Fairly price storage costs for transactions #2456

iamalwaysuncomfortable · 2024-05-16T20:27:21Z

Motivation

Transactions are currently limited to being 128kb or less in size.

Most execution transactions are small (for instance, transfer_public is 2880 bytes in size). However, transaction size can be greatly increased if the function being executed has very large inputs or very large outputs in the form of nested structs or arrays.

SnarkOS nodes must allocate enough memory to verify all transactions and carry out speculate operations on them. If many executions near the 128kb limit are sent to a network in a short timeframe, it can lead to a large amount of memory growth, long block times, long sync times and fast ledger growth.

Below is an example of memory growth on a validator observed from sending 3 TPS of 124kb transactions over 2 hours.

Currently a 128kb transaction would only cost .128 credits to execute, which makes it relatively cheap to flood the network with them. Given the deleterious effects of large transactions to network health, it makes sense to impose higher storage costs on larger transactions to ensure users of them are paying the appropriate costs for the effort it takes to process them.

This PR proposes that the storage cost of execution transactions above 5kb in size be penalized with a quadratic function.

$$\text{fee\_microcredits} = \begin{cases} \text{num\_bytes} & \text{if } 0 < \text{num\_bytes} \leq 5 \\\ \frac{{num\_bytes}^2}{1000} & \text{otherwise} & \text{if } \text{num\_bytes} > 5 \\\ \end{cases}$$

An exact function to price storage remains open for debate, and alternatives can certainly be proposed, but it is strongly encouraged storage costs be increased to properly pay for the extra effort the network must perform verify and store them.

NOTE: This would be a breaking change for Testnet, so this should be a pre-mainnet change and integrated into mainnet-staging.

synthesizer/process/src/cost.rs

raychu86 · 2024-05-24T16:28:43Z

When a decision is made, can you post the before and after fees amounts for the credits.aleo functions?

iamalwaysuncomfortable · 2024-05-25T02:23:19Z

When a decision is made, can you post the before and after fees amounts for the credits.aleo functions?

The decision on the function has been made and posted above on the PR description:

The costs of major credits.aleo functions are as follows:

bond_public: 306229 (storage: 1519, finalize: 304710)
join: 2087 (storage: 2087, finalize: 0)
split: 2131 (storage: 2131, finalize: 0)
transfer_private: 2242 (storage: 2242, finalize: 0)
transfer_private_to_public: 26961 (storage: 2141, finalize: 24820)
transfer_public: 51060 (storage: 1420, finalize: 49640)
transfer_public_to_private: 26439 (storage: 1619, finalize: 24820)
unbond_public: 325149 (storage: 1309, finalize: 323840)

There haven't been any changes in storage costs in credits.aleo functions under this pricing model.

console/network/src/lib.rs

Co-authored-by: vicsn <victor.s.nicolaas@protonmail.com>

vicsn

Await merge until Howard approved and CI passes

howardwu · 2024-06-04T17:19:43Z

synthesizer/process/src/cost.rs

+    // Compute a quadratic storage cost penalty if above the size penalty threshold.
+    if storage_cost > N::EXECUTION_STORAGE_PENALTY_THRESHOLD {
+        storage_cost = storage_cost
+            .saturating_pow(N::EXECUTION_STORAGE_FEE_DEGREE)


I strongly advise we use a quadratic function otherwise use an approximation algorithm for pow, similar to how to do retarget in the puzzle.

The concern here is determinism across architectures and across programming languages.

May I ask how this would be non-determinism here? Both storage_cost and the fee degree are integers which power in rust is deterministic for.

And also, what is meant by other languages? WASM compatibility?

And also, what is meant by other languages? WASM compatibility?

WASM is certainly a good example. Wasm does have 64-bit types but those are represented in 32-bit memory addresses so "saturation" operations might not translate here.

howardwu · 2024-06-04T17:21:07Z

synthesizer/process/src/cost.rs

+    let mut storage_cost = execution.size_in_bytes()?;
+
+    // Compute a quadratic storage cost penalty if above the size penalty threshold.
+    if storage_cost > N::EXECUTION_STORAGE_PENALTY_THRESHOLD {


I would strongly advise introducing unit tests that explicitly clamp the expected upper bound and lower bound numeric value behaviors prior to the approval and merge of this PR.

These should be able to added before the end of this week.

Meshiest · 2024-06-05T22:15:54Z

synthesizer/process/src/cost.rs

+    // Compute a quadratic storage cost penalty if above the size penalty threshold.
+    if storage_cost > N::EXECUTION_STORAGE_PENALTY_THRESHOLD {
+        storage_cost = storage_cost
+            .saturating_pow(N::EXECUTION_STORAGE_FEE_DEGREE)


Is there a reason why the penalty is applied to the entire storage_cost rather than just overage?

A jump in fee by 20,010 for 1 additional byte seems a little concerning

$(storageCost ^{degree})/scalingFactor$ new storage_cost

$(5001 ^2)/1000$ 25,010

$(5100 ^2)/1000$ 26,010

$(6000 ^2)/1000$ 36,000

$(32,000 ^2)/1000$ 1,024,000

$(64,000 ^2)/1000$ 4,096,000

$(128,000 ^2)/1000$ 16,384,000

Is there a reason why the penalty is applied to the entire storage_cost rather than just overage?

Namely because as transaction size grows, block size does as well. Large blocks with many large transactions require each node to fetch these transmissions to complete their view of the dag which causes stress on the BFT and syncing processes.

While perhaps not the most mathematically beautiful, it prioritizes network health by ensuring penalties are applied to larger transactions starting early. It's likely that most functions on the network will have transaction sizes under 5kb, which is the reasoning for the initially linear pricing.

Pricing information like the table you crafted above can be added to network documentation ensure this is explained.

Add an extra cost penalty for executions above 15kb

5bbdee2

iamalwaysuncomfortable changed the base branch from testnet-beta to mainnet-staging May 16, 2024 20:31

vicsn reviewed May 17, 2024

View reviewed changes

synthesizer/process/src/cost.rs Outdated Show resolved Hide resolved

Move price constants into the network trait

89e0614

iamalwaysuncomfortable added 2 commits May 24, 2024 16:52

Update storage fee equation to be quadratic after the penalty threshold

2ad22c0

Merge branch 'mainnet-staging' into storage-price-penalty

56501d8

Update test expectations that reflect new fees

14d033c

iamalwaysuncomfortable marked this pull request as ready for review May 25, 2024 02:26

vicsn reviewed May 25, 2024

View reviewed changes

console/network/src/lib.rs Outdated Show resolved Hide resolved

Clarify units in Network trait comments

ed9872f

Co-authored-by: vicsn <victor.s.nicolaas@protonmail.com>

vicsn approved these changes May 30, 2024

View reviewed changes

iamalwaysuncomfortable added the requires-ledger-reset Merging this PR will require a ledger reset label May 31, 2024

howardwu reviewed Jun 4, 2024

View reviewed changes

Meshiest reviewed Jun 5, 2024

View reviewed changes

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

Fairly price storage costs for transactions #2456

Fairly price storage costs for transactions #2456

iamalwaysuncomfortable commented May 16, 2024 •

edited

raychu86 commented May 24, 2024

iamalwaysuncomfortable commented May 25, 2024

vicsn left a comment

howardwu Jun 4, 2024 •

edited

gluax Jun 5, 2024 •

edited

iamalwaysuncomfortable Jun 6, 2024

howardwu Jun 4, 2024

iamalwaysuncomfortable Jun 6, 2024

Meshiest Jun 5, 2024

iamalwaysuncomfortable Jun 6, 2024

$(storageCost ^{degree})/scalingFactor$	new storage_cost
$(5001 ^2)/1000$	25,010
$(5100 ^2)/1000$	26,010
$(6000 ^2)/1000$	36,000
$(32,000 ^2)/1000$	1,024,000
$(64,000 ^2)/1000$	4,096,000
$(128,000 ^2)/1000$	16,384,000

Fairly price storage costs for transactions #2456

Are you sure you want to change the base?

Fairly price storage costs for transactions #2456

Conversation

iamalwaysuncomfortable commented May 16, 2024 • edited

Motivation

raychu86 commented May 24, 2024

iamalwaysuncomfortable commented May 25, 2024

vicsn left a comment

Choose a reason for hiding this comment

howardwu Jun 4, 2024 • edited

Choose a reason for hiding this comment

gluax Jun 5, 2024 • edited

Choose a reason for hiding this comment

iamalwaysuncomfortable Jun 6, 2024

Choose a reason for hiding this comment

howardwu Jun 4, 2024

Choose a reason for hiding this comment

iamalwaysuncomfortable Jun 6, 2024

Choose a reason for hiding this comment

Meshiest Jun 5, 2024

Choose a reason for hiding this comment

iamalwaysuncomfortable Jun 6, 2024

Choose a reason for hiding this comment

iamalwaysuncomfortable commented May 16, 2024 •

edited

howardwu Jun 4, 2024 •

edited

gluax Jun 5, 2024 •

edited