Large batch sizes with SupraSeal #320
Draft
Conversation
This file contains bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
This is the first step of make more code share between the native and the SupraSeal implementation possible.
Refactor things similar to the SupraSeal code.
SupraSeal is so efficient as it leverages the fact that the Filecoin circuits have a specific shape. This can also be used for a more memory efficient representation during synthesis. This makes the synthesis less memory intensive and thus makes it possible to synthesize more circuits in parallel. The prover still needs the normal representation, hence is still memory intensive. We run a batch size of 5 instead of 10, which even reduces that memory footprint a lot which justified the additional computation costs. Those changes enables proving a large number of proofs pretty efficiently, which makes it suitable for Filecoins non-interactive PoRep. All those optimization are only for the SupraSeal code path as they are kind of Filecoin specific.
vmx
force-pushed
the
pipelining-lessmem
branch
from
772ce85 to
69b83f6
Compare
|
Sorry I accidentally pushed a change, but I reverted back to what it was. |
|
I've converted it to draft, as it needs more work. The CPU usage is not ideal, sometimes the Rayon thread pool is used by the synthesis, which then blocks the representation conversion. Ideally the Synthesis runs whenever the GPU is used. |
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Add this suggestion to a batch that can be applied as a single commit.
This suggestion is invalid because no changes were made to the code.
Suggestions cannot be applied while the pull request is closed.
Suggestions cannot be applied while viewing a subset of changes.
Only one suggestion per line can be applied in a batch.
Add this suggestion to a batch that can be applied as a single commit.
Applying suggestions on deleted lines is not supported.
You must change the existing code in this line in order to create a valid suggestion.
Outdated suggestions cannot be applied.
This suggestion has been applied or marked resolved.
Suggestions cannot be applied from pending reviews.
Suggestions cannot be applied on multi-line comments.
Suggestions cannot be applied while the pull request is queued to merge.
Suggestion cannot be applied right now. Please check back later.
When using proof batch sizes > 10, things will take a lot of memory. There's now a memory optimized code path which is only enabled for SupraSeal as it takes advantage of the shape of the Filecoin circuits. It may not be as efficient with other circuits.
This PR doesn't change the current PoRep behaviour, it is a different code path only if the proofs size is >10, which is the case for e.g. non-interactive PoRep. With these changes, non-interactive PoRep, which has about 12.6x times the proofs to do compared to the interactive PoRep, runs only about 8x slower, while not taking more memory.