Sky currently supports users requesting X amount of GPUs/TPUs, e.g.
resources:
accelerators:
V100: 8
Users should also be able to specify how many CPU cores they want, e.g.
resources:
cores: 16
accelerators:
V100: 8
Additional changes are needed to the optimizer (to consider this additional resource constraint). In additional, while Google decouples # of CPUs and # of GPUs, AWS advertises their GPUs in fixed packages of (CPUs, GPUs) (e.g. p3.2x (16 CPU, 1 V100), 3.8x, ...), which is not as flexible as Google's model. The optimizer needs to account for this and select the right packages in AWS that meet the user's constraint.
Sky currently supports users requesting X amount of GPUs/TPUs, e.g.
Users should also be able to specify how many CPU cores they want, e.g.
Additional changes are needed to the optimizer (to consider this additional resource constraint). In additional, while Google decouples # of CPUs and # of GPUs, AWS advertises their GPUs in fixed packages of (CPUs, GPUs) (e.g.
p3.2x (16 CPU, 1 V100), 3.8x, ...), which is not as flexible as Google's model. The optimizer needs to account for this and select the right packages in AWS that meet the user's constraint.