[bitcoin-dev] Capacity increases for the Bitcoin system.

[Anthony Towns]

TLDR:

  1.7MB effective block size is a better estimate than 1.6MB for p2pkh
  with segwit. 2MB for 2/2 multisig still seems accurate.

  Additional post-segwit soft forked script improvements can improve
  the effective block size for p2pkh txns from 1.7MB to 1.9MB, and for
  2/2 multisig from 2MB to 2.5MB/3MB.

  (To the best of my knowledge, anyway; if I've made a mistake in my
  maths or assumptions, corrections appreciated)

On Tue, Dec 08, 2015 at 02:58:03PM +1000, Anthony Towns via bitcoin-dev wrote:
> So from IRC, this doesn't seem quite right -- capacity is constrained as
>   base_size + witness_size/4 <= 1MB
..
> That would be 1.6MB and 2MB of total actual data if you hit the limits
> with real transactions, so it's more like a 1.8x increase for real
> transactions afaics, even with substantial use of multisig addresses.

I think these numbers are slightly mistaken -- I was only aware of version
1 segwit scripts at the time, and assumed 256 bit hashes would be used
for all segwit transaction, however version 0 segwit txns would be more
efficient for p2pkh, with the same security as bitcoin currently has
(which seems fine).

Also, segwit will make two additional soft-fork improvements possible that
would have a positive effect on transactions per block without requiring
more data per block: ecdsa public key recovery (more space efficient for
*both* multisig and p2pkh) and schnorr signatures (more space efficient
multisig) which might also improve things. I don't know how soon they're
planned to be worked on post segwit's roll out; basic Schnorr signatures
are in the Elements sidechain, but I don't think key recovery has been
implemented anywhere? (Actually, I guess they could both be done already
via softforking OP_NOP opcodes, though segwit makes them slightly
cleaner)

Anyhoo here's some revised figures, working explained in the footnotes.
If I've made mistakes, corrections appreciated, of course.

p2pkh:

  now: 10+146i+34o [0]
  segwit: 10+41i+36o + 0.25*105*i [1]
  ecdsa recovery: 10+41i+33o + 0.25*71*i [2]
  80-bit schnorr: 10+41i+33o + 0.25*71*i (same as ecdsa recovery imo [3])
  128-bit schnorr: 10+41i+43o + 0.25*106*i [4]

(128-bit schnorr provides a not very useful increase in security here)

2-of-2 multisig:

  now: 10+254i+32o [5]
  segwit: 10+43i+43o + 0.25*213*i [6]
  ecdsa recovery: 10+43i+43o + 0.25+187*i [7]
  80-bit schnorr: 10+41i+33o + 0.25*71*i (same as p2pkh)
  128-bit schnorr: 10+41i+43o + 0.25*106*i (same as p2pkh)

(segwit, ecdsa recovery and 128-bit schnorr all provide a beneficial
security increase here, as per the "Time to worry about 80-bit collision
attacks" thread; 80-bit schnorr provides the same security as current
p2sh multisig)

Using the same assumptions in the previous mail, ie that over the long
term number inputs is about the same as number of outputs, these
simplify to:

        p2pkh           2-of-2 msig
now     10+180i         10+286i
segwit  10+104i         10+140i
recov   10+92i          10+133i
sch80   10+92i          10+92i
sch128  10+111i         10+111i

Translating "now" to 100%, the scaling factors work out to be:

i=1, i->inf

        p2pkh           2-of-2 msig
now     100%            100%
segwit  166%-173%       197%-204%
recov   186%-195%       207%-215%
sch80   186%-195%       290%-310%
sch128  157%-162%       244%-257%

So 170% for p2pkh (rather my original estimate of 160%) and 200% for
multisig (same as my original estimate), which can rise via further
soft-forks up to 190% for p2pkh and 250% or 300% for 2-of-2 multisig
(depending on whether you want additional security for 2/2 multisig
beyond what's currently available).

(I'm assuming people are mostly interested in the number of transactions
per block (or tx/second or tx/day); if miners are worried about the
actual data per block (which effects orphan rates) implied by the above,
but don't want to work it out themselves, I could do the maths for that
too pretty easily. Let me know)


If a 2MB hard fork is done first, then the 1/4 discount for segwit could
mean up to 8MB of total data per block -- from what I understand this
is currently infeasible; so I presume that segwit on top of a hardfork
and prior to IBLT/weak blocks would need to have a smaller discount or
no discount applied so as to ensure total data per block remains at 4MB
or less. With no discount for witness data (ie, no "accounting tricks")
those figures look like:

        p2pkh           2-of-2 msig
now     100%            100%
segwit  99%             95%
recov   122%-124%       104%
sch80   122%-124%       191%-198%
sch128  94%-95%         148%-150%

That is, without discounting, segwit comes at a slight cost in
transactions per block, and additional soft forks will only result in
25% gain for p2pkh (via key recovery) and 50%-100% for 2-of-2 multisig
(through the use of schnorr sigs and key recovery, and depending on
whether you want 128 bits of security rather than 80 bits).

(So without the discounting factor, with a 2MB block size, 2MB per block
with segwit and key recovery gives you 25% more p2pkh transactions than
just 2MB per block now; while segwit and schnorr signatures gives you
50%-100% more 2/2 multisig transactions in the same 2MB. Likewise with
1MB or 4MB and no discounting. Discounting has the indirect benefit of
providing a monetary incentive to limit UTXO sizes however)


(2 of 3 multisig for escrow payments would probably be interesting to
work out too; I think ecdsa key recovery combined with 1/4 discounting
would provide a substantial improvement there. I don't think Schnorr
helps at all for that case, unfortunately; and it's probably too small
scale for merkle-ised abstract syntax trees to do any good either)


A caveat: I'm only counting the script data from witnesses here; but it's
possible that additional metadata (such as a length for each witness
signature, or the value of the input, or even some/all of the merkle
hashes) should also be accounted for. I don't think any of them need to
be accounted for segwit as proposed, but I'm not sure. And it might well
be different for a hardforked segwit; there I have no idea at all. I
don't think a byte or two for length would make much difference, at least.

Cheers,
aj

[0] 10 bytes for version (4), input count (1), output count (1) and
    locktime (4); 146 bytes per input consisting of tx hash (32), txout
    index (4), script length (1), scriptsig (signature and pubkey =
    105), CHECKSIG = 25), and sequence number (4); 34 bytes per output
    consisting of value (8), script length (1) and scriptpubkey (DUP
    HASH160 PUSH20 EQVERIFY CHECKSIG = 25).

[1] Same as now, except two extra bytes per output script (segwit push and
    segwit version byte), and moving the 105 bytes of signature script
    directly into the segregated witness

[2] Allowing ECDSA recovery requires an additional soft-fork post segwit
    to change the CHECKSIG operation; this requires bumping the
    segwit script version to 2 or higher and possibly using a different
    opcode, but allows the scriptsig to just be the 70 byte signature,
    without also including the 33 byte pubkey. The 33 byte pubkey is
    automatically calculated from the signature, and verified against
    the hash provided in the scriptpubkey to maintain security, with a
    scriptpubkey like: [PUSH (20 byte pubkey hash) CHECKSIG_RECOVER] (22
    bytes versus 25 bytes), and a scriptsig like [PUSH (70 byte sig)]
    (71 bytes versus 105 bytes).

[3] libsecp256k1 has a function to recover a pubkey from a schnorr
    signature, so I'm assuming that means pubkey recovery with schnorr
    is possible :) -- I haven't actually verified the maths
    https://github.com/bitcoin/secp256k1/blob/master/include/secp256k1_schnorr.h

[4] The witness scriptpubkey is limited to 32 bytes (plus push op and
    version byte for a total of 34 bytes, so 128 bit security requires
    version 1 segwit, and p2sh-style constuction. Hence: 10 bytes
    (version, input and output counts and locktime); 41 base bytes per
    input (tx hash, tx index, script length, and sequence number); 106
    witness bytes per input (sig (70 bytes) plus witness script (PUSH
    schnorr merged pubkey (32 bytes) plus CHECKSCHNORR), plus PUSH ops);
    and 43 bytes per output (value, script length, and 34 bytes for the
    v1-style witness script).

[5] Per input is (32 bytes tx hash, 4 bytes tx index, 4 bytes nsequence,
    1 byte scriptsig length, 143 bytes for actual signature (2x70
    for the sigs, 3 bytes for OP_0 and two OP_PUSH), and 70 bytes for
    the redeemscript (2 pub pub 2 OP_CHECKMULTISIG)) for 254 bytes;
    per output is (8 bytes value, 1 byte length, 23 bytes for HASH160
    [20 byte hash] OP_EQUAL) for 32 bytes.

[6] Per input is (34 bytes tx hash, 4 bytes tx index, 4 bytes nsequence,
    1 byte scriptsig length) for 43 bytes in the base block and (143
    bytes for the actual signature, plus 70 bytes for the redeemscript)
    for 213 bytes of witness data; per output is (8 bytes value, 1 byte
    length, and 34 bytes for version 1 segwit scriptpubkey) for 43
    bytes.

[7] Same as [6], but with key recovery on a MULTISIG op, rather than 33
    bytes per pubkey, this could be reduced to a 20 byte pubkey hash
    per pubkey, for a saving of 26 bytes of witness data.