowulf nodes creating a cluster), which distributed database model/engine would you suggest and how could we modify that preprocessing code? It would allow to have nodes collaborating from all over the world, run locally, contribute with other supercomputing nodes in a heterogenous platform. What do you think?

Gy-Lu commented 1 year ago

Hi, it sounds like a really interesting idea. However, I am not familiar with DB, and I can not give you a suggestion.