Conversation

Notices

1. Embed this notice
   Corey Quinn (quinnypig@awscommunity.social)'s status on Wednesday, 31-Jan-2024 08:45:35 JST Corey Quinn

   in reply to

   "Well what about Spot?"

   HPC workloads generally hate nodes going away mid-computation, plus there's no way to model the cost in advance given Spot's variable nature.

   For some of these runs it's hard enough to get on-demand capacity, let alone finding enough *excess*.

   • Haelwenn /элвэн/ :triskell: likes this.
   • Embed this notice
     Corey Quinn (quinnypig@awscommunity.social)'s status on Wednesday, 31-Jan-2024 08:45:36 JST Corey Quinn

     Okay, I've been dancing around it for years--it's time to call it:

     There's no economic story around using the cloud for steady-state HPC workloads. It's not even close to cost competitive.