Network propagation speeds
9 messages, 24 to 27 Nov 2013. Started by Christian Decker.
Christian Decker (4), Mike Hearn (2), Greg Maxwell, Peter Todd, Michael Gronager
Since this came up again during the discussion of the Cornell paper I thought I'd dig up my measurement code from the Information Propagation paper and automate it as much as possible.
The result is the Network Propagation page on bitcoinstats.com (http://bitcoinstats.com/network/propagation/). It takes a daily snapshot of the situation, then calculates the time until blocks and transactions reach a certain percentile of the nodes in the network. There is also a detailed page showing the density function describing at what times nodes learn about the existence of a block/transaction (for example yesterdays distribution: http://bitcoinstats.com/network/propagation/2013/11/23).
I intend to add more information and plots over time, but I wanted to push this out quickly as there were some people asking for it. Hope this helps getting the blockchain fork rate down :-)
Regards, Chris
On Sun, Nov 24, 2013 at 8:20 AM, Christian Decker <decker.christian@gmail•com> wrote:
Quoted text (3 lines)
> Since this came up again during the discussion of the Cornell paper I > thought I'd dig up my measurement code from the Information > Propagation paper and automate it as much as possible.
Could you publish the block ids and timestamp sets for each block?
It would be useful in correlating propagation information against block characteristics.
Sure thing, I'm looking for a good way to publish these measurements, but I haven't found a good option yet. They are rather large in size, so I'd rather not serve them along with the website as it hasn't got the capacity. Any suggestions? If the demand is not huge I could provide them on a per user basis.
This is great, thanks for doing it. Tip sent your way.
Graphs of how propagation data change over time would also be helpful (as well as raw data so we can calculate overhead per kilobyte and so on). I know there are only two days worth of data, but for future, it'd be good.
I think the next part of figuring out why there's such huge disparity is instrumenting bitcoind to find out where the time goes when relaying a block.
On Sun, Nov 24, 2013 at 5:26 PM, Gregory Maxwell <gmaxwell@gmail•com> wrote:
Quoted text (24 lines)
> On Sun, Nov 24, 2013 at 8:20 AM, Christian Decker > <decker.christian@gmail•com> wrote: > > Since this came up again during the discussion of the Cornell paper I > > thought I'd dig up my measurement code from the Information > > Propagation paper and automate it as much as possible. > > Could you publish the block ids and timestamp sets for each block? > > It would be useful in correlating propagation information against > block characteristics. > > > ------------------------------------------------------------------------------ > Shape the Mobile Experience: Free Subscription > Software experts and developers: Be at the forefront of tech innovation. > Intel(R) Software Adrenaline delivers strategic insight and game-changing > conversations that shape the rapidly evolving mobile landscape. Sign up > now. > http://pubads.g.doubleclick.net/gampad/clk?id=63431311&iu=/4140/ostg.clktrk > _______________________________________________ > Bitcoin-development mailing list > Bitcoin-development@lists•sourceforge•net > https://lists.sourceforge.net/lists/listinfo/bitcoin-development >
On Sun, Nov 24, 2013 at 05:20:22PM +0100, Christian Decker wrote:
Quoted text (16 lines)
> Since this came up again during the discussion of the Cornell paper I > thought I'd dig up my measurement code from the Information > Propagation paper and automate it as much as possible. > > The result is the Network Propagation page on bitcoinstats.com > (http://bitcoinstats.com/network/propagation/). It takes a daily > snapshot of the situation, then calculates the time until blocks and > transactions reach a certain percentile of the nodes in the network. > There is also a detailed page showing the density function describing > at what times nodes learn about the existence of a block/transaction > (for example yesterdays distribution: > http://bitcoinstats.com/network/propagation/2013/11/23). > > I intend to add more information and plots over time, but I wanted to > push this out quickly as there were some people asking for it. Hope > this helps getting the blockchain fork rate down :-)
Do you have the resources to save the raw log data? You'll also need to save transaction timestamp data - whether or not a given node has a transaction already matters re: propagation.
Of course given pool centralization the moment pools start peering directly with each other all these stats might not mean all that much.
Note that the number that's important isn't seconds, rather rather seconds/actual block interval as long as hashing power is growing. Unfortunately actually determining that is tricky - block interval is inherently noisy so you'll want to use a fairly agressively smoothed average.
So here's a rough calculation: right now blocks are happening roughly %15 faster than they would at equilibrium, and blockchain.info reports about 2 orphans a day. 2/166=1.2% orphan rate.
Now with a simplistic model where it takes exactly t seconds for a block to propagate to 100% of the hashing power, and until then 0% has it, you'd get:
orphan rate = t / actual block interval -> t = rate * interval
Or 6.2 seconds with our orphan rate data. Now whether or not blockchain.info succesfully captures all orphans I don't know, but given you're reporting 4.5 to 9.4 seconds for 50th and 75th percentile respectively that number 6.2s seems "ballpark" reasonable - remember that hashing power is definitely not distributed evenly among the nodes you are sampling from.
Which is another point... it may be the case that your propagation data doesn't actually give any insight into real-world orphan rates because the distribution of hashing power is concentrated into pools.
Hi Christian,
Cool - thanks for posting - agree, that it would be nice to normalize the results with block size - so divide by size and: 1. see if there is a correlation (we all presume there still is) 2. plot the delay graph as e.g. normalized to the averaged blocksize or lets define a "standard block size" of 200kb or what ever so we can compare the plot btw days.
Also, does the correlation of propagation times hold for transaction sizes as well (would be ice to find the logical t0 and the constant - I guess the interesting measure is not kb but signatures, so number of inputs - some correlation with size though).
Best,
Michael
On 24/11/13, 17:37 , Christian Decker wrote:
Quoted text (32 lines)
> Sure thing, I'm looking for a good way to publish these measurements, > but I haven't found a good option yet. They are rather large in size, > so I'd rather not serve them along with the website as it hasn't got > the capacity. Any suggestions? If the demand is not huge I could > provide them on a per user basis. > -- > Christian Decker > > > On Sun, Nov 24, 2013 at 5:26 PM, Gregory Maxwell <gmaxwell@gmail•com> wrote: >> On Sun, Nov 24, 2013 at 8:20 AM, Christian Decker >> <decker.christian@gmail•com> wrote: >>> Since this came up again during the discussion of the Cornell paper I >>> thought I'd dig up my measurement code from the Information >>> Propagation paper and automate it as much as possible. >> >> Could you publish the block ids and timestamp sets for each block? >> >> It would be useful in correlating propagation information against >> block characteristics. > > ------------------------------------------------------------------------------ > Shape the Mobile Experience: Free Subscription > Software experts and developers: Be at the forefront of tech innovation. > Intel(R) Software Adrenaline delivers strategic insight and game-changing > conversations that shape the rapidly evolving mobile landscape. Sign up now. > http://pubads.g.doubleclick.net/gampad/clk?id=63431311&iu=/4140/ostg.clktrk > _______________________________________________ > Bitcoin-development mailing list > Bitcoin-development@lists•sourceforge•net > https://lists.sourceforge.net/lists/listinfo/bitcoin-development >
Thanks Mike for the Tip :-)
I will definitely extend the calculations to include a size-normalized version. As for transaction propagations, being much smaller the measurements tend to be much noisier, but given enough samples we might be able to reconstruct some of the system parameters.
Good idea to attempt to correlate propagation speed and number of inputs/outputs, might be interesting to see whether processing at the nodes has an influence.
Regards, Chris
Hey Christian,
Could you sort the snapshots by date? At the moment they're kind of in a random order.
Sometimes I wish we had real-time stats too but this is a great start.
On Mon, Nov 25, 2013 at 8:27 PM, Christian Decker < decker.christian@gmail•com> wrote:
Quoted text (104 lines)
> Thanks Mike for the Tip :-) > > I will definitely extend the calculations to include a size-normalized > version. As for transaction propagations, being much smaller the > measurements tend to be much noisier, but given enough samples we > might be able to reconstruct some of the system parameters. > > Good idea to attempt to correlate propagation speed and number of > inputs/outputs, might be interesting to see whether processing at the > nodes has an influence. > > Regards, > Chris > -- > Christian Decker > > > On Mon, Nov 25, 2013 at 9:51 AM, Michael Gronager <gronager@ceptacle•com> > wrote: > > Hi Christian, > > > > Cool - thanks for posting - agree, that it would be nice to normalize > > the results with block size - so divide by size and: > > 1. see if there is a correlation (we all presume there still is) > > 2. plot the delay graph as e.g. normalized to the averaged blocksize or > > lets define a "standard block size" of 200kb or what ever so we can > > compare the plot btw days. > > > > Also, does the correlation of propagation times hold for transaction > > sizes as well (would be ice to find the logical t0 and the constant - I > > guess the interesting measure is not kb but signatures, so number of > > inputs - some correlation with size though). > > > > Best, > > > > Michael > > > > On 24/11/13, 17:37 , Christian Decker wrote: > >> Sure thing, I'm looking for a good way to publish these measurements, > >> but I haven't found a good option yet. They are rather large in size, > >> so I'd rather not serve them along with the website as it hasn't got > >> the capacity. Any suggestions? If the demand is not huge I could > >> provide them on a per user basis. > >> -- > >> Christian Decker > >> > >> > >> On Sun, Nov 24, 2013 at 5:26 PM, Gregory Maxwell <gmaxwell@gmail•com> > wrote: > >>> On Sun, Nov 24, 2013 at 8:20 AM, Christian Decker > >>> <decker.christian@gmail•com> wrote: > >>>> Since this came up again during the discussion of the Cornell paper I > >>>> thought I'd dig up my measurement code from the Information > >>>> Propagation paper and automate it as much as possible. > >>> > >>> Could you publish the block ids and timestamp sets for each block? > >>> > >>> It would be useful in correlating propagation information against > >>> block characteristics. > >> > >> > ------------------------------------------------------------------------------ > >> Shape the Mobile Experience: Free Subscription > >> Software experts and developers: Be at the forefront of tech innovation. > >> Intel(R) Software Adrenaline delivers strategic insight and > game-changing > >> conversations that shape the rapidly evolving mobile landscape. Sign up > now. > >> > http://pubads.g.doubleclick.net/gampad/clk?id=63431311&iu=/4140/ostg.clktrk > >> _______________________________________________ > >> Bitcoin-development mailing list > >> Bitcoin-development@lists•sourceforge•net > >> https://lists.sourceforge.net/lists/listinfo/bitcoin-development > >> > > > > > > > ------------------------------------------------------------------------------ > > Shape the Mobile Experience: Free Subscription > > Software experts and developers: Be at the forefront of tech innovation. > > Intel(R) Software Adrenaline delivers strategic insight and game-changing > > conversations that shape the rapidly evolving mobile landscape. Sign up > now. > > > http://pubads.g.doubleclick.net/gampad/clk?id=63431311&iu=/4140/ostg.clktrk > > _______________________________________________ > > Bitcoin-development mailing list > > Bitcoin-development@lists•sourceforge•net > > https://lists.sourceforge.net/lists/listinfo/bitcoin-development > > > ------------------------------------------------------------------------------ > Shape the Mobile Experience: Free Subscription > Software experts and developers: Be at the forefront of tech innovation. > Intel(R) Software Adrenaline delivers strategic insight and game-changing > conversations that shape the rapidly evolving mobile landscape. Sign up > now. > http://pubads.g.doubleclick.net/gampad/clk?id=63431311&iu=/4140/ostg.clktrk > _______________________________________________ > Bitcoin-development mailing list > Bitcoin-development@lists•sourceforge•net > https://lists.sourceforge.net/lists/listinfo/bitcoin-development >
Damn, that happens if I do the overview as an afterthought. Fixed :-)
Real time (last 24 hours, last week, last month) are in the pipeline, just need to find the time to implement access to the collector from the webpage.