Skip to content
New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Jump to bottom

Reseeding perf #76

Merged
merged 4 commits into from
Jan 1, 2018
Merged

Reseeding perf #76

Show file tree
Hide file tree
Changes from 2 commits
Commits
Show all changes
4 commits
Select commit Hold shift + click to select a range
402e673
Add benchmarks for `ReseedingRng`
pitdicker Dec 16, 2017
4d10328
Improve performance of `ReseedingRng`
pitdicker Dec 16, 2017
07717bc
Simplify reseeding erro logic
pitdicker Dec 31, 2017
14f02a2
Benchmark reseeding HC-128 instead of Xorshift
pitdicker Dec 31, 2017
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
38 changes: 38 additions & 0 deletions benches/generators.rs
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -11,6 +11,7 @@ use test::{black_box, Bencher};

use rand::{Rng, NewSeeded, Sample, SeedableRng, StdRng, OsRng, JitterRng};
use rand::prng::*;
use rand::reseeding::{ReseedingRng, ReseedWithNew};

macro_rules! gen_bytes {
($fnn:ident, $gen:ident) => {
Expand Down Expand Up @@ -102,3 +103,40 @@ fn init_jitter(b: &mut Bencher) {
black_box(JitterRng::new().unwrap());
});
}



#[bench]
fn reseeding_xorshift_bytes(b: &mut Bencher) {
let mut rng = ReseedingRng::new(XorShiftRng::new().unwrap(),
Copy link
Owner

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Wouldn't it make more sense to benchmark the PRNG we're interested in (ISAAC or HC128)? Especially given that your BlockRng idea integrates tighter with the PRNG algorithm.

Copy link
Author

@pitdicker pitdicker Dec 31, 2017 •
edited
Loading

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

With HC-128 the overhead of reseeding is much larger:

test gen_bytes_hc128       ... bench:     445,483 ns/iter (+/- 16,674) = 2298 MB/s
test gen_u32_hc128         ... bench:       2,809 ns/iter (+/- 194) = 1423 MB/s
test gen_u64_hc128         ... bench:       4,254 ns/iter (+/- 360) = 1880 MB/s
test init_hc128            ... bench:       4,539 ns/iter (+/- 412)
test reseeding_hc128_bytes ... bench:     451,584 ns/iter (+/- 25,463) = 2267 MB/s
test reseeding_hc128_u32   ... bench:       3,690 ns/iter (+/- 125) = 1084 MB/s
test reseeding_hc128_u64   ... bench:       5,907 ns/iter (+/- 157) = 1354 MB/s

And before this PR:

test reseeding_hc128_bytes ... bench:     449,635 ns/iter (+/- 5,755) = 2277 MB/s
test reseeding_hc128_u32   ... bench:       6,418 ns/iter (+/- 84) = 623 MB/s
test reseeding_hc128_u64   ... bench:       7,693 ns/iter (+/- 124) = 1039 MB/s

This makes sense, because the overhead of checking and indexing in the results array makes up 20~40%. With ReseedingRng that percentage gets doubled because it also does checks for reseeding.

You are right, this benchmark is testing something nonsensical. Reseeding Xorshift ?!. I don't think it matters much because they both show the overhead of ReseedingRng, but I'll change it.

128*1024*1024*1024,
ReseedWithNew);
let mut buf = [0u8; BYTES_LEN];
b.iter(|| {
for _ in 0..RAND_BENCH_N {
rng.fill_bytes(&mut buf);
black_box(buf);
}
});
b.bytes = BYTES_LEN as u64 * RAND_BENCH_N;
}

macro_rules! reseeding_uint {
($fnn:ident, $ty:ty) => {
#[bench]
fn $fnn(b: &mut Bencher) {
let mut rng = ReseedingRng::new(XorShiftRng::new().unwrap(),
128*1024*1024*1024,
ReseedWithNew);
b.iter(|| {
for _ in 0..RAND_BENCH_N {
black_box(rng.gen::<$ty>());
}
});
b.bytes = size_of::<$ty>() as u64 * RAND_BENCH_N;
}
}
}

reseeding_uint!(reseeding_xorshift_u32, u32);
reseeding_uint!(reseeding_xorshift_u64, u64);
131 changes: 71 additions & 60 deletions src/reseeding.rs
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -11,14 +11,13 @@
//! A wrapper around another RNG that reseeds it after it
//! generates a certain number of random bytes.

use core::cmp::max;
use {Rng, SeedableRng, Error, ErrorKind};
#[cfg(feature="std")]
use NewSeeded;

/// How many bytes of entropy the underling RNG is allowed to generate
/// before it is reseeded
const DEFAULT_GENERATION_THRESHOLD: u64 = 32 * 1024;
const DEFAULT_RESEEDING_THRESHOLD: i64 = 32 * 1024;

/// A wrapper around any RNG which reseeds the underlying RNG after it
/// has generated a certain number of random bytes.
Expand All @@ -32,8 +31,8 @@ const DEFAULT_GENERATION_THRESHOLD: u64 = 32 * 1024;
#[derive(Debug, Clone)]
pub struct ReseedingRng<R, Rsdr: Reseeder<R>> {
rng: R,
generation_threshold: u64,
bytes_generated: u64,
threshold: i64,
bytes_until_reseed: i64,
/// Controls the behaviour when reseeding the RNG.
pub reseeder: Rsdr,
}
Expand All @@ -44,13 +43,14 @@ impl<R: Rng, Rsdr: Reseeder<R>> ReseedingRng<R, Rsdr> {
/// # Arguments
///
/// * `rng`: the random number generator to use.
/// * `generation_threshold`: the number of bytes of entropy at which to reseed the RNG.
/// * `threshold`: the amount of generated bytes after which to reseed the RNG.
Copy link
Owner

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

I believe "the number of [generated] bytes" is correct English; amount is typically used for "uncountable" things (e.g. water, money, food). But I'm not really fussed (I know I accepted something similar recently anyway).

Copy link
Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Please keep correcting my English, It is not my first language and it is better if the language in the documentation is correct.

/// * `reseeder`: the reseeding object to use.
pub fn new(rng: R, generation_threshold: u64, reseeder: Rsdr) -> ReseedingRng<R,Rsdr> {
pub fn new(rng: R, threshold: u64, reseeder: Rsdr) -> ReseedingRng<R,Rsdr> {
assert!(threshold <= ::core::i64::MAX as u64);
ReseedingRng {
rng: rng,
generation_threshold: generation_threshold,
bytes_generated: 0,
threshold: threshold as i64,
bytes_until_reseed: threshold as i64,
reseeder: reseeder
}
}
Expand All @@ -59,101 +59,112 @@ impl<R: Rng, Rsdr: Reseeder<R>> ReseedingRng<R, Rsdr> {
/// generated exceed the threshold.
///
/// On error, this may delay reseeding or not reseed at all.
pub fn reseed_if_necessary(&mut self) {
if self.bytes_generated >= self.generation_threshold {
let mut err_count = 0;
loop {
if let Err(e) = self.reseeder.reseed(&mut self.rng) {
// TODO: log?
if e.kind.should_wait() {
// Delay reseeding
let delay = max(self.generation_threshold >> 8, self.bytes_generated);
self.bytes_generated -= delay;
break;
} else if e.kind.should_retry() {
if err_count > 4 { // arbitrary limit
// TODO: log details & cause?
break; // give up trying to reseed
}
err_count += 1;
continue; // immediate retry
} else {
#[inline(never)]
pub fn reseed(&mut self) {
let mut err_count = 0;
loop {
if let Err(e) = self.reseeder.reseed(&mut self.rng) {
// TODO: log?
if e.kind.should_wait() {
// Delay reseeding
self.bytes_until_reseed = self.threshold >> 8;
break;
Copy link
Owner

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

This sets bytes_until_reseed twice. Should it return instead?

I admit that the logic of this function is weird but I guess something like this is useful.

Copy link
Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Hee, you're right.

} else if e.kind.should_retry() {
if err_count > 4 { // arbitrary limit
// TODO: log details & cause?
break; // give up trying to reseed
}
err_count += 1;
continue; // immediate retry
} else {
break; // no reseeding
break; // give up trying to reseed
}
} else {
break; // no reseeding
}
self.bytes_generated = 0;
}
self.bytes_until_reseed = self.threshold;
}
/// Reseed the internal RNG if the number of bytes that have been
/// generated exceed the threshold.
///
/// If reseeding fails, return an error with the original cause. Note that
/// if the cause has a permanent failure, we report a transient error and
/// skip reseeding.
pub fn try_reseed_if_necessary(&mut self) -> Result<(), Error> {
if self.bytes_generated >= self.generation_threshold {
if let Err(err) = self.reseeder.reseed(&mut self.rng) {
let newkind = match err.kind {
a @ ErrorKind::NotReady => a,
b @ ErrorKind::Transient => b,
_ => {
self.bytes_generated = 0; // skip reseeding
ErrorKind::Transient
}
};
return Err(Error::with_cause(newkind, "reseeding failed", err));
}
self.bytes_generated = 0;
#[inline(never)]
pub fn try_reseed(&mut self) -> Result<(), Error> {
if let Err(err) = self.reseeder.reseed(&mut self.rng) {
let newkind = match err.kind {
a @ ErrorKind::NotReady => a,
b @ ErrorKind::Transient => b,
_ => {
self.bytes_until_reseed = self.threshold; // skip reseeding
ErrorKind::Transient
}
};
return Err(Error::with_cause(newkind, "reseeding failed", err));
}
self.bytes_until_reseed = self.threshold;
Ok(())
}
}


impl<R: Rng, Rsdr: Reseeder<R>> Rng for ReseedingRng<R, Rsdr> {
fn next_u32(&mut self) -> u32 {
self.reseed_if_necessary();
self.bytes_generated += 4;
self.rng.next_u32()
let value = self.rng.next_u32();
self.bytes_until_reseed -= 4;
if self.bytes_until_reseed <= 0 {
self.reseed();
}
value
}

fn next_u64(&mut self) -> u64 {
self.reseed_if_necessary();
self.bytes_generated += 8;
self.rng.next_u64()
let value = self.rng.next_u64();
self.bytes_until_reseed -= 8;
if self.bytes_until_reseed <= 0 {
self.reseed();
}
value
}

#[cfg(feature = "i128_support")]
fn next_u128(&mut self) -> u128 {
self.reseed_if_necessary();
self.bytes_generated += 16;
self.rng.next_u128()
let value = self.rng.next_u128();
self.bytes_until_reseed -= 16;
if self.bytes_until_reseed <= 0 {
self.reseed();
}
value
}

fn fill_bytes(&mut self, dest: &mut [u8]) {
self.reseed_if_necessary();
self.bytes_generated += dest.len() as u64;
self.rng.fill_bytes(dest);
self.bytes_until_reseed -= dest.len() as i64;
if self.bytes_until_reseed <= 0 {
self.reseed();
}
}

fn try_fill(&mut self, dest: &mut [u8]) -> Result<(), Error> {
self.try_reseed_if_necessary()?;
self.bytes_generated += dest.len() as u64;
self.rng.try_fill(dest)
self.rng.try_fill(dest)?;
self.bytes_until_reseed -= dest.len() as i64;
if self.bytes_until_reseed <= 0 {
self.try_reseed()?;
}
Ok(())
}
}

impl<R: SeedableRng, Rsdr: Reseeder<R>> ReseedingRng<R, Rsdr> {
/// Create a new `ReseedingRng` from the given reseeder and
/// seed. This uses a default value for `generation_threshold`.
/// seed. This uses a default value for `threshold`.
pub fn from_reseeder(rsdr: Rsdr, seed: <R as SeedableRng>::Seed) -> ReseedingRng<R, Rsdr> {
ReseedingRng {
rng: SeedableRng::from_seed(seed),
generation_threshold: DEFAULT_GENERATION_THRESHOLD,
bytes_generated: 0,
threshold: DEFAULT_RESEEDING_THRESHOLD,
bytes_until_reseed: DEFAULT_RESEEDING_THRESHOLD,
reseeder: rsdr
}
}
Expand Down