[bitcoin-dev] Small Nodes: A Better Alternative to Pruned Nodes

Gregory Maxwell greg at xiph.org
Fri Apr 21 20:38:36 UTC 2017


On Mon, Apr 17, 2017 at 6:54 AM, David Vorick via bitcoin-dev
<bitcoin-dev at lists.linuxfoundation.org> wrote:
> Rationale:
>
> A node that stores the full blockchain (I will use the term archival node)
> requires over 100GB of disk space, which I believe is one of the most
> significant barriers to more people running full nodes. And I believe the
> ecosystem would benefit substantially if more users were running full nodes.

Hi David,

I've been thinking about this subject for a long time (Tier Nolan
linked some of the threads; see also the coloration part of
https://en.bitcoin.it/wiki/User:Gmaxwell/block_network_coding) and
about using FEC with the history for the last year or so. (we use FEC
in practice in fibre for relay at the tip now, and that has a design
history going back to 2013).

As Jonas points out, we're all set to also offer the 'recent blocks'
separately, so that is obviously going to happen and will help. The
free design parameter is the definition of recent, but we have good
measurements for setting that now. But about history...

I think I might have designed myself into a corner and perhaps you've
shown a way out-- I think there are some serious limits in your
proposal but perhaps we can fix them.  Let me give you what I had been
thinking about, then hand you at least a couple improvements to your
thinking:

As you hopefully now know (per Tier Nolan's) post: I'd previously been
thinking about nodes keeping a deterministic random, independently
selected subset which is self-leveling based on a small seed.   The
connection overhead can can be mitigated by working with chunks of
blocks rather than single blocks.

But as you've observed, the failure probabilities are rather high,
especially if an active attacker targets nodes carrying less commonly
available blocks.  So I thought, okay we can use FEC to help improve
the recovery odds.

So I considered using a sparse random code (E.g. an LDPC erasure code
like in RFC 5170) the advantage of these codes is that they are very
fast to decode, and they support having enormous codewords, so you can
a high probability of every peer having a unique ID, so there would
effectively never need to be any duplication.

But then I ran into the problem that I had no reasonable way to
recover from bad data, short of pulling a single group from an archive
then banning whatever peers gave you bad chunks.

So that is kind of where I got stuck.

I didn't even consider the advantage of only being able to use a few
peers total, as I still assumed that it would be doing the random
groups thing as well. That's a great point.

So on your design:

Being able to have a lower bound of 20% seems like a serious
limitation to me: it would be fine today, but the chain will quite
possibly be twice the current size in a year.... and in four years
we're back to where we are now. What I'd been thinking would be able
to scale arbitrarily low.

20% is small, but is it small enough? -- today that would get us back
into common SSDs being able to hold the whole chain, but that property
will probably be lost again in a couple years. I think with any fixed
fraction we'll constantly be fighting against the opportunity cost of
upgrading storage, if not the cost of the storage itself. (and I agree
this is the vast majority of the response from actual users,  sync
time and ongoing bandwidth usage seeming to tie for second; the latter
of which can be mitigated in other ways but for the former see my
remarks at the end)

The fact that having only a few required blocks needed lets you brute
force out the decode is a great point.  I hadn't considered that, and
the schemed I'd been considering are not communications efficient with
only a few blocks required e.g. they sometimes require a couple extra
blocks to successfully decode, which is a lot of overhead if you're
only splitting 5 ways.

> I also believe that alternative erasure coding schemes exist which actually
> are able to identify the bad pieces given sufficient good pieces, however I
> don't know if they have the same computational performance as the best
> Reed-Solomon coding implementations.

Actually RS codes are _the_ codes you want to use for with decoding
with errors but the 'computationally efficient' error correcting
decoding (which is itself heinously slow) requires 2*errors + data
number of blocks to decode. Not so useful if you're only looking at 5
parts.

RS decoding is pretty slow generally, esp compared to binary sparse
codes.  We were unable to make RS coding make sense for use in fast
block relay for this reason-- decoding time bottlenecked
reconstruction. The most highly optimized RS codes make a special
optimization which is not useful for your proposal: They're much
faster to decode if you're decoding from the first couple correction
blocks.  Without these optimizations speeds from from 1GB/s to more
like 100MB/s or worse.  Though I suppose with assumevalid initial sync
now is pretty cpu-idle on most hardware.  (If 100MB/s sounds fast,
keep in mind that time spent decoding is time that can't be spent
hashing/verifying/etc.. and on a lot hardware with fast broadband with
signature validation enabled we're cpu bound already)

> Another option would be to have the small nodes store a cryptographic
> checksum of each piece. Obtaining the cryptographic checksum for all 256
> pieces would incur a nontrivial amount of hashing (post segwit, as much as
> 100MB of extra hashing per block), and would require an additional ~4kb of
> storage per block. The hashing overhead here may be prohibitive.

This was a complete non-starter when thinking about using these sparse
codes where the number of possible blocks is effectively unlimited.

Your 4kb assumption isn't correct though: How you do it is that after
downloading a block you compute the 255 fragments (as an aside, you're
saying 256 but the most you can get is 255 for an 8-bit RS code).
then you compute a hashtree over them. Every node remembers the root,
and the membership proofs for their chunk(s)... this is 256 bytes of
extra storage.

When you decode you use a majority to decide what root you are trying
to decode. If it fails to result in valid blocks, you blacklist that
root, ban all those peers, and try the next.  Worst case cost is the
number of invalid roots rather than peers choose 5.

I'm not sure if combinitoral decode or the above would be easier to implement.

> This represents a non-trivial amount of code, but I believe that the result
> would be a non-trivial increase in the percentage of users running full
> nodes, and a healthier overall network.

The RS coder stuff is small, even doing the fancy w/ error decodes it
isn't huge. I expect complexity mostly will show up in the garbage
input handling.

A couple other thoughts:

The coded blocks are not useful for things like bloom scanning or
other lookup.  With the committed filter proposals you could still
keep the committed filters (even 5 way shared themselves...) so
perhaps not that much of a concern.

Coding the blocks will make their encoding normative. The current P2P
one we use is by no means more efficient,  Pieter and I have an
encoding that works on a transaction by transaction basis and
decreases the size of the whole chain by ~28%, block-wide entropy
encoding could reduce the size even further.  We'd hoped to at least
start using this per transaction coding locally, converting on the fly
to the original serialization where needed (the idea would be to
eventually support the compacted serialization on the wire too).
Maybe the answer there is to just get in whatever improvements we
think are reasonable before doing the coded block.

I think the 20% floor is still too high, and too many nodes will be
forced to just do the recent blocks things. perhaps not today, but
eventually.   I suppose we could just assume that the block is split
10 ways, and the default is two indexes, but there is flexibility to
go down to one in the future, at the cost of needing more peers...
could run numbers on the decode failure odds for the 10 of 255 case...
But that would only improve it to 10%.   I suppose the proposal could
be combined with sparse chain storage like I was thinking years ago,
but much less sparsity would be needed so the downsides would be less
severe.

E.g. if you want to store <10% of the chain you'd keep some 10% blocks
for random groups.  such a feature could be introduced later when it
was more important to keep less than 10 or 20 percent.

Recovery odds could be improved with a second level of coding. E.g. if
your ID was not a constant but a seed into PRNG(height)%250+5  then
you also have a fraction of nodes store the xor between adjacent pairs
then you can fill in unrecoverable groups.  the limit of this thinking
is exactly the sparse random code ECC schemes) Probably not worth the
complexity, but just a thing to keep in mind...

Probably the next step is to do some more focused benchmarks. I have
some code I can recommend offline.

I can't help but feel that this might be a little bit of a waste of
time. I think the long term survival of the system is likely going to
ultimately depend on doing an assumevalid like gated sync of a UTXO
set.  Ethereum already does something far more reckless (they let
miners pick the state and blindly trust it with almost no depth
required) and no one cares.  If that is done then all these concerns
are greatly reduced, along with the 100(+++?)GB/history-year transfer
costs.  You'd still want to have archives, but do you care about 20%
archives?

Replying to some other comments in the thread:

> A user may not want to run a node at home, but rather on a digital ocean or AWS server

This is almost if not quite entirely pointless. There are already many
nodes on AWS and DO, adding more does not improve your security (you
must trust DO/AWS), does not improve your reliability (DO/AWS), does
not improve the network's capacity, etc.  About the only arguable
benefit I can see there beyond making more money for these hosts is
feel good points for (incorrectly) thinking you're helping the
network, and negligibly reducing the density of spy nodes (but wait:
AWS/DO may well just be spying on your connections too..). and it it
isn't like fast storage on these services is cheap.

> Perhaps it is not, but I would think that it would be pretty straightforward to configure a global bandwidth limit within Bitcoin.

We have outbound transmission limits though not by default. Setting
them sensibly is something of a black art. Obviously these will
improve over time... and are more reasons that I agree with your
relative cost assumptions except for the sync-time part.

> Fingerprinting?

Unavoidable, but should be made inconsequential by making transaction
announcement more private independent of any of this. There are
already _MANY_ ways to fingerprint nodes, please don't think that
existing software has any real immunity here. We avoid adding new
fingerprinting where we can... but they're very fingerprintable
already. Transaction announcement privacy MUST be fixed.  I assume if
any of this is worth doing, it will also be worth the increase in
fingerprinting. And at least this proposal would 'only' create 255
node classes.

> SPV?

As I pointed out above, Vorick's idea is totally compatible with
committed filters, and the filters can even be shared like the blocks
are.

> [People who fail at math]

The SNR on this list really sucks.  If you aren't spending a couple
hours thinking about your responses or at least citing research that
took days of effort or more then you are probably wasting people's
time. Please respect the other users of the list.

> Could there be a slider?

Absolutely, I assume if Vorick's proposal were implemented that nodes
would have the follow options: Pruned [UTXO + recent two weeks of
blocks], 20%, 40%, 60%, 80%, 100% (archive).


More information about the bitcoin-dev mailing list