MIME-Version: 1.0
From: Olaoluwa Osuntokun <laolu32@gmail.com>
Date: Thu, 01 Jun 2017 19:01:14 +0000
Message-ID: <CAO3Pvs8ccTkgrecJG6KFbBW+9moHF-FTU+4qNfayeE3hM9uRrg@mail.gmail.com>
To: Arnoud Kouwenhoven - Pukaki Corp via bitcoin-dev
	<bitcoin-dev@lists.linuxfoundation.org>
Content-Type: multipart/alternative; boundary="94eb2c190790d124ca0550eaa932"
Subject: [bitcoin-dev] BIP Proposal: Compact Client Side Filtering for Light
	Clients
Precedence: list

--94eb2c190790d124ca0550eaa932
Content-Type: text/plain; charset="UTF-8"

Hi y'all,

Alex Akselrod and I would like to propose a new light client BIP for
consideration:
   * https://github.com/Roasbeef/bips/blob/master/gcs_light_client.mediawiki

This BIP proposal describes a concrete specification (along with a
reference implementations[1][2][3]) for the much discussed client-side
filtering reversal of BIP-37. The precise details are described in the
BIP, but as a summary: we've implemented a new light-client mode that uses
client-side filtering based off of Golomb-Rice coded sets. Full-nodes
maintain an additional index of the chain, and serve this compact filter
(the index) to light clients which request them. Light clients then fetch
these filters, query the locally and _maybe_ fetch the block if a relevant
item matches. The cool part is that blocks can be fetched from _any_
source, once the light client deems it necessary. Our primary motivation
for this work was enabling a light client mode for lnd[4] in order to
support a more light-weight back end paving the way for the usage of
Lightning on mobile phones and other devices. We've integrated neutrino
as a back end for lnd, and will be making the updated code public very
soon.

One specific area we'd like feedback on is the parameter selection. Unlike
BIP-37 which allows clients to dynamically tune their false positive rate,
our proposal uses a _fixed_ false-positive. Within the document, it's
currently specified as P = 1/2^20. We've done a bit of analysis and
optimization attempting to optimize the following sum:
filter_download_bandwidth + expected_block_false_positive_bandwidth. Alex
has made a JS calculator that allows y'all to explore the affect of
tweaking the false positive rate in addition to the following variables:
the number of items the wallet is scanning for, the size of the blocks,
number of blocks fetched, and the size of the filters themselves. The
calculator calculates the expected bandwidth utilization using the CDF of
the Geometric Distribution. The calculator can be found here:
https://aakselrod.github.io/gcs_calc.html. Alex also has an empirical
script he's been running on actual data, and the results seem to match up
rather nicely.

We we're excited to see that Karl Johan Alm (kallewoof) has done some
(rather extensive!) analysis of his own, focusing on a distinct encoding
type [5]. I haven't had the time yet to dig into his report yet, but I
think I've read enough to extract the key difference in our encodings: his
filters use a binomial encoding _directly_ on the filter contents, will we
instead create a Golomb-Coded set with the contents being _hashes_ (we use
siphash) of the filter items.

Using a fixed fp=20, I have some stats detailing the total index size, as
well as averages for both mainnet and testnet. For mainnet, using the
filter contents as currently described in the BIP (basic + extended), the
total size of the index comes out to 6.9GB. The break down is as follows:

    * total size:  6976047156
    * total avg:  14997.220622758816
    * total median:  3801
    * total max:  79155
    * regular size:  3117183743
    * regular avg:  6701.372750217131
    * regular median:  1734
    * regular max:  67533
    * extended size:  3858863413
    * extended avg:  8295.847872541684
    * extended median:  2041
    * extended max:  52508

In order to consider the average+median filter sizes in a world worth
larger blocks, I also ran the index for testnet:

    * total size:  2753238530
    * total avg:  5918.95736054141
    * total median:  60202
    * total max:  74983
    * regular size:  1165148878
    * regular avg:  2504.856172982827
    * regular median:  24812
    * regular max:  64554
    * extended size:  1588089652
    * extended avg:  3414.1011875585823
    * extended median:  35260
    * extended max:  41731

Finally, here are the testnet stats which take into account the increase
in the maximum filter size due to segwit's block-size increase. The max
filter sizes are a bit larger due to some of the habitual blocks I
created last year when testing segwit (transactions with 30k inputs, 30k
outputs, etc).

     * total size:  585087597
     * total avg:  520.8839608674402
     * total median:  20
     * total max:  164598
     * regular size:  299325029
     * regular avg:  266.4790836307566
     * regular median:  13
     * regular max:  164583
     * extended size:  285762568
     * extended avg:  254.4048772366836
     * extended median:  7
     * extended max:  127631

For those that are interested in the raw data, I've uploaded a CSV file
of raw data for each block (mainnet + testnet), which can be found here:
     * mainnet: (14MB):
https://www.dropbox.com/s/4yk2u8dj06njbuv/mainnet-gcs-stats.csv?dl=0
     * testnet: (25MB):
https://www.dropbox.com/s/w7dmmcbocnmjfbo/gcs-stats-testnet.csv?dl=0


We look forward to getting feedback from all of y'all!

-- Laolu


[1]: https://github.com/lightninglabs/neutrino
[2]: https://github.com/Roasbeef/btcd/tree/segwit-cbf
[3]: https://github.com/Roasbeef/btcutil/tree/gcs/gcs
[4]: https://github.com/lightningnetwork/lnd/

-- Laolu

--94eb2c190790d124ca0550eaa932
Content-Type: text/html; charset="UTF-8"
Content-Transfer-Encoding: quoted-printable

<div dir=3D"ltr"><div>Hi y&#39;all,=C2=A0</div><div><br></div><div>Alex Aks=
elrod and I would like to propose a new light client BIP for</div><div>cons=
ideration:=C2=A0</div><div>=C2=A0 =C2=A0* <a href=3D"https://github.com/Roa=
sbeef/bips/blob/master/gcs_light_client.mediawiki">https://github.com/Roasb=
eef/bips/blob/master/gcs_light_client.mediawiki</a></div><div><br></div><di=
v>This BIP proposal describes a concrete specification (along with a</div><=
div>reference implementations[1][2][3]) for the much discussed client-side<=
/div><div>filtering reversal of BIP-37. The precise details are described i=
n the</div><div>BIP, but as a summary: we&#39;ve implemented a new light-cl=
ient mode that uses</div><div>client-side filtering based off of Golomb-Ric=
e coded sets. Full-nodes</div><div>maintain an additional index of the chai=
n, and serve this compact filter</div><div>(the index) to light clients whi=
ch request them. Light clients then fetch</div><div>these filters, query th=
e locally and _maybe_ fetch the block if a relevant</div><div>item matches.=
 The cool part is that blocks can be fetched from _any_</div><div>source, o=
nce the light client deems it necessary. Our primary motivation</div><div>f=
or this work was enabling a light client mode for lnd[4] in order to</div><=
div>support a more light-weight back end paving the way for the usage of</d=
iv><div>Lightning on mobile phones and other devices. We&#39;ve integrated =
neutrino</div><div>as a back end for lnd, and will be making the updated co=
de public very</div><div>soon.</div><div><br></div><div>One specific area w=
e&#39;d like feedback on is the parameter selection. Unlike</div><div>BIP-3=
7 which allows clients to dynamically tune their false positive rate,</div>=
<div>our proposal uses a _fixed_ false-positive. Within the document, it&#3=
9;s</div><div>currently specified as P =3D 1/2^20. We&#39;ve done a bit of =
analysis and</div><div>optimization attempting to optimize the following su=
m:</div><div>filter_download_bandwidth + expected_block_false_positive_band=
width. Alex</div><div>has made a JS calculator that allows y&#39;all to exp=
lore the affect of</div><div>tweaking the false positive rate in addition t=
o the following variables:</div><div>the number of items the wallet is scan=
ning for, the size of the blocks,</div><div>number of blocks fetched, and t=
he size of the filters themselves. The</div><div>calculator calculates the =
expected bandwidth utilization using the CDF of</div><div>the Geometric Dis=
tribution. The calculator can be found here:</div><div><a href=3D"https://a=
akselrod.github.io/gcs_calc.html">https://aakselrod.github.io/gcs_calc.html=
</a>. Alex also has an empirical</div><div>script he&#39;s been running on =
actual data, and the results seem to match up</div><div>rather nicely.</div=
><div><br></div><div>We we&#39;re excited to see that Karl Johan Alm (kalle=
woof) has done some</div><div>(rather extensive!) analysis of his own, focu=
sing on a distinct encoding</div><div>type [5]. I haven&#39;t had the time =
yet to dig into his report yet, but I</div><div>think I&#39;ve read enough =
to extract the key difference in our encodings: his</div><div>filters use a=
 binomial encoding _directly_ on the filter contents, will we</div><div>ins=
tead create a Golomb-Coded set with the contents being _hashes_ (we use</di=
v><div>siphash) of the filter items.</div><div><br></div><div>Using a fixed=
 fp=3D20, I have some stats detailing the total index size, as</div><div>we=
ll as averages for both mainnet and testnet. For mainnet, using the</div><d=
iv>filter contents as currently described in the BIP (basic + extended), th=
e</div><div>total size of the index comes out to 6.9GB. The break down is a=
s follows:</div><div><br></div><div>=C2=A0 =C2=A0 * total size: =C2=A069760=
47156</div><div>=C2=A0 =C2=A0 * total avg: =C2=A014997.220622758816</div><d=
iv>=C2=A0 =C2=A0 * total median: =C2=A03801</div><div>=C2=A0 =C2=A0 * total=
 max: =C2=A079155</div><div>=C2=A0 =C2=A0 * regular size: =C2=A03117183743<=
/div><div>=C2=A0 =C2=A0 * regular avg: =C2=A06701.372750217131</div><div>=
=C2=A0 =C2=A0 * regular median: =C2=A01734</div><div>=C2=A0 =C2=A0 * regula=
r max: =C2=A067533</div><div>=C2=A0 =C2=A0 * extended size: =C2=A0385886341=
3</div><div>=C2=A0 =C2=A0 * extended avg: =C2=A08295.847872541684</div><div=
>=C2=A0 =C2=A0 * extended median: =C2=A02041</div><div>=C2=A0 =C2=A0 * exte=
nded max: =C2=A052508</div><div><br></div><div>In order to consider the ave=
rage+median filter sizes in a world worth</div><div>larger blocks, I also r=
an the index for testnet:=C2=A0</div><div><br></div><div>=C2=A0 =C2=A0 * to=
tal size: =C2=A02753238530</div><div>=C2=A0 =C2=A0 * total avg: =C2=A05918.=
95736054141</div><div>=C2=A0 =C2=A0 * total median: =C2=A060202</div><div>=
=C2=A0 =C2=A0 * total max: =C2=A074983</div><div>=C2=A0 =C2=A0 * regular si=
ze: =C2=A01165148878</div><div>=C2=A0 =C2=A0 * regular avg: =C2=A02504.8561=
72982827</div><div>=C2=A0 =C2=A0 * regular median: =C2=A024812</div><div>=
=C2=A0 =C2=A0 * regular max: =C2=A064554</div><div>=C2=A0 =C2=A0 * extended=
 size: =C2=A01588089652</div><div>=C2=A0 =C2=A0 * extended avg: =C2=A03414.=
1011875585823</div><div>=C2=A0 =C2=A0 * extended median: =C2=A035260</div><=
div>=C2=A0 =C2=A0 * extended max: =C2=A041731</div><div><br></div><div>Fina=
lly, here are the testnet stats which take into account the increase</div><=
div>in the maximum filter size due to segwit&#39;s block-size increase. The=
 max</div><div>filter sizes are a bit larger due to some of the habitual bl=
ocks I</div><div>created last year when testing segwit (transactions with 3=
0k inputs, 30k</div><div>outputs, etc).</div><div><br></div><div>=C2=A0 =C2=
=A0 =C2=A0* total size: =C2=A0585087597</div><div>=C2=A0 =C2=A0 =C2=A0* tot=
al avg: =C2=A0520.8839608674402</div><div>=C2=A0 =C2=A0 =C2=A0* total media=
n: =C2=A020</div><div>=C2=A0 =C2=A0 =C2=A0* total max: =C2=A0164598</div><d=
iv>=C2=A0 =C2=A0 =C2=A0* regular size: =C2=A0299325029</div><div>=C2=A0 =C2=
=A0 =C2=A0* regular avg: =C2=A0266.4790836307566</div><div>=C2=A0 =C2=A0 =
=C2=A0* regular median: =C2=A013</div><div>=C2=A0 =C2=A0 =C2=A0* regular ma=
x: =C2=A0164583</div><div>=C2=A0 =C2=A0 =C2=A0* extended size: =C2=A0285762=
568</div><div>=C2=A0 =C2=A0 =C2=A0* extended avg: =C2=A0254.4048772366836</=
div><div>=C2=A0 =C2=A0 =C2=A0* extended median: =C2=A07</div><div>=C2=A0 =
=C2=A0 =C2=A0* extended max: =C2=A0127631</div><div><br></div><div>For thos=
e that are interested in the raw data, I&#39;ve uploaded a CSV file</div><d=
iv>of raw data for each block (mainnet + testnet), which can be found here:=
</div><div>=C2=A0 =C2=A0 =C2=A0* mainnet: (14MB): <a href=3D"https://www.dr=
opbox.com/s/4yk2u8dj06njbuv/mainnet-gcs-stats.csv?dl=3D0">https://www.dropb=
ox.com/s/4yk2u8dj06njbuv/mainnet-gcs-stats.csv?dl=3D0</a></div><div>=C2=A0 =
=C2=A0 =C2=A0* testnet: (25MB): <a href=3D"https://www.dropbox.com/s/w7dmmc=
bocnmjfbo/gcs-stats-testnet.csv?dl=3D0">https://www.dropbox.com/s/w7dmmcboc=
nmjfbo/gcs-stats-testnet.csv?dl=3D0</a></div><div><br></div><div><br></div>=
<div>We look forward to getting feedback from all of y&#39;all!</div><div><=
br></div><div>-- Laolu</div><div><br></div><div><br></div><div>[1]: <a href=
=3D"https://github.com/lightninglabs/neutrino">https://github.com/lightning=
labs/neutrino</a></div><div>[2]: <a href=3D"https://github.com/Roasbeef/btc=
d/tree/segwit-cbf">https://github.com/Roasbeef/btcd/tree/segwit-cbf</a></di=
v><div>[3]: <a href=3D"https://github.com/Roasbeef/btcutil/tree/gcs/gcs">ht=
tps://github.com/Roasbeef/btcutil/tree/gcs/gcs</a></div><div>[4]: <a href=
=3D"https://github.com/lightningnetwork/lnd/">https://github.com/lightningn=
etwork/lnd/</a></div><div><br></div><div>-- Laolu</div><div><br></div></div=
>

--94eb2c190790d124ca0550eaa932--