bump crate to 0.6.1

- Use array-init for the lowest level vector operations. Makes them nicer and less bug-prone, with zero performance impact.
- Remove unnecessary/unused ::new() functions on Vector and HeightVector.
- Update README with new dependency.
- Fix typo in README.

Signed-off-by: Stephen Waits <steve@waits.net>

.build.yml

@@ -1,26 +1,48 @@
 image: archlinux
 packages:
   - rustup
+  - cargo-tarpaulin
 sources:
   - https://git.sr.ht/~swaits/vivaldi-nc
+artifacts:
+  - vivaldi-nc/tarpaulin-report.html
 tasks:
   - setup: |
       rustup toolchain install nightly stable
       cd vivaldi-nc/
       rustup run stable cargo fetch
-  - stable: |
+  - format: |
       rustup default stable
       cd vivaldi-nc/
       cargo fmt --verbose --check
+  - lint: |
+      rustup default stable
+      cd vivaldi-nc/
       cargo clippy --verbose
+  - coverage: |
+      rustup default stable
+      cd vivaldi-nc/
+      cargo tarpaulin --verbose --all-features --out Html
+  - stable: |
+      rustup default stable
+      cd vivaldi-nc/
       cargo build --verbose
-      cargo test --verbose
+      cargo test --verbose --all-features
   - nightly: |
       rustup default nightly
       cd vivaldi-nc/
       cargo clippy --verbose ||:
       cargo build --verbose  ||:
-      cargo test --verbose   ||:
+      cargo test --verbose --all-features ||:
+  - bench: |
+      rustup default stable
+      cd vivaldi-nc/
+      cargo bench --verbose --all-features
+  - examples: |
+      rustup default stable
+      cd vivaldi-nc/
+      cargo run --verbose --all-features --example planetlab
+      cargo run --verbose --all-features --example planetlab --release
   - docs: |
       cd vivaldi-nc/
       rustup run stable cargo doc --no-deps

.gitignore

@@ -1,2 +1,4 @@
+/flamegraph.svg
 /target
 Cargo.lock
+/tarpaulin-report.html

Cargo.toml

@@ -1,6 +1,6 @@
 [package]
 name = "vivaldi-nc"
-version = "0.5.3"
+version = "0.6.1"
 edition = "2021"
 authors = ["Stephen Waits <steve@waits.net>"]
 description = "Vivaldi network coordinates for fast, distributed latency estimates in multinode networks, with a clean and simple interface."
@@ -16,17 +16,28 @@ exclude = [
 ]
 
 [features]
-default = ["f32"]
+default = []
 f32 = []
-f64 = []
+
+[profile.bench]
+debug = true
+
+[profile.release]
+debug = true
 
 [dev-dependencies]
 assert_approx_eq = "1.1.0"
+criterion = "0.4.0"
 proptest = "1.0.0"
 serde_json = "1.0.91"
 
+[[bench]]
+name = "million-updates"
+harness = false
+
 [dependencies]
-nanorand = {version = "0.7.0", features = ["wyrand", "getrandom"]}
+array-init = "2.1.0"
 num-traits = "0.2.15"
+rand = "0.8.5"
 serde = {version = "1.0.152", features = ["serde_derive"]}
 serde_with = "2.2.0"

README.md

@@ -112,18 +112,20 @@ usage examples.
 
 ### Cargo Features
 
-By default, the internal data structures and operations all use `f32`. If you want to use `f64`, you can enable it as a cargo feature:
+By default, the internal data structures and operations all use `f64`, which is
+slightly faster on modern architectures. If you want to use `f32` instead, you
+can enable it as a cargo feature:
 
 ```toml
 [dependencies]
-vivaldi-nc = { version = "0.5.2", features = ["f64"] }
+vivaldi-nc = { version = "(version)", features = ["f32"] }
 ```
 
 ### Examples
 
-The repository includes an example which loads a n-to-n latency sample from
-PlanetLab and iterates on `NetworkCoordinate`s until some low-enough mean error
-is reached. The output is a JSON array of elements which contain the NC's
+The repository includes an example which loads a 490 node N-to-N latency sample
+from PlanetLab and iterates on `NetworkCoordinate`s until some low-enough mean
+error is reached. The output is a JSON array of elements which contain the NC's
 position, its height (or stem latency estimate), and its estimation of error
 (lower is better) that look like this:
 
@@ -172,10 +174,11 @@ One design goal of this crate is to minimize dependencies. When dependencies
 are required, I try to be very selective about them (re: bloat and licensing).
 This crate depends on:
 
-- [`nanorand`](https://crates.io/crates/nanorand): A fast PRNG based on WyRand.
-  It pulls in `getrandom` as a portable source of entropy. I chose this over
-  the more commonly used, but heavyweight `rand` because it's significantly
-  smaller and does just what I need, and not much more.
+- [`rand`](https://crates.io/crates/rand): I started with `nanorand`; however,
+  this is faster because of `rand::thread_rng()`; whereas with `nanorand` we
+  ended up needing to reimplement that same functionality through something
+  like `lazy_static!` plus `Mutex`, which is not worth it for a few percentage
+  points of performance.
 - [`num-traits`](https://crates.io/crates/num-traits): A brilliant crate which
   makes it easier to operate on numbers in generics; like using `Float` as a
   constraint on a generic type. Its convenience outweighs its cost.
@@ -185,12 +188,17 @@ This crate depends on:
 - [`serde_with`](https://crates.io/crates/serde_with): Because the inner
   `Vector<T,N>` uses a const generic length, we use this to help derive
   `Deserialize`.
+- [`array-init`](https://crates.io/crates/array-init): Makes the array
+  operations for the internal `Vector` so much nicer. Once we have
+  `feature(array_zip)` stabilized, we could use that instead. Until then, this
+  works well with zero performance impact.
 
 ## Design Goals & Alternatives
 
 Several crates implemented Vivaldi NC before this one. So, why another?
 
-I had several design goals which the existing crates didn't satisfy:
+I had several design goals which the existing crates didn't satisfy, in order
+of priority:
 
 1. Provide the *simplest interface possible*. I just want to have some sort of
    Coordinate struct and be able to update it, and then use it to estimate
@@ -202,7 +210,10 @@ I had several design goals which the existing crates didn't satisfy:
 3. Serializable/Deserializable by default. These things are useful for sending
    across networks. The intent is to be able to do that without a bunch of
    rigamarole. Support `serde` traits by default.
-4. Well documented. Well tested. This varies by the crates.
+4. Well documented. Well tested. This varies in the existing crates.
+5. Performance. While it's a low priority, it should be reasonable. The library
+   currently does about 10,000 `NetworkCoordinate::update()` calls per
+   millisecond on my Macbook Pro (Intel i7-9750H (12) @ 2.60GHz).
 
 ### Other Vivaldi NC Implementations
 

benches/million-updates.rs

@@ -0,0 +1,52 @@
+use std::time::Duration;
+
+use criterion::{criterion_group, criterion_main, Criterion};
+use vivaldi_nc::{NetworkCoordinate2D, NetworkCoordinate3D};
+
+const NUM_NODES: usize = 1_000;
+
+fn million_vivaldi_2d_updates() {
+    // create NUM_NODES NCs
+    let mut nc: Vec<NetworkCoordinate2D> =
+        (0..NUM_NODES).map(|_| NetworkCoordinate2D::new()).collect();
+
+    // loop, update each item for every other item
+    for i in 0..NUM_NODES {
+        let nc_i = nc[i].clone();
+        for (j, nc_j) in nc.iter_mut().enumerate().take(NUM_NODES) {
+            if i == j {
+                continue;
+            }
+            let rtt = if i > j { i - j } else { j - i };
+            // let nc_j = nc[j].clone();
+            nc_j.update(&nc_i, Duration::from_millis(rtt.try_into().unwrap()));
+        }
+    }
+}
+
+fn million_vivaldi_3d_updates() {
+    // create NUM_NODES NCs
+    let mut nc: Vec<NetworkCoordinate3D> =
+        (0..NUM_NODES).map(|_| NetworkCoordinate3D::new()).collect();
+
+    // loop, update each item for every other item
+    for i in 0..NUM_NODES {
+        let nc_i = nc[i].clone();
+        for (j, nc_j) in nc.iter_mut().enumerate().take(NUM_NODES) {
+            if i == j {
+                continue;
+            }
+            let rtt = if i > j { i - j } else { j - i };
+            // let nc_j = nc[j].clone();
+            nc_j.update(&nc_i, Duration::from_millis(rtt.try_into().unwrap()));
+        }
+    }
+}
+
+fn run_benchmarks(c: &mut Criterion) {
+    c.bench_function("million 2D updates", |b| b.iter(million_vivaldi_2d_updates));
+    c.bench_function("million 3D updates", |b| b.iter(million_vivaldi_3d_updates));
+}
+
+criterion_group!(benches, run_benchmarks);
+criterion_main!(benches);

examples/planetlab.rs

@@ -2,13 +2,25 @@
 //
 // for info on the format, see `NetLatency-Data/README.md`
 
+//
+// **** Features ****
+//
+
+/// `FloatType` is a type alias for either `f32` or `f64` depending on cargo features
+#[cfg(feature = "f32")]
+type FloatType = f32;
+
+/// `FloatType` is a type alias for either `f32` or `f64` depending on cargo features
+#[cfg(not(feature = "f32"))]
+type FloatType = f64;
+
 use std::{
     fs::File,
     io::{BufReader, Read},
     time::Duration,
 };
 
-use nanorand::{Rng, WyRand};
+use rand::Rng;
 use vivaldi_nc::NetworkCoordinate;
 
 // this will return a Vec<Vec<Duration>>
@@ -66,10 +78,10 @@ fn main() {
     // 4. adjust local NC
     let mut error = 0.0;
     for _ in 0..30_000 {
-        let mut rng = WyRand::new();
+        let mut rng = rand::thread_rng();
         // 1. choose a local and remote NC
-        let i_local = rng.generate::<usize>() % NUM_NODES;
-        let i_remote = rng.generate::<usize>() % NUM_NODES;
+        let i_local = rng.gen_range(0..NUM_NODES);
+        let i_remote = rng.gen_range(0..NUM_NODES);
         if i_local == i_remote {
             continue;
         }
@@ -85,7 +97,7 @@ fn main() {
         nc[i_local].update(&nc_remote, rtt_measured);
 
         // complete, now let's sum up the error
-        error = nc.iter().map(|n| n.error()).sum::<f32>() / NUM_NODES as f32;
+        error = nc.iter().map(|n| n.error()).sum::<FloatType>() / NUM_NODES as FloatType;
         assert!(error.is_finite());
         if error < 5.0 {
             break;

proptest-regressions/height_vector.txt

@@ -9,3 +9,4 @@ cc 38ed1e2c313317809f5469d856d2c1c61e4851c64a156cd1bdd0757fa7c77128 # shrinks to
 cc b7d1b5e9fe3e33066b583c4532865373179972e9b5786ecc62438f23bc939e87 # shrinks to x = -0.0, y = 0.0, h = -3.448118e-33, m = -0.0
 cc 3c4fc4f21c82ffe5bcbc740ca9ab598e0894226acb86a3046a3a24cbbcdbf8e6 # shrinks to x = 0.0, y = 0.0, h = -0.0
 cc d1a551d7eb92bf69e3d0cd0b677baac5aaa5a95d6a016fccf67cf57af71c8b0c # shrinks to x = -5.098216e26, y = 0.0, h = 0.0
+cc 7fa6f2f56f2398bae6804c5f0c0f3f47a7304e505c34f3addb676546cefd97f2 # shrinks to x = -0.0, y = 1.500389e38, h = 2.828499e38

src/height_vector.rs

@@ -44,7 +44,7 @@
 
 use core::ops::{Add, Mul, Sub};
 
-use nanorand::{Rng, WyRand};
+use rand::prelude::*;
 use serde::{Deserialize, Serialize};
 
 use crate::vector::{self, Vector};
@@ -58,7 +58,7 @@ use crate::vector::{self, Vector};
 type FloatType = f32;
 
 /// `FloatType` is a type alias for either `f32` or `f64` depending on cargo features
-#[cfg(feature = "f64")]
+#[cfg(not(feature = "f32"))]
 type FloatType = f64;
 
 //
@@ -92,19 +92,14 @@ pub struct HeightVector<const N: usize> {
 //
 
 impl<const N: usize> HeightVector<N> {
-    /// A new height vector is a random unit vector
-    pub(crate) fn new() -> Self {
-        Self::random()
-    }
-
     /// A new height vector is a random unit vector
     pub(crate) fn random() -> Self {
-        let mut rng = WyRand::new();
+        let mut rng = rand::thread_rng();
         let mut vec = [0.0; N];
         for i in vec.iter_mut().take(N) {
-            *i = rng.generate::<FloatType>();
+            *i = rng.gen::<FloatType>() - 0.5;
         }
-        let height = rng.generate::<FloatType>().abs();
+        let height = rng.gen::<FloatType>().abs();
         Self {
             position: Vector::<FloatType, N>::from(vec),
             height,
@@ -122,15 +117,16 @@ impl<const N: usize> HeightVector<N> {
     /// by the inveerse of its length.
     pub(crate) fn normalized(&self) -> Self {
         let len = self.len();
-        if len < FloatType::EPSILON {
-            // if we have a bad vector, generate a new random vector
-            Self::new()
+        // scale it by the inverse of its magnitude like any normal vector
+        let ret = Self {
+            position: self.position / len,
+            height: self.height / len,
+        };
+        if ret.is_valid() {
+            ret
         } else {
-            // otherwise scale it by the inverse of its magnitude like any normal vector
-            Self {
-                position: self.position / len,
-                height: self.height / len,
-            }
+            // if we have a bad vector, generate a new random vector
+            Self::random()
         }
     }
 
@@ -255,7 +251,11 @@ mod tests {
                 assert_approx_eq!(a.len(), 1.0);
             } else {
                 // `HeightVector` we created from proptest values should be valid and have a length
-                assert_approx_eq!(a.len(), len);
+                if len.is_nan() || len.is_infinite() {
+                    assert!(a.len().is_nan() || a.len().is_infinite());
+                } else {
+                    assert_approx_eq!(a.len(), len);
+                }
             }
         }
 
@@ -313,7 +313,7 @@ mod tests {
     #[test]
     fn test_new() {
         // new gives us a random unit length vec
-        let a = HeightVector::<3>::new();
+        let a = HeightVector::<3>::random();
         assert_approx_eq!(a.len(), 1.0);
     }
 
@@ -373,4 +373,36 @@ mod tests {
         let t = serde_json::to_string(&a);
         assert_eq!(t.as_ref().unwrap(), s);
     }
+
+    #[test]
+    fn test_sub_invalid() {
+        let valid = HeightVector::<2>::from(([1.0, 2.0], 3.0));
+        let mut invalid = HeightVector::<2>::from(([1.0, 2.0], 3.0));
+        // force invalidity here - because `::from()` actually also catches this
+        invalid.height = 1.0 / 0.0;
+        let result = valid - invalid;
+        assert_approx_eq!(result.len(), 1.0);
+    }
+
+    #[test]
+    fn test_add_invalid() {
+        let valid = HeightVector::<2>::from(([1.0, 2.0], 3.0));
+        let mut invalid = HeightVector::<2>::from(([1.0 / 0.0, 2.0], 3.0));
+
+        // make sure `from()` caught and replaced the invalid vec
+        assert_approx_eq!(invalid.len(), 1.0);
+
+        // force invalidity here - because `::from()` actually also catches this
+        invalid.height = 1.0 / 0.0;
+        let result = valid + invalid;
+        assert_approx_eq!(result.len(), 1.0);
+    }
+
+    #[test]
+    fn test_zero_norm() {
+        let a = HeightVector::<2>::from(([0.0, 0.0], 0.0));
+        let b = a.normalized();
+        let len = b.len();
+        assert_approx_eq!(len, 1.0);
+    }
 }

src/lib.rs

@@ -108,10 +108,6 @@
 mod height_vector;
 mod vector;
 
-// make sure we don't have conflicting features
-#[cfg(all(feature = "f32", feature = "f64"))]
-compile_error!("feature \"f32\" and feature \"f64\" cannot be enabled at the same time");
-
 // publish our interface
 pub mod network_coordinate;
 pub use network_coordinate::NetworkCoordinate;

src/network_coordinate.rs

@@ -14,7 +14,7 @@ use crate::height_vector::HeightVector;
 #[cfg(feature = "f32")]
 type FloatType = f32;
 
-#[cfg(feature = "f64")]
+#[cfg(not(feature = "f32"))]
 type FloatType = f64;
 
 //
@@ -127,7 +127,7 @@ impl<const N: usize> NetworkCoordinate<N> {
         #[cfg(feature = "f32")]
         return Duration::from_secs_f32((self.heightvec - rhs.heightvec).len() / 1000.0);
 
-        #[cfg(feature = "f64")]
+        #[cfg(not(feature = "f32"))]
         return Duration::from_secs_f64((self.heightvec - rhs.heightvec).len() / 1000.0);
     }
 
@@ -195,14 +195,17 @@ impl<const N: usize> NetworkCoordinate<N> {
         #[cfg(feature = "f32")]
         let rtt_estimated_ms = self.estimated_rtt(rhs).as_secs_f32() * 1000.0;
 
-        #[cfg(feature = "f64")]
+        #[cfg(not(feature = "f32"))]
         let rtt_ms = rtt.as_secs_f64() * 1000.0;
-        #[cfg(feature = "f64")]
+        #[cfg(not(feature = "f32"))]
         let rtt_estimated_ms = self.estimated_rtt(rhs).as_secs_f64() * 1000.0;
 
         // rtt needs to be positive
         if rtt_ms < 0.0 {
-            return self;
+            // Note: `rtt` is guaranteed to be positive because `Duration` enforces it.
+            //       If this panics, something changed where `Duration` now allows for negative
+            //       values.
+            unreachable!();
         }
 
         // Sample weight balances local and remote error. (1)
@@ -233,6 +236,10 @@ impl<const N: usize> NetworkCoordinate<N> {
         // error
         if self.heightvec.is_invalid() {
             *self = Self::new();
+
+            // We should never get here because the call to `normalized()` above should catch any
+            // invalid `heightvec`
+            unreachable!();
         }
 
         // return reference to updated self
@@ -376,4 +383,11 @@ mod tests {
         let estimate = a.estimated_rtt(&b);
         assert_approx_eq!(estimate.as_secs_f32(), 0.080_099);
     }
+
+    #[test]
+    fn test_error_getter() {
+        let s = "{\"position\":[1.5,0.5,2.0],\"height\":25.0,\"error\":1.0}";
+        let a: NetworkCoordinate<3> = serde_json::from_str(s).unwrap();
+        assert_approx_eq!(a.error(), 1.0);
+    }
 }

src/vector.rs

@@ -3,6 +3,7 @@
 //! This is the simplest possible implementation of a vector to support the needs of a
 //! Vivaldi `HeightVector`.
 
+use array_init::array_init;
 use num_traits::Float;
 use serde::{Deserialize, Serialize};
 use serde_with::serde_as;
@@ -37,11 +38,6 @@ impl<T, const N: usize> Vector<T, N>
 where
     T: Float + Serialize + for<'d> Deserialize<'d>,
 {
-    /// Create a new Zero vector.
-    pub(crate) fn new() -> Self {
-        Self::default()
-    }
-
     /// Compute magnitude of vector.
     pub(crate) fn len(&self) -> T {
         // use hypot() instead of sqrt() of sum of squares to avoid overflows
@@ -90,11 +86,9 @@ where
 
     /// Vector+Vector addition.
     fn add(self, rhs: Self) -> Self::Output {
-        let mut ret = [T::zero(); N]; //Self::Output::new();
-        for i in 0..N {
-            ret[i] = self[i] + rhs[i];
+        Self::Output {
+            inner: array_init(|i| self[i] + rhs[i]),
         }
-        Self::Output::from(ret)
     }
 }
 
@@ -106,11 +100,9 @@ where
 
     /// Vector-Vector subtraction.
     fn sub(self, rhs: Self) -> Self::Output {
-        let mut ret = [T::zero(); N]; //Self::Output::new();
-        for i in 0..N {
-            ret[i] = self[i] - rhs[i];
+        Self::Output {
+            inner: array_init(|i| self[i] - rhs[i]),
         }
-        Self::Output::from(ret)
     }
 }
 
@@ -122,11 +114,9 @@ where
 
     /// Vector*scalar multiplication.
     fn mul(self, rhs: T) -> Self::Output {
-        let mut ret = Self::Output::new();
-        for i in 0..N {
-            ret.inner[i] = self.inner[i] * rhs;
+        Self::Output {
+            inner: array_init(|i| self[i] * rhs),
         }
-        ret
     }
 }
 
@@ -138,11 +128,9 @@ where
 
     /// Vector/scalar division.
     fn div(self, rhs: T) -> Self::Output {
-        let mut ret = Self::Output::new();
-        for i in 0..N {
-            ret.inner[i] = self.inner[i] / rhs;
+        Self::Output {
+            inner: array_init(|i| self[i] / rhs),
         }
-        ret
     }
 }
 
@@ -235,7 +223,7 @@ mod tests {
     #[test]
     fn test_default_new() {
         let a = Vector::<f32, 3>::default();
-        let b = Vector::<f32, 3>::new();
+        let b = Vector::<f32, 3>::from([1.0; 3]) * 0.0;
         let c = Vector::<f32, 3>::from([0.0; 3]);
         assert_eq!(a, b);
         assert_eq!(b, c);