Just a thought on this -- I'm not saying this is a good idea or a bad idea, because I have spent about zero time thinking about it, but something did come to mind as I read this. Reading 20 GB of data for every hash might be a bit excessive. And as the blockchain grows, it will become infeasible to continue. However, what comes to mind is the ROMix algorithm defined by Colin Percival, which was the pre-cursor to scrypt. Which is actually what Armory uses for key stretching because it's far simpler than scrypt itself while maintaining the memory-hard properties (the downside is that it's much less flexible in allowing the user to trade-off compute time vs memory usage).
ROMix works by taking N sequential hashes and storing the results into a single N*32 byte lookup table. So if N is 1,000,000, you are going to compute 1,000,000 and store the results into 32,000,000 sequential bytes of RAM. Then you are going to do 1,000,000 lookup operations on that table, using the hash of the previous lookup result, to determine the location of next lookup (within that 32,000,000 bytes). Assuming a strong hash function, this means its impossible to know in advance what needs to be available in RAM to lookup, and it's easiest if you simply hold all 32,000,000 bytes in RAM. Something similar could be applied to your idea. We use the hash of a prevBlockHash||nonce as the starting point for 1,000,000 lookup operations. The output of the previous lookup is used to determine which block and tx (perhaps which chunk of 32 bytes within that tx) is used for the next lookup operation. This means that in order to do the hashing, you need the entire blockchain available to you, even though you'll only be using a small fraction of it for each "hash". This might achieve what you're describing without actually requiring the full 20 GB of reading on ever hash. -Alan On 07/04/2014 06:27 AM, Andy Parkins wrote: > Hello, > > I had a thought after reading Mike Hearn's blog about it being impossible to > have an ASIC-proof proof of work algorithm. > > Perhaps I'm being dim, but I thought I'd mention my thought anyway. > > It strikes me that he's right that it's impossible for any algorithm to exist > that can't be implemented in an ASIC. However, that's only because it's > trying to pick an algorithm that is CPU bound. You could protect against > ASCI > mining (or rather, make it irrelevant that it was being used) by making the > algorithm IO-bound rather than CPU-bound. > > For example, what if the proof-of-work hash for a block were no longer just > "hash of block", which contains the hash of the parent block, but instead > were > hash of > > [NEW_BLOCK] [ALL_PREVIOUS_BLOCKS] [NEW_BLOCK] > > [ALL_PREVIOUS_BLOCKS] is now 20GB (from memory) and growing. By prefixing > and > suffixing the new block, you have to feed every byte of the blockchain > through > the hashing engine (the prefix prevents you caching the intermediate result). > > Whatever bus you're using to feed your high speed hashing engine, it will > always be faster than the bus -- hence you're now IO-bound, not CPU-bound, > and > any hashing engine will, effectively, be the same. > > I'm making the assumption that SHA-256 is not cacheable from the middle > outwards, so the whole block-chain _has_ to be transferred for every hash. > > Apologies in advance if this is a stupid idea. > > > > Andy ------------------------------------------------------------------------------ Open source business process management suite built on Java and Eclipse Turn processes into business applications with Bonita BPM Community Edition Quickly connect people, data, and systems into organized workflows Winner of BOSSIE, CODIE, OW2 and Gartner awards http://p.sf.net/sfu/Bonitasoft _______________________________________________ Bitcoin-development mailing list Bitcoin-development@lists.sourceforge.net https://lists.sourceforge.net/lists/listinfo/bitcoin-development
