Re: [Bitcoin-development] Request for review/testing: headers-first synchronization in Bitcoin Core

[Jameson Lopp <jameson.lopp@gmail.com>]

Great work, Pieter. I've been spooling up several nodes per week lately and can testify that stalled downloads during initial syncing are a pain. I usually forgo bootstrapping on VPSes because I don't want to have to adjust the disk space allocation.

With headers-first I'm saturating my home cable connection with download rates of 4 MB/s until block 295,000 at which point CPU becomes the bottleneck and it settles down in the 1 MB/s range.

It took 6 minutes for my node to sync to block height 100,000
22 minutes to reach height 200,000
62 minutes to reach height 250,000
125 minutes to reach height 295,000
144 minutes to reach height 300,000
248 minutes to reach height 325,000

- Jameson

On 10/11/2014 07:34 PM, Pieter Wuille wrote:
> Hi all,
> 
> I believe that a large change that I've been working on for Bitcoin
> Core is ready for review and testing: headers-first synchronization.
> In short, it changes the way the best chain is discovered, downloaded
> and verified, with several advantages:
> * Parallel block downloading (much faster sync on typical network connections).
> * No more stalled downloads.
> * Much more robust against unresponsive or slow peers.
> * Removes a class of DoS attacks related to peers feeding you
> low-difficulty valid large blocks on a side branch.
> * Reduces the need for checkpoints in the code.
> * No orphan blocks stored in memory anymore (reducing memory usage during sync).
> * A major step step towards an SPV mode using the reference codebase.
> 
> Historically, this mode of operation has been known for years (Greg
> Maxwell wrote up a description of a very similar method in
> https://en.bitcoin.it/wiki/User:Gmaxwell/Reverse_header-fetching_sync
> in early 2012, but it was known before that), but it took a long time
> to refactor these code enough to support it.
> 
> Technically, it works by replacing the single-peer blocks download by
> a single-peer headers download (which typically takes seconds/minutes)
> and verification, and simultaneously fetching blocks along the best
> known headers chain from all peers that are known to have the relevant
> blocks. Downloading is constrained to a moving window to avoid
> unbounded unordering of blocks on disk (which would interfere with
> pruning later).
> 
> At the protocol level, it increases the minimally supported version
> for peers to 31800 (corresponding to bitcoin v3.18, released in
> december 2010), as earlier versions did not support the getheaders P2P
> message.
> 
> So, the code is available as a github pull request
> (https://github.com/bitcoin/bitcoin/pull/4468), or packaged on
> http://bitcoin.sipa.be/builds/headersfirst, where you can also find
> binaries to test with.
> 
> Known issues:
> * At the very start of the sync, especially before all headers are
> processed, downloading is very slow due to a limited number of blocks
> that are requested per peer simultaneously. The policies around this
> will need some experimentation can certainly be improved.
> * Blocks will be stored on disk out of order (in the order they are
> received, really), which makes it incompatible with some tools or
> other programs. Reindexing using earlier versions will also not work
> anymore as a result of this.
> * The block index database will now hold headers for which no block is
> stored on disk, which earlier versions won't support. If you are fully
> synced, it may still be possible to go back to an earlier version.
> 
> Unknown issues:
> * Who knows, maybe it will replace your familiy pictures with Nyan
> Cat? Use at your own risk.
> 
> TL;DR: Review/test https://github.com/bitcoin/bitcoin/pull/4468 or
> http://bitcoin.sipa.be/builds/headersfirst.
>