Skip to content

Latest commit

 

History

History
135 lines (86 loc) · 6.27 KB

eip-1884.md

File metadata and controls

135 lines (86 loc) · 6.27 KB
Raw
eip title author type category discussions-to status created
1884
Repricing for trie-size-dependent opcodes
Martin Holst Swende (@holiman)
Standards Track
Core
Draft
2019-03-28

Simple Summary

This EIP proposes repricing certain opcodes, to obtain a good balance between gas expenditure and resource consumption.

Abstract

The growth of the Ethereum state has caused certain opcodes to be more resource-intensive at this point than they were previously. This EIP proposes to raise the gasCost for those opcodes.

Motivation

An imbalance between the price of an operation and the resource consumption (CPU time, memory etc) has several drawbacks:

  • It could be used for attacks, by filling blocks with underpriced operations which causes excessive block processing time.
  • Underpriced opcodes cause a skewed block gas limit, where sometimes blocks finish quickly but other blocks with similar gas use finish slowly.

If operations are well-balanced, we can maximise the block gaslimit and have a more stable processing time.

Specification

At block N,

  • The SLOAD operation changes from 200 to 800 gas,
  • The BALANCE operation changes from 400 to 700 gas,
  • A new opcode, SELFBALANCE is introduced at 0x46.
    • SELFBALANCE pops 0 arguments off the stack,
    • SELFBALANCE pushes the balance of the current address to the stack,
    • SELFBALANCE is priced as GasFastStep, at 5 gas.

Rationale

Here are two charts, taken from a full sync using Geth. The execution time was measured for every opcode, and aggregated for 10K blocks. These bar charts show the top 25 'heavy' opcodes in the ranges 5M to 6M and 6M to 7M:

bars1 bars2

Note: It can also be seen that the SLOAD moves towards the top position. The GASPRICE opcode has position one which I believe can be optimized away within the client -- which is not the case with SLOAD/BALANCE.

Here is another chart, showing a full sync with Geth. It represents the blocks 0 to 5.7M, and highlights what the block processing time is spent on.

geth

It can be seen that storage_reads and account_reads are the two most significant factors contributing to the block processing time.

SLOAD

SLOAD was repriced at EIP-150, from 50 to 200. The following graph shows a go-ethereum full sync, where each data point represents 10K blocks. During those 10K blocks, the execution time for the opcode was aggregated.

graph

It can be seen that the repricing at EIP-150 caused a steep drop, from around 67 to 23. Around block 5M, it started reaching pre-EIP-150 levels, and at block 7M it was averaging on around 150 - more than double pre-eip-150 levels.

Increasing the cost of SLOAD by 4 would bring it back down to around 40. It is to be expected that it will rise again in the future, and may need future repricing, unless state clearing efforts are implemented before that happens.

BALANCE

BALANCE (a.k.a EXTBALANCE) is an operation which fetches data from the state trie. It was repriced at EIP-150 from 20 to 400.

graph

It is comparable to EXTCODESIZE and EXTCODEHASH, which are priced at 700 already.

It has a built-in high variance, since it is often used for checking the balance of this, which is a inherently cheap operation, however, it can be used to lookup the balance of arbitrary account which often require trie (disk) access.

In hindsight, it might have been a better choice to have two opcodes: EXTBALANCE(address) and SELFBALANCE, and have two different prices.

  • This EIP proposes to extend the current opcode set.
    • Unfortunately, the opcode span 0x3X is already full, hence the suggestion to place SELFBALANCE in the 0x4X range.
    • As for why it is priced at 5 (GasFastStep) instead of 2 (GasQuickStep), like other similar operations: the EVM execution engine still needs a lookup into the (cached) trie, and balance, unlike gasPrice or timeStamp, is not constant during the execution, so it has a bit more inherent overhead.

Backwards Compatibility

The changes require a hardfork. The changes have the following consequences:

  • Certain calls will become more expensive.
  • Default-functions which access the storage and may in some cases require more than2300 gas (the minimum gas that is always available in calls).
  • Contracts that assume a certain fixed gas cost for calls (or internal sections) may cease to function.
    • However, these operations have already been repriced earlier, so there is a historical precedent that 'the gascost for these operations may change', which should have prevented such fixed-gas-cost assumptions from being implemented.

I expect that certain patterns will be less used, for example the use of multiple modifiers which SLOADs the same opcode will be merged into one. It may also lead to less log operations containing SLOADed values that are not strictly necessary.

Test Cases

No test cases are implemented as of yet.

Implementation

This EIP has not yet been implemented in any client. Both these opcodes have been repriced before, and the client internals for managing reprices are already in place.

SELFBALANCE

This is the implementation for the new opcode in go-ethereum:

func opSelfBalance(pc *uint64, interpreter *EVMInterpreter, contract *Contract, memory *Memory, stack *Stack) ([]byte, error) {
	stack.push(interpreter.intPool.get().Set(interpreter.evm.StateDB.GetBalance(contract.Address())
	return nil, nil
}

Security considerations

  • See backwards compatibility section.
  • There are no special edgecases regarding SELFBALANCE, if we define it as BALANCE with address instead of popping an address from the stack -- since BALANCE is already well-defined.
  • It should be investigated if Solidity contains any hardcoded expectations on the gas cost of these operations.

Copyright

Copyright and related rights waived via CC0.