@gcolvin no the v8 benchmarks aren't using native bigint host functions. they're using wasm for the bigint math. v8 with host function bigints should be faster than the interpreter, but unfortunately v8 has a terrible host function overhead so it's a slowdown right now. if the host function overhead for v8 gets fixed/optimized, then it should be faster
the wasm code for the bigints are the hand-optimized wasm that's in websnark. i'll just drop the charts in here (pending a writeup to be done later).
here's the code: https://github.com/cdetrio/scout.ts/tree/bn128-pairing
Hi, I was trying out the project (https://medium.com/wasm/compile-and-deploy-an-erc20-contract-on-ewasm-d91d890665e6). I failed to connect to ewasm test network from metamask. It tries to connect for some time and shows that it cannot be connected.
Can someone help me point out the error also please let me know whether the public testnet URL http://ewasm.ethereum.org:8545 is up and running. Thanks in advance.
Hey, have few questions on Gas metering. (referred https://ewasm.readthedocs.io/en/mkdocs/determining_wasm_gas_costs/)
- Cycles are calculated based on bare metal CPU cycle or it is calculated based on wasm virtual machine cycle?
- How the cycles are calculated ?
@mishaaaa you can try running a local geth node, using this: https://github.com/jwasinger/ewasm-dev-env , after that you can connect Metamask to the local node:
http://localhost:8545.
right now the gas cost table is a best guess. It needs to be evaluated by benchmarking, to see how accurately those gas costs can predict runtime in various engines. that's one of our current projects
@mishaaaa you can import the premined account in Metamask using this pk: https://github.com/jwasinger/ewasm-dev-env/blob/master/geth/keys/prefunded/prefunded-priv.txt
Instruction tables by Agner Fog: http://www.agner.org/optimize/instruction_tables.pdf
In the eWASM readme books it is mentioned that
Assumption 2: 1 second of CPU execution equals to 10 million gas (i.e. 1 gas equals to 0.1 us).
- On what basis this assumptions are made?
- And how the cycles are mentioned? Checked in book http://www.agner.org/optimize/instruction_tables.pdf in Haswell section couldn't able see those cycles for those operation
In Haswell table for DIV r32- it is mentioned the latency range 22-29
and for DIV r64- it is mentioned the latency range 32-96
But in gas-cost table for both i32 and i64 type it is mentioned 80 cycle
I have a small bench to measure it here, should be easily ported to C/C++: https://github.com/status-im/nim-beacon-chain/blob/nbench/nbench/platforms/x86.nim#L100-L122
just make sure to use volatile load/store
@AchalaSB the 80 came from the haswell specs and not third party sources (it should be linked in the markdown)
oops I meant this one
I wish we had some notion of how far off we can be and not get DoSed. If the code is interpreted these numbers will get drowned to some extent by the interpreter overhead, and if it is compiled the compiler itself and the tricks CPUs play will make cycle-counting only a guess. And the counting itself can be a significant overhead that will depend on how smart the injection is and how well the compiler can optimize it.
@jakelang you might know the reference book for CPU cycle, as you have commited for eWASM gas cost table.
https://github.com/ewasm/design/blob/mkdocs/docs/determining_wasm_gas_costs.md
https://github.com/ewasm/design/blob/mkdocs/docs/determining_wasm_gas_costs.md
This is the original source of the document: https://gist.github.com/axic/5bf506728005864461454dd4d37a7f37
And it refers to “Appendix C” of https://software.intel.com/sites/default/files/managed/9e/bc/64-ia-32-architectures-optimization-manual.pdf of which @cetrio copy pasted.
Anyway, I wrote it over 3 years so don’t remember where the 80 came from.
And search for wrc20 for more examples