Fastest retrieval | Minimum computational cost | Security model | Assumptions | Confidence to deploy | |
NSE | could be <10s | 10x cost of storage of polling time | cost | Cost of SHA, DRG | high |
Holy Grail | can’t be <30s | cheap | latency | VDF | medium-low |
ChainPoSt | can be <1s | cheap | latency | VDF | medium-low |
OngoingPoSt | can be <1s | cheap | latency | VDF | low |
Extra copy | <1s | 0 | any | extra storage requirement | high |
Cost model - NSE
NSE: Parallelized proof of space
Pro
1min retrieval. Potentially down to O(1s)
Can do it today
Con
cost model (versus latency model) - limits us to linear checks
expensive to retrieve (cost equivalent to 1 day of storage)
One 32gb retrieval would cost $0.00715
so very parallelizable, but you need it to be “a lot” of computation, to force this cost.
Potential future direction
reduce polling time (so if you challenge every hour) then your cost of unsealing can be reduced as well
Latency model - naive (aka holy grail)
Ratio between honest and malicious is currently 50. want to bring it down to 1-2.
include this proof as part of winningPoST
this would use a VDF to get to cheaper probability
Pros
Cheap
1min retrieval
cons
can’t go below 1 min
don’t know how to do it
Other (less plausible) options below:
Execution Extension - chain post
interleaving VDF with a proof of space
pro
potential sub second retrieval
con
have to deal VDFs
depending on potential adversarial speed-up of VDFs this could be hard
proof construction is non-trivial
Execution extension - ongoing post
always generate proofs. at time of winning do a summary over the ones so far
(always doing vdf work. occasionally need to proof of space)
pro
winning post is simple.
con
have to deal VDFs
(same cons as chain post)
Better parameters for SDR
pro
very easy to implement
could be more secure
con
not going to get us that far. best speed ~10min-30min
Proof of extra copy
include convolution of the unencrypted copy
pro
no network upgrade
no unsealing cost
con
more storage
hard to make mandatory
Asymmetric porep
if the input to your sector is public data rather than another replica then you can do this.
pro
fast retrieval (potentially sub-second)
con
only works for public data