The probability for a success forward of x amount on a channel is now
computed as:
P_success(x) = b * P_rene(x)
where P_rene(x) is the probability distribution proposed by Rene
Pickhardt with a uniform distribution on the known liquidity of the
channel and 'b' is a new parameter we add in this PR, by the name of
"base probability". The "base probability" represents the probability
for a channel in the network choosen at random to be able to forward at
least 1msat, ie. of being alive, non-depleted and with at least one HTLC
slot. We don't know the value of 'b', but for the moment we assume that
it is 0.98 and use that as default.
As a consequence the probability cost becomes non-linear and non-convex
because of the additional constant term:
Cost(x) = - log P_success(x)
= - log b - - log P_rene(x)
= - log b + Cost_rene(x)
We currently don't handle well base fees and neither this additional
"base probability" but we can as a first approximation linearize the
cost function in this way:
Cost_linear(x) = (- log b)*x + Cost_rene_linear(x)
Changelog-Added: renepay: Add a dev parameter to renepay that represents
a constant probability of availability for all channels in the network.
Signed-off-by: Lagrang3 <lagrang3@protonmail.com>
Expecting to have more than just a bunch of disabled channels if we
prune the lightning network heavily I am changing the internal data
structure of disabledmap from a simple array to a hashtable.
Have a finer control over disabled channels by targeting
short_channel_id_dir instead of short_channel_id,
ie. we can disable a single direction of a channel.
Signed-off-by: Lagrang3 <lagrang3@protonmail.com>
The route-builder checks the liquidity bounds of each route one at a
time. Every route that satisfy the contraints is recorded in the
uncertainty network and produces an HTLC burden on the channels it uses,
so that the following routes cannot count on the same liquidity twice.
gcc-12 (Ubuntu 12.3.0-9ubuntu2) 12.3.0 with -O3 gives:
```
In file included from plugins/renepay/test/../mcf.c:5,
from plugins/renepay/test/run-arc.c:13:
ccan/ccan/tal/str/str.h: In function ‘minflow’:
ccan/ccan/tal/str/str.h:43:9: error: ‘errmsg’ may be used uninitialized [-Werror=maybe-uninitialized]
43 | tal_fmt_(ctx, TAL_LABEL(char, "[]"), __VA_ARGS__)
| ^~~~~~~~
plugins/renepay/test/../mcf.c:1565:15: note: ‘errmsg’ was declared here
1565 | char *errmsg;
| ^~~~~~
cc1: all warnings being treated as errors
```
Signed-off-by: Rusty Russell <rusty@rustcorp.com.au>
In the Min. Cost Flow solver we put a constraint on arcs capacities,
such that the total flow that can be allocated through a channel does
not exceed htlc_max. This solves two previous problem of the htlc_max
constraint at the MCF level.
In some weird situations it may happen that some channel along the route
could have htlcmax=htlcmin, so that the supremum of htlcmin and the
infimum of htlcmax are the same number. In that case there is only one
allowed amount that can go through that route.
Without this patch renepay would not handle correctly this cornercase.
Min. Cost Flow does not take into account fees when computing a flow
with liquidity constraints.
This is a work-around solution that reduces the amount on every route to
respect the liquidity bound. The deficity in the delivered amount is
solved by running MCF once again.
Changes:
1. the function `flow_complete` allocates amounts to send over the set of routes
computed by the MCF algorithm, but it does not allocate more than liquidity
bound of the route. For this reason `minflow` returns a set of routes that
satisfy the liquidity bounds but it is not guaranteed that the total payment
reaches the destination therefore there could a deficit in the delivery:
`deficit = amount_to_deliver - delivering`.
2. in the function `add_payflows` after `minflow` returns a set of routes we
call `flows_fit_amount` that tries to a allocate the `deficit` in the routes
that the MCF have computed.
3. if the resulting flows pass all payment constraints then we update
`amount_to_deliver = amount_to_deliver - delivering`, and the loop
repeats as long as `amount_to_deliver` is not zero.
In other words, the excess amount, beyond the liquidity bound,
in the routes is removed and then we try to allocate it
into known routes, otherwise we do a whole MCF again just for the
remaining amount.
Fixes issue #6599
Adds a "canonical" error handling in the module mcf.c,
adhering to the same convention we have adopted in flow.c.
So that when a function fails it signals the caller with an invalid
return value, or bool false, and it communicates an error message.
It's expressed in bits, but really it's clearer as a quantity, given
how it's used.
Suggested-by: @Lagrang3
Signed-off-by: Rusty Russell <rusty@rustcorp.com.au>
- adopt "const <type> *"convention
- remove use_shadow option for some pyln tests
- show prob. information of flows into paynotes
- show prob. of success of entire payment flow in paynotes
- minflow: We were not releasing the memory of flow arrays when replacing
them with a new canditate.
- use memleak_scan_obj in memleak_check
- replace u64 with size_t
Signed-off-by: Lagrang3 <eduardo.quintana@pm.me>
The global is an *internal* hack because dijkstra_item_mover doesn't
take a context arg! It should be used with care.
Easy, since all the accessors exist: we just hand in the struct dijkstra.
Signed-off-by: Rusty Russell <rusty@rustcorp.com.au>
- remove internal gheap checks
- add check for arc_t.chanidx overflow
- remove outdated comments
- check the delta flow bounds before augmenting along a path
- get_flow_paths uses a dynamic tal array instead of a list.
- fix a unit test that depended on the order of returned flows
- fix bug: lightnind doesn't like if I reuse the partid of a failed
flow, therefore use a higher partid than any of the previous attempts.
- plugin_err instead of LOG_BROKEN if sendpay fails and we cannot get a
an error code.
- fix wrong comments.
- remove the background timer.
- This is a bugfix. Previous to this the MCF network was built using the
knowledge of the min and max liquidity but it didn't take into account
pending HTLCs.
- Also remove the min_prob_success option but hardcode a 90% value.
Removing some options that are not relevant to the user, they're kept
for developer mode only:
- base_fee_penalty
- min_prob_success
- prob_cost_factor
- remove heap.h, not used
Signed-off-by: Lagrang3 <eduardo.quintana@pm.me>
