← Back to team overview

mimblewimble team mailing list archive

Re: Elapsed-Time-Scaled Block Size

 

Responding to Philbert:

"To be on the safe side, it would probably need to be based not so much on
the time_from_last_block, but more on a median timestamp of the last N
blocks"

The problem with using the median in this situation is that the guarantee
of cumulative size limit being the same as would be in the
constant-block-size case goes out the window. Also, it's more complex to
reason about and figure out the limit in one's head. The big benefit of
using the elapsed time since last block is its simplicity.
Also, can you elaborate on "to be on the safe side"? How would this be
safer?


"If I understand you correctly, we're basically saying that if hash power
cut in half all of a sudden, then as soon as the protocol/nodes realized
it, the allowable blocksize would be doubled."

The way you put it doesn't sound right. Protocol/nodes do not look at hash
rate to realize it was cut in half and react to that. Block size limit is
just a linear formula of timestamp delta. A better way to look at it is
saying if block was not found for 20 minutes, then the valid block size
limit for the next block, at this time, is double the nominal size.
Also, note that block times vary by quite a bit due to the nature of
probability, so this idea applies not only when hash rate varies, it
applies at all times.


"The mechanism would need to be symmetric (more time_from_last_block -->
larger max_blocksize, and less time_from_last_block --> smaller max
blocksize)"

Yes. Linear scaling is symmetric around the nominal size in that sense. So
no issues here.


"then when they [large chunk of miners] returned the max blocksize would
shrink down to less than it was before they left."

Can you expand on what you mean here? If hash power returns to the same
level as before it left, block size limit would be the same on average.
This is because the proposed size function does not accumulate past the
last found block. For example... Hashing power drops 75%. Now remaining
miners get a block every 40 minutes on average, but still including the
same amount of transactions. Lets say after 4 hours (~6 blocks) hash power
returns. Block rate is back to one block per 10 minutes. Nothing is
changed, block size limit is back to nominal size on average.
This essentially decouples tx (and therefore network) throughput from pow
variations. Pow variations then affect only new coin creation rate (as
usual).
Maybe coin creation can also po pro-rated to block size, so that inflation
is stabilized and not gameable too. I leave that up to others to ponder.


"P.P.S. Can this mechanism be gamed by large miners who collude to "cycle"
large amounts of hash power offline, let the max blocksize grow, then come
back online to grab the larger block?"

The disincentive is two-fold here. First, as you mentioned, they're losing
out on block reward; second, they're risking someone else getting the
larger block with all those tx fees, so there's no guarantee they'd get
those bigger blocks consistently.


On Nov 21, 2017 2:30 PM, "Dustin Dettmer" <dustinpaystaxes@xxxxxxxxx> wrote:

> It is a cool idea.
>
> On Tue, Nov 21, 2017 at 12:23 PM Philbert Wallace <
> philbertwallace4@xxxxxxxxx> wrote:
>
>> Hi Tomas,
>> I agree with Dustin that we don't want to have to trust the miners at the
>> individual level (ie: choosing timestamps) too much, but part of what these
>> protocols do is create mechanisms where we can (hopefully) trust them in
>> the aggregate. That being said, it's an interesting thought that you had
>> and I've been thinking about something similar. There is a (I hope)
>> particularly compelling argument presented at the end that I think lends at
>> least some merit to the concept. So, let's work through it a little:
>>
>> To be on the safe side, it would probably need to be based not so much on
>> the time_from_last_block, but more on a median timestamp of the last N
>> blocks but N would need to be less than the number of blocks in a
>> difficulty adjustment period (so this whole thing would happen totally
>> within the normal difficulty cycle and not affect the difficulty calcs),
>> but we'll just call it time_from_last_block for now. If I understand you
>> correctly, we're basically saying that if hash power cut in half all of a
>> sudden, then as soon as the protocol/nodes realized it, the allowable
>> blocksize would be doubled. And if that same hashpower returned, then as
>> soon as the protocol realized it, the allowable max size would be halved.
>> The mechanism would need to be symmetric (more time_from_last_block -->
>> larger max_blocksize, and less time_from_last_block --> smaller max
>> blocksize) as otherwise you run the risk of growing the total size of the
>> chain in an unpredictable fashion.
>>
>> One of the positive side effects (and I'm sure there are some negative
>> ones too) is that whenever a large chunk of hashpower unexpectedly leaves,
>> this mechanism which basically targets a "constant maximum transactional
>> throughput rate" would disproportionately reward those miners that stick
>> around and slog through the glut of hashpower, and proportionally punish
>> all miners (including those that stuck around) when they return. As an
>> example: in the recent BTC / BCH situation, where hashpower left and BTC's
>> blocktimes became slower than normal, there was one pool, Slush's pool,
>> that seemed to stick around BTC regardless. The mechanism we're discussing
>> here would have allowed Slush to mine larger blocks (and thereby earn more
>> fees for itself) during the time that the hashpower was absent.
>>
>> Another interesting feature (bug?) of a mechanism like this is that it
>> would actually have punished all miners for temporarily defecting because,
>> since Slush was able to grow the allowable blocksize while they were gone,
>> then when they returned the max blocksize would shrink down to less than it
>> was before they left. Those miners returning to BTC would be causing the
>> time-between-blocks to be increasing, and therefore, under this proposed
>> mechanism, the max blocksize would be decreasing. Perhaps such a punitive
>> arrangement would have removed/lessened the incentive that any of those
>> miners had to temporarily leave in the first place? While that may be
>> desirable, its hard to know without further analysis whether this mechanism
>> would lessen the propensity for defection.
>>
>> Now for the reason I do like the mechanism, presuming it could be
>> implemented safely. If we follow this logic all the way down and imagine a
>> situation where there is catastrophic loss of hashing power (governments
>> shut down all known large-scale mining operations), I think it's pretty
>> much assumed that a PoW network would do a hard-fork difficulty adjustment
>> (or in the worst case change PoW algorithms all together) in an effort to
>> broaden the number of people/nodes that can carry to torch forward. Let's
>> assume the network would try a difficulty adjustment first and try to coax
>> the current non-mining nodes into becoming miners in order to "save the
>> chain." With a mechanism like we're discussing here in place, the desired
>> effect of the hard-fork difficulty adjustment (namely to open mining up to
>> a larger group of nodes) I think might happen more naturally and with less
>> downtime/chaos (and at-minimum without a hard fork). This is because the
>> more time that passes, the more the mempool would grow, and the large the
>> acceptable max block size. In the extreme case, where only 1% of the
>> original hashpower remained, then, sure, even though your node would still
>> need to meet the difficulty requirement, the temptation to build a block
>> that is 100x bigger than normal (and earn all those fees!) so as to keep
>> the network's max throughput on track at a constant could become large
>> enough that literally most (previously non-mining) nodes that didn't have a
>> shot when compared to the "big guys" would begin mining. Similarly the big
>> guys would be in a mad scramble to get back online and nab mining that
>> giant goldmine of a large block!
>>
>> Lastly, the big winner here would be (hopefully): users. This is because
>> while the miners contemplated in the above are playing games the honest
>> miners were able to hold down the fort and make sure that the max
>> transactional throughput rate stayed steady.
>>
>> While I'm not certain this is the outcome/mechanism we want for grin's
>> network (especially since grin is very focused on keeping things simple
>> right now), it at least is an interesting direction to consider.
>>
>> P.S. I'm also not certain that this mechanism doesn't pervert miner
>> incentives such that they actively try to reduce each other's network
>> connectivity. Then again that perversion exists in bitcoin as-is anyway,
>> and so far it seems to be ok.
>>
>> P.P.S. Can this mechanism be gamed by large miners who collude to "cycle"
>> large amounts of hash power offline, let the max blocksize grow, then come
>> back online to grab the larger block? Perhaps, but since they would only be
>> getting one block subsidy for the block (the larger reward would be due to
>> fees), it seems they would be better off by not doing so and instead just
>> staying online as much as possible (which is what we want in the first
>> place!).
>>
>>
>> On Tue, Nov 21, 2017 at 1:00 AM, Dustin Dettmer <
>> dustinpaystaxes@xxxxxxxxx> wrote:
>>
>>> I wouldn't put so much trust in miners timestamps. I'm much more of a
>>> fan of building systems where we don't have to trust them.
>>>
>>> On Tue, Nov 21, 2017 at 12:53 AM Tomas Juočepis <tomasjuocepis@xxxxxxxxx>
>>> wrote:
>>>
>>>> Hello, grinners,
>>>>
>>>> what if block size limit of each newly found block would be linearly
>>>> proportional to time elapsed since last block? Stated another way, nodes
>>>> would consider a new block valid only if timestamp delta (new block
>>>> timestamp - last block timestamp) multiplied by some parameter of size/time
>>>> ratio is greater than the size of the new block. It seems that something
>>>> like this could produce a more constant transaction throughput in cases of
>>>> quickly varying applied pow rate ("hash rate") without affecting difficulty
>>>> adjustment.
>>>>
>>>> For example, consider the following block timestamp deltas (in minutes):
>>>> 10, 20, 30, 3, 5, 7, 5, 5, 5, 10 (total 100, average 10 (nominal))
>>>> With block sizes being constant, we get 10 blocks worth of size.
>>>> With block sizes linearly proportional to time delta, with slope
>>>> coefficient set so that nominal (target) time produces nominal block size,
>>>> we'd get same cumulative size (and therefore same cumulative consumed
>>>> network bandwidth), but transaction rate would be more even/consistent.
>>>>
>>>> Let's say we can fit 1000 txs in a block designed to be mined every 10
>>>> minutes. With the previous example, with constant size we'd have the
>>>> following:
>>>> t=10, 1000 txs confirmed
>>>> t=30, 2000 txs confirmed
>>>> t=60, 3000 txs confirmed
>>>> t=63, 4000
>>>> t=68, 5000
>>>> t=75, 6000
>>>> t=80, 7000
>>>> t=85, 8000
>>>> t=90, 9000
>>>> t=100, 10000
>>>> With time-delta-scaled size, we'd have:
>>>> t=10, 1000 txs
>>>> t=30, 3000 txs
>>>> t=60, 6000 txs
>>>> t=63, 6300 txs
>>>> t=68, 6800 txs
>>>> t=75, 7500 txs
>>>> t=80, 8000 txs
>>>> t=90, 9000 txs
>>>> t=100, 10000 txs
>>>>
>>>> Any thoughts? The main issue I see is gaming timestamps, but can't that
>>>> be solved by nodes not propagating blocks until timestamps are no longer
>>>> dated in the future? Miners could still risk and add a timestamp few
>>>> minutes into the future, but they would risk their block being orphaned if
>>>> another miner finds another block with a valid timestamp, thus propagating
>>>> through the network (before the future dated block can propagate).
>>>> --
>>>> Mailing list: https://launchpad.net/~mimblewimble
>>>> Post to     : mimblewimble@xxxxxxxxxxxxxxxxxxx
>>>> Unsubscribe : https://launchpad.net/~mimblewimble
>>>> More help   : https://help.launchpad.net/ListHelp
>>>>
>>>
>>> --
>>> Mailing list: https://launchpad.net/~mimblewimble
>>> Post to     : mimblewimble@xxxxxxxxxxxxxxxxxxx
>>> Unsubscribe : https://launchpad.net/~mimblewimble
>>> More help   : https://help.launchpad.net/ListHelp
>>>
>>>
>>

Follow ups

References