The rest of this post discusses (2), the effects on a single node. The purpose of this analysis is to understand how calldata was priced at the network launch.
Update on Local Tests (Standalone Node) and Gas Cost Comparison
The main goal of performing local tests is to understand how the decrease in calldata gas cost would impact the system, assuming the network has zero delay (the effect on network delays is investigated separately). Recall that each node must validate the header of each new block and then, process the block. We want to ensure that the implementation of this EIP does not harm the network: for example, that the mere increase in blocksize does not cause significant problems in computing block headers and increase processing time, even if we assume zero network delays.
Recall that the current cost for calldata is 68 gas for each non-zero byte and 4 gas for each zero byte.
An increase in headers validation would impact security whereas an increase in block processing, all other things being equal, would impact throughput.
A quick gas comparison is interesting here:
When a validator receives a block, it computes the Keccak hash of the whole block to verify that the header used for the PoW is correct.
For a calldata input of size N bytes, if this verification operations were done in the EVM, the gas cost would be 30 gas for the first invocation of keccak and 6 gas for each additional word (32 bytes), totaling 30+6 * N / 32 gas. Asymptotically (neglecting the first 30 gas) this gives a gas price of roughly 6 for every 32 bytes, or 0.2 gas per byte. So, under current pricing, transmission and history for nonzero bytes accounts for all 68 gas.
One could argue that Keccak might be underpriced too. This is debunked by Holiman vm analysis, based on a historical analysis of opcode runtime over Ethereum lifetime. According to this work, the Keccak opcode’s gas price is on par with its processing time.
Our local measurements on our machine go along with this analysis.
So far, we compared calldata to computation cost. One could also try to compare calldata and storage. We tried but failed to find similar analogies there. This goes along with several other reports on SLOAD and SSTORE, stating that their pricing is complex and must be understood as a design choice for the network future 1
To conclude, assuming zero network delay, the fair and consistent pricing of calldata has a lower bound at 0.2 gas per byte. Practically, this means that the new gas cost recommendation should be based on its effects on network delay and history. This is also what was done in the original system design.