Solution to error "mined block in the past"

[phoenixzhua]

This PR gives a solution to error "mined block in the past" due to windowpost and winpost happening at the same time.

• Add a flag to identify whether it is winning post or not when doing fft and multiexp calculation.

• Add an env. variable: WINPOST_TMPDIR as a path to hold the gpu_lock_file and priority_lock_file when doing winning post.

  Make sure to set the env. variable: WINPOST_TMPDIR, and not to set the env. variable: CUDA_VISIBLE_DEVICES, when running lotus-miner.

It is suggested to put two GPUs on the miner box. It would select the first GPU for winning post and the second GPU for window post.

related discussion: https://github.com/filecoin-project/lotus/discussions/7797

There is another related PR for this solution: https://github.com/filecoin-project/rust-fil-proofs/pull/1549

[vmx]

Sorry for the delay in responding. Thanks for your PR, it makes the desire for more advanced scheduling very clear. I wouldn't expect this to be merged soon, but it is a good starting point for further discussions.

I'm currently in the process of refactoring FFT and Multiexp into a general purpose library, so adding parameters directly to those calls, might not work anymore. Though I'll keep that use case in mind and perhaps there is a way of doing it.

Though perhaps we might want to spend time on a more general purpose solution.

[vmx]

@phoenixzhua I've discussed this issue with the team and we tried to find out which underlying issue this tries to solve. Can you please have a look at the new issue I opened (https://github.com/filecoin-project/bellperson/issues/251) and see if such a feature would meet your requirements.