avoid unnecessary async copies in broadcast #3830
Conversation
| try: snappy.encode(uncompressed) | ||
| except InputTooLarge: | ||
| raiseAssert "More than 4gb? not likely.." | ||
| proc broadcast(node: Eth2Node, topic: string, msg: seq[byte]): |
There was a problem hiding this comment.
Is the claim that this PR is reducing the number of copies? There will be still some copying here (both topic and msg will be copied)
There was a problem hiding this comment.
it avoids the (extremely slow) copy of msg via genericAssign as well as the lifetime extension of keeping copies of the unserialized object and uncompessed in the closure environment
There was a problem hiding this comment.
We can perhaps start optimizing away even the string and seq copies like the ones here by using more ref parameters.
There was a problem hiding this comment.
not worth it imo, this is by far not the biggest culprit (most copying happens in libp2p which would require major changes) - this was just a small drive-by fix that doesn't change semantics or introduce extra mutability like ref would - it gets rid of the extra try/catch layers that also were introducing overhead and noise (no longer needed thanks to the new snappy impl)
No description provided.