Hi Poinsot,
Started to review a bit the code branch on inquisition, and while doing so I was specifically
thinking about the proposed 2500 sigops limit, and how it weights on a multi-dimensional matrix
of a full-node performace (e.g fees, CPU time, disk space, etc).
Currently, in a simple model, a DoS adversary could constitute a 1-MB (it's pre-segwit acoutning)
transaction with 80_000 sigops from a 1-sat UTXO. A full-node to validate that would have to SHA256
the 1MB tx 80_000 times, thus the O(n^2) "bad" complexity.
Assuming the novel per-tx 2500 sigops limit, still a simple DoS adversary could constitute 32 * 32_150
virtual bytes tx (it's still a 1 MB block) spending from _32_ 1-sat UTXOs. A full-node to validate that
would have to fetch the 32 UTXOs.
This is the 1 UTXO from 32 UTXOs trade-off, I would like to draw awareness on, as fair the O(n^2)
complexity is "bad" but quid of the UTXO memory fetches if there are not in your high-hierarchy cache
and they have to be fetched from RAM, or even worst from disk (i7 core have RAM bigger than the current
UTXO set, not necessarily all range of RasPi).
From the viewpoint of a defending full-node, sure I can limit the number of per-tx sigops, but if an
adversary can achieve the same DoS efficiency now that more than UTXOs have to be fetched to validate
the same per-block total number of sigops, it's a legitimate wonder about the efficiency limit, and
more interestingly if there wouldn't be a better value to be selected.
So it's a bit my interrogation about this 2500 proposal, if worst-case transactions samples binding
to the 2500 limit have been crafted to maximize the number of UTXOs fetches. One can make the hypothesis
that UTXO fetches are "free", but I don't think it's necessarily true, while on the other hand modern ISAs
have dedicated hashing instructions.
Current BIP54 is a bit silent if full-node performance metrics like CPU cycles, IO disk operations or
bandwidth consumptions have been weighted in to select the proposed 2500 value of the limit. This would
be a fair point to modify BIP54 to say that the new sigops limit is only aimed to mitigate CPU DoS
and that others dimensions like memory management have not been emperically observed to be downgraded.
Best,
Antoine
OTS hash: 975674252060994d92eecd63a924e7530623ee737e33c5646d382f0f8c04ec74
Hi everyone,
I'd like to give an update on my Consensus Cleanup work, now BIP54.
I opened an implementation against Bitcoin Inquisition v29.1 at [0]. It contains extensive testing
of each of the four proposed mitigations, and was used as a basis to generate test vectors for
BIP54. I opened a PR against the BIPs repository to add them to BIP54 [1].
The test vectors for the transaction-level sigops limit contain a wide variety of usage combinations
as well as ways of running into the limit. They also include some historical violations as well as
pathological transactions demonstrating the implementation details of the sigop accounting logic
(which was itself borrowed from that of BIP16, which all Bitcoin implementations presumably already
have).
The test vectors for the new witness-stripped transaction size restriction similarly exercise the
bounds of the check under various conditions (e.g. transactions with/without a witness). All
historical violations were also added to the test vectors, thanks to Chris Stewart for digging those
up.
Because the new timestamp restrictions are tailor-made to the mainnet difficulty adjustment
parameters, the test vectors for those contain a number of chains of mainnet headers (from genesis).
Each test case contains a full header chain and whether it is valid according to BIP54. These chains
were generated using a custom miner available in [2] and added to the implementation as a JSON data
file.
The test vectors for the coinbase restriction similarly include a chain of mainnet blocks, because
the timelock check is context-dependent. These were generated using a similar miner also available
at [2].
I'm seeking feedback on these test vectors from everybody but in particular developers of
alternative Bitcoin clients, as compatibility with other Bitcoin implementations than Bitcoin Core
was a design goal.
Best,
Antoine Poinsot
[0]: https://github.com/bitcoin-inquisition/bitcoin/pull/99
[1]: https://github.com/bitcoin/bips/pull/2015
[2]: https://github.com/darosior/bitcoin/commits/bip54_miner