[syscalls] Update mod exp compute cost #2807
Merged
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
|
The Firedancer team maintains a line-for-line reimplementation of the |
samkim-crypto
force-pushed
the
modexp-cu
branch
from
b8d7d68
to
6f267cb
Compare
Lichtso
reviewed
Aug 31, 2024
| .unwrap_or(u64::MAX), | ||
| .checked_div(2) | ||
| .unwrap_or(u64::MAX) | ||
| .saturating_add(budget.big_modular_exponentiation_cost), |
Lichtso
reviewed
Aug 31, 2024
| consume_compute_meter( | ||
| invoke_context, | ||
| budget.syscall_base_cost.saturating_add( | ||
| input_len | ||
| .saturating_mul(input_len) | ||
| .checked_div(budget.big_modular_exponentiation_cost) | ||
| .unwrap_or(u64::MAX), | ||
| .checked_div(2) |
0x0ece
approved these changes
Aug 31, 2024
Lichtso
reviewed
Sep 2, 2024
compute-budget/src/compute_budget.rs
Outdated
| @@ -180,7 +182,8 @@ impl ComputeBudget { | |||
| alt_bn128_multiplication_cost: 3_840, | |||
| alt_bn128_pairing_one_pair_cost_first: 36_364, | |||
| alt_bn128_pairing_one_pair_cost_other: 12_121, | |||
| big_modular_exponentiation_cost: 33, | |||
| big_modular_exponentiation_base_cost: 190, | |||
| big_modular_exponentiation_cost_multiplicative_factor: 2, | |||
There was a problem hiding this comment.
Isn't it a "divisor" or "denominator" instead of a multiplicative_factor?
There was a problem hiding this comment.
Yeah you are right. 1/2 is the multiplicative factor. I'll rename it to divisor.
Lichtso
approved these changes
Sep 2, 2024
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Outdated suggestions cannot be applied.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.
Problem
The big integer modular exponentiation (EIP-198) syscall was added in solana-labs#28503. However, the compute units were significantly under-assigned and there is currently a security advisory regarding this.
The syscalls are feature-gated (on the blocked list) and not activated on any of the networks. There are currently no projects that are actively waiting for these syscalls to land so these syscalls are not at the top priority especially given the current active development of firedancer. However, there are project that plans to use these syscalls in the future, so it would be best to address the compute-unit issue and close the security advisory even if the feature gate remain in the blocked list for some time.
Summary of Changes
I updated the compute cost.
The initial compute unit was assigned by benching the modular exponentiation and dividing the time by 33ns per CU. The initial bench not only failed to represent an average cluster performance in executing these modular exponentiation functions, but also failed to take into account certain edge cases in the modular exponentiation algorithm that performs significantly slower.
The modular exponentiation syscall implementation uses the rust
num_bigintcrate. In this crate, the modular exponentiation is implemented with the following logic:The initial bench of the syscalls only took into account the case when the modulus is odd even though the performance is much slower when the modulus is even due to the unsuitability of the montgomery arithmetic.
I re-ran the benches in my devserver for the cases when the modulus is even (power-of-two) and re-calculated the compute units (33ns per CU). I interpolated the CUs in relation to the input size (
N) and it seems that the functionN^2/2 + 190suitably approximates the CUs. This function does slightly over-assigns the CUs for small inputs (up to ~32 bytes), but given that most uses of these syscalls will be on much higher numbers, I think this is okay.Note: There is actually a better algorithm for computing modular exponentiation that takes advantage of the Chinese remainder theorem (https://www.people.vcu.edu/~jwang3/CMSC691/j34monex.pdf). I tested the Aurora implementation of this algorithm (https://github.com/aurora-is-near/aurora-engine/tree/develop/engine-modexp), but it seems that this specific implementation actually performs slower than the
num_bigintimplementation.Since this syscall is currently not at the top-priority list and will remain in the blocked list for some time, I think we can just bump-up the compute units to address and close the security advisory for now. In the future, we should revisit this to use the better algorithm to improve the syscall performance.
Fixes #