1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 370 371 372 373 374 375 376 377 378 379 380 381 382 383 384 385 386 387 388 389 390 391 392 393 394 395 396 397 398 399 400 401 402 403 404 405 406 407 408 409 410 411 412 413 414 415 416 417 418 419 420 421 422 423 424 425 426 427 428 429 430 431 432 433 434 435 436 437 438 439 440 441 442 443 444 445 446 447 448 449 450 451 452 453 454 455 456 457 458 459 460 461 462 463 464 465 466 467 468 469 470 471 472 473 474 475 476 477 478 479 480 481 482 483 484 485 486 487 488 489 490 491 492 493 494 495 496 497 498 499 500
use kitsune_p2p_dht_arc::DhtArc;
use num_traits::Zero;
use crate::spacetime::{SpaceOffset, Topology};
use super::{is_full, Arq, ArqClamping, ArqStrat};
/// A "view" of the peers in a neighborhood. The view consists of a few
/// observations about the distribution of peers within a particular arc, used
/// to make inferences about the rest of the (out-of-view) DHT, ultimately
/// enabling the calculation of the target arc size for the agent who has this View.
///
/// The enum allows us to add different views (and different calculations of
/// target arc length) over time.
#[derive(derive_more::From)]
pub enum PeerView {
/// The quantized PeerView
Quantized(PeerViewQ),
}
impl PeerView {
/// Given the current view of a peer and the peer's current coverage,
/// this returns the next step to take in reaching the ideal coverage.
pub fn update_arc(&self, dht_arc: &mut DhtArc) -> bool {
match self {
Self::Quantized(v) => {
let mut arq = Arq::from_dht_arc_approximate(&v.topo, &v.strat, dht_arc);
let updated = v.update_arq(&mut arq);
*dht_arc = arq.to_dht_arc(&v.topo);
updated
}
}
}
}
/// The Quantized PeerView
pub struct PeerViewQ {
/// The strategy which generated this view
strat: ArqStrat,
/// The topology of the network space
pub topo: Topology,
/// The peers in this view (TODO: replace with calculated values)
peers: Vec<Arq>,
#[cfg(feature = "test_utils")]
/// Omit the arq at this index from all peer considerations.
/// Useful for tests which update all arqs, without needing to
/// construct a new PeerView for each arq needing to be updated
pub skip_index: Option<usize>,
}
impl PeerViewQ {
/// Constructor
pub fn new(topo: Topology, strat: ArqStrat, peers: Vec<Arq>) -> Self {
Self {
strat,
topo,
peers,
#[cfg(feature = "test_utils")]
skip_index: None,
}
}
/// The actual coverage of all arcs in this view.
/// TODO: this only makes sense when the view contains all agents in the DHT.
/// So, it's more useful for testing. Probably want to tease out some
/// concept of a test DHT from this.
pub fn actual_coverage(&self) -> f64 {
actual_coverage(&self.topo, self.peers.iter())
}
/// Extrapolate the coverage of the entire network from our local view.
pub fn extrapolated_coverage(&self, filter: &Arq) -> f64 {
self.extrapolated_coverage_and_filtered_count(filter).0
}
/// Return the extrapolated coverage and the number of arqs which match the filter.
/// These two are complected together simply for efficiency's sake, to
/// minimize computation
///
/// TODO: this probably will be rewritten when PeerView is rewritten to
/// have the filter baked in.
pub fn extrapolated_coverage_and_filtered_count(&self, filter: &Arq) -> (f64, usize) {
let filter = filter.to_dht_arc(&self.topo);
if filter.is_empty() {
// More accurately this would be 0, but it's handy to not have
// divide-by-zero crashes
return (1.0, 1);
}
let filter_len = filter.length();
let initial = (0, 0);
// FIXME: We can't just filter arcs on the fly here, because we might be
// trying to get coverage info for an area we don't have arcs for
// (because we don't store arcs for agents outside of our arc).
// So, we need to extrapolate the arcs we do have to extend into the
// unknown area outside the filter.
// For now though, just filter arcs on the fly so we have something to test.
// But, this means that the behavior for growing arcs is going to be a bit
// different in the future.
let (sum, count) = self
.filtered_arqs(filter)
.fold(initial, |(sum, count), arq| {
(sum + arq.absolute_length(&self.topo), count + 1)
});
let cov = sum as f64 / filter_len as f64;
(cov, count)
}
/// Compute the total coverage observed within the filter interval.
pub fn raw_coverage(&self, filter: &Arq) -> f64 {
self.extrapolated_coverage(filter) * filter.to_dht_arc_range(&self.topo).length() as f64
/ 2f64.powf(32.0)
}
/// Mutate the arq to its ideal target
pub fn update_arq(&self, arq: &mut Arq) -> bool {
let topo = &self.topo;
let strat = &self.strat;
match strat.local_storage.arc_clamping {
Some(ArqClamping::Empty) => {
let changed = arq.is_empty();
*arq.count_mut() = 0;
changed
}
Some(ArqClamping::Full) => {
let changed = arq.is_full(topo);
*arq = Arq::new_full(topo, arq.start, topo.max_space_power(strat));
changed
}
None => self.update_arq_with_stats(arq).changed,
}
}
fn is_slacking(&self, cov: f64, num_peers: usize) -> bool {
num_peers as f64 <= cov * self.strat.slacker_ratio
}
/// The "slacker" factor. If our observed coverage is significantly
/// greater than the number of peers we see, it's an indication
/// that we may need to pick up more slack.
///
/// This check helps balance out stable but unequitable situations where
/// all peers have a similar estimated coverage, but some peers are
/// holding much more than others.
pub fn slack_factor(&self, cov: f64, num_peers: usize) -> f64 {
if self.is_slacking(cov, num_peers) {
if num_peers.is_zero() {
// Prevent a NaN.
// This value gets clamped anyway, so it will never actually
// lead to an infinite value.
f64::INFINITY
} else {
cov / num_peers as f64
}
} else {
1.0
}
}
fn growth_factor(&self, cov: f64, num_peers: usize, median_power_diff: i8) -> f64 {
let np = num_peers as f64;
let under = cov < self.strat.min_coverage;
let over = cov > self.strat.max_coverage();
// The ratio of ideal coverage vs actual observed coverage.
// A ratio > 1 indicates undersaturation and a need to grow.
// A ratio < 1 indicates oversaturation and a need to shrink.
let cov_diff = if over || under {
let ratio = self.strat.midline_coverage() / cov;
// We want to know which of our peers are likely to be making a similar
// update to us, because that will affect the overall coverage more
// than the drop in the bucket that we can provide.
//
// If all peers have seen the same change as us since their last update,
// they will on average move similarly to us, and so we should only make
// a small step in the direction of the target, trusting that our peers
// will do the same.
//
// Conversely, if all peers are stable, e.g. if we just came online to
// find a situation where all peers around us are under-representing,
// but stable, then we want to make a much bigger leap.
let peer_dampening_factor = 1.0 / (1.0 + np);
(ratio - 1.0) * peer_dampening_factor + 1.0
} else {
1.0
};
// The "slacker" factor. If our observed coverage is significantly
// greater than the number of peers we see, it's an indication
// that we may need to pick up more slack.
//
// This check helps balance out stable but unequitable situations where
// all peers have a similar estimated coverage, but some peers are
// holding much more than others.
let slack_factor = self.slack_factor(cov, num_peers);
let unbounded_growth = cov_diff * slack_factor;
// The difference between the median power and the arq's power helps
// determine some limits on growth.
// If we are at the median growth, then it makes sense to cap
// our movement by 2x in either direction (1/2 to 2)
//
// If we are 1 below the median, then our range is (1/2 to 4)
// If we are 2 below the median, then our range is (1/2 to 8)
// If we are 1 above the median, then our range is (1/4 to 2)
// If we are 2 above the median, then our range is (1/8 to 2)
//
// Note that there is also a hard limit on growth described by
// ArqStrat::max_power_diff, enforced elsewhere.
let mpd = median_power_diff as f64;
let min = 2f64.powf(mpd).min(0.5);
let max = 2f64.powf(mpd).max(2.0);
unbounded_growth.clamp(min, max)
}
/// Take an arq and potentially resize and requantize it based on this view.
///
/// This represents an iterative step towards the ideal coverage, based on
/// the observed coverage.
/// This makes many assumptions, including:
/// - this arc resizing algorithm is a good one, namely that the coverage
/// at any point of the DHT is close to the target range
/// - all other peers are following the same algorithm
/// - if we see a change that we need to make, we assume that a number of
/// peers are about to make a similar change, and that number is on
/// average the same as our target coverage
///
/// More detail on these assumptions here:
/// <https://hackmd.io/@hololtd/r1IAIbr5Y/https%3A%2F%2Fhackmd.io%2FK_fkBj6XQO2rCUZRRL9n2g>
/// TODO: make the above link to something publicly available, preferably in the repo
pub fn update_arq_with_stats(&self, arq: &mut Arq) -> UpdateArqStats {
let topo = &self.topo;
let (cov, num_peers) = self.extrapolated_coverage_and_filtered_count(arq);
let old_count = arq.count();
let old_power = arq.power();
let power_stats = self.power_stats(topo, arq);
let PowerStats {
median: median_power,
..
} = power_stats;
let median_power_diff = median_power as i8 - arq.power() as i8;
let growth_factor = self.growth_factor(cov, num_peers, median_power_diff);
let new_count = if growth_factor < 1.0 {
// Ensure we shrink by at least 1
(old_count as f64 * growth_factor).floor() as u32
} else {
// Ensure we grow by at least 1 (if there is any growth at all)
(old_count as f64 * growth_factor).ceil() as u32
};
if new_count != old_count {
let mut tentative = *arq;
tentative.count = SpaceOffset(new_count);
// If shrinking caused us to go below the target coverage,
// or to start "slacking" (not seeing enough peers), then
// don't update. This happens when we shrink too much and
// lose sight of peers.
let (new_cov, new_num_peers) =
self.extrapolated_coverage_and_filtered_count(&tentative);
if new_count < old_count
&& (new_cov < self.strat.min_coverage
|| (!self.is_slacking(cov, num_peers)
&& self.is_slacking(new_cov, new_num_peers)))
{
return UpdateArqStats {
changed: false,
desired_delta: new_count as i32 - old_count as i32,
power: None,
num_peers,
};
}
}
// Commit the change to the count
arq.count = SpaceOffset(new_count);
let power_above_min = |pow| {
// not already at the minimum
pow > topo.min_space_power()
// don't power down if power is already too low
&& (median_power as i8 - pow as i8) < self.strat.max_power_diff as i8
};
loop {
// check for power downshift opportunity
if *arq.count < self.strat.min_chunks() {
if power_above_min(arq.power) {
*arq = arq.downshift();
} else {
// If we could not downshift due to other constraints, then we cannot
// shrink any smaller than the min_chunks.
arq.count = SpaceOffset(self.strat.min_chunks());
}
} else {
break;
}
}
let power_below_max = |pow| {
// not already at the maximum
pow < topo.max_space_power(&self.strat)
// don't power up if power is already too high
&& (pow as i8 - median_power as i8) < self.strat.max_power_diff as i8
};
loop {
// check for power upshift opportunity
if *arq.count > self.strat.max_chunks() {
if power_below_max(arq.power) {
// Attempt to requantize to the next higher power.
// If we only grew by one chunk, into an odd count, then don't
// force upshifting, because that would either require undoing
// the growth, or growing by 2 instead of 1. In this case, skip
// upshifting, and we'll upshift on the next update.
let force = new_count as i32 - old_count as i32 > 1;
if let Some(a) = arq.upshift(force) {
*arq = a
} else {
break;
}
} else {
// If we could not upshift due to other constraints, then we cannot
// grow any larger than the max_chunks.
arq.count = SpaceOffset(self.strat.max_chunks());
}
} else {
break;
}
}
if is_full(topo, arq.power(), arq.count()) {
*arq = Arq::new_full(topo, arq.start_loc(), arq.power());
}
// check if anything changed
let changed = !(arq.power() == old_power && arq.count() == old_count);
UpdateArqStats {
changed,
desired_delta: new_count as i32 - old_count as i32,
power: Some(power_stats),
num_peers,
}
}
/// Gather statistics about the power levels of all arqs in this view.
/// Used to make inferences about what your neighbors are up to.
pub fn power_stats(&self, topo: &Topology, filter: &Arq) -> PowerStats {
use statrs::statistics::*;
let mut powers: Vec<_> = self
.filtered_arqs(filter.to_dht_arc(topo))
.filter(|a| *a.count > 0)
.map(|a| a.power as f64)
.collect();
powers.push(filter.power() as f64);
let powers = statrs::statistics::Data::new(powers);
let median = powers.median() as u8;
let std_dev = powers.std_dev().unwrap_or_default();
if std_dev > self.strat.power_std_dev_threshold {
// tracing::warn!("Large power std dev: {}", std_dev);
}
PowerStats { median, std_dev }
}
/// Filter to return only **non-zero** arcs whose start lies in the filtering arc
fn filtered_arqs(&self, filter: DhtArc) -> impl Iterator<Item = &Arq> {
let it = self.peers.iter();
#[cfg(feature = "test_utils")]
let it = it
.enumerate()
.filter(|(i, _)| self.skip_index.as_ref() != Some(i))
.map(|(_, arq)| arq);
it.filter(move |arq| !arq.is_empty() && filter.contains(arq.start_loc()))
}
}
/// A summary of what happened while updating an Arq
#[derive(Debug, Clone)]
pub struct UpdateArqStats {
/// Did the arq change?
pub changed: bool,
/// How much did the arq "want" to change?
pub desired_delta: i32,
/// PowerStats, if calculated.
pub power: Option<PowerStats>,
/// Number of peers initially visible from this arq.
pub num_peers: usize,
}
/// The actual coverage provided by these peers. Assumes that this is the
/// entire view of the DHT, all peers are accounted for here.
pub fn actual_coverage<'a, P: Iterator<Item = &'a Arq>>(topo: &Topology, peers: P) -> f64 {
peers
.map(|a| a.absolute_length(topo) as f64 / 2f64.powf(32.0))
.sum()
}
/// Statistics about the power levels of all arqs in a view.
/// Used to make inferences about what your neighbors are up to.
#[derive(Debug, Clone)]
pub struct PowerStats {
/// The median power level
pub median: u8,
/// The standard deviation of power levels
pub std_dev: f64,
}
#[cfg(test)]
mod tests {
use std::convert::identity;
use kitsune_p2p_dht_arc::DhtArcRange;
use crate::arq::{pow2, print_arqs};
use crate::spacetime::Topology;
use crate::ArqBounds;
use super::*;
fn make_arq(topo: &Topology, pow: u8, lo: u32, hi: u32) -> Arq {
let (a, _) = ArqBounds::from_interval_rounded(
topo,
pow,
DhtArcRange::from_bounds(
pow2(pow) * lo,
((pow2(pow) as u64 * hi as u64) as u32).wrapping_sub(1),
),
);
a.to_arq(topo, identity)
}
#[test]
fn test_filtered_arqs() {
let topo = Topology::unit_zero();
let pow = 25;
let a = make_arq(&topo, pow, 0, 0x20);
let b = make_arq(&topo, pow, 0x10, 0x30);
let c = make_arq(&topo, pow, 0x20, 0x40);
let arqs = vec![a, b, c];
print_arqs(&topo, &arqs, 64);
let view = PeerViewQ::new(topo.clone(), ArqStrat::default(), arqs);
let get = |b: Arq| {
view.filtered_arqs(b.to_dht_arc(&topo))
.cloned()
.collect::<Vec<_>>()
};
assert_eq!(get(make_arq(&topo, pow, 0, 0x10)), vec![a]);
assert_eq!(get(make_arq(&topo, pow, 0, 0x20)), vec![a, b]);
assert_eq!(get(make_arq(&topo, pow, 0, 0x40)), vec![a, b, c]);
assert_eq!(get(make_arq(&topo, pow, 0x10, 0x20)), vec![b]);
}
#[test]
fn test_coverage() {
let topo = Topology::unit_zero();
let pow = 24;
let arqs: Vec<_> = (0..0x100)
.step_by(0x10)
.map(|x| make_arq(&topo, pow, x, x + 0x20))
.collect();
print_arqs(&topo, &arqs, 64);
let view = PeerViewQ::new(topo.clone(), ArqStrat::default(), arqs);
assert_eq!(
view.extrapolated_coverage_and_filtered_count(&make_arq(&topo, pow, 0, 0x10)),
(2.0, 1)
);
assert_eq!(
view.extrapolated_coverage_and_filtered_count(&make_arq(&topo, pow, 0, 0x20)),
(2.0, 2)
);
assert_eq!(
view.extrapolated_coverage_and_filtered_count(&make_arq(&topo, pow, 0, 0x40)),
(2.0, 4)
);
// TODO: when changing PeerView logic to bake in the filter,
// this will probably change
assert_eq!(
view.extrapolated_coverage_and_filtered_count(&make_arq(&topo, pow, 0x10, 0x20)),
(2.0, 1)
);
}
}