storage_benchmarks/
io_benchmarks.rs

// Copyright 2022 The Fuchsia Authors. All rights reserved.
// Use of this source code is governed by a BSD-style license that can be
// found in the LICENSE file.

use crate::{Benchmark, CacheClearableFilesystem, Filesystem, OperationDuration, OperationTimer};
use async_trait::async_trait;
use rand::seq::SliceRandom;
use rand::{Rng, SeedableRng};
use rand_xorshift::XorShiftRng;
use std::fs::OpenOptions;
use std::io::{Seek, SeekFrom, Write};
use std::os::unix::fs::FileExt;
use std::os::unix::io::AsRawFd;

const RNG_SEED: u64 = 0xda782a0c3ce1819a;
/// How many blocks to skip after each used block in sparse read benchmarks. This is set to thwart
/// the effects of readahead on these benchmarks. As high as we can go while staying under the file
/// size limit of 4GiB in minfs.
const BLOCK_SKIP: usize = 255;

/// A benchmark that measures how long `read` calls take to a file that should not already be cached
/// in memory.
#[derive(Clone)]
pub struct ReadSequentialCold {
    op_size: usize,
    op_count: usize,
}

impl ReadSequentialCold {
    pub fn new(op_size: usize, op_count: usize) -> Self {
        Self { op_size, op_count }
    }
}

#[async_trait]
impl<T: CacheClearableFilesystem> Benchmark<T> for ReadSequentialCold {
    async fn run(&self, fs: &mut T) -> Vec<OperationDuration> {
        storage_trace::duration!(
            c"benchmark",
            c"ReadSequentialCold",
            "op_size" => self.op_size,
            "op_count" => self.op_count
        );
        let file_path = fs.benchmark_dir().join("file");

        // Setup
        let mut file = OpenOptions::new().write(true).create_new(true).open(&file_path).unwrap();
        write_file(&mut file, self.op_size, self.op_count);
        std::mem::drop(file);
        fs.clear_cache().await;

        // Benchmark
        let mut file = OpenOptions::new().read(true).open(&file_path).unwrap();
        read_sequential(&mut file, self.op_size, self.op_count)
    }

    fn name(&self) -> String {
        format!("ReadSequentialCold/{}", self.op_size)
    }
}

/// A benchmark that measures how long `read` calls take to a file that should already be cached in
/// memory.
#[derive(Clone)]
pub struct ReadSequentialWarm {
    op_size: usize,
    op_count: usize,
}

impl ReadSequentialWarm {
    pub fn new(op_size: usize, op_count: usize) -> Self {
        Self { op_size, op_count }
    }
}

#[async_trait]
impl<T: Filesystem> Benchmark<T> for ReadSequentialWarm {
    async fn run(&self, fs: &mut T) -> Vec<OperationDuration> {
        storage_trace::duration!(
            c"benchmark",
            c"ReadSequentialWarm",
            "op_size" => self.op_size,
            "op_count" => self.op_count
        );
        let file_path = fs.benchmark_dir().join("file");

        // Setup
        let mut file =
            OpenOptions::new().write(true).read(true).create_new(true).open(&file_path).unwrap();
        write_file(&mut file, self.op_size, self.op_count);
        file.seek(SeekFrom::Start(0)).unwrap();

        // Benchmark
        read_sequential(&mut file, self.op_size, self.op_count)
    }

    fn name(&self) -> String {
        format!("ReadSequentialWarm/{}", self.op_size)
    }
}

/// A benchmark that measures how long random `pread` calls take to a file that should not already
/// be cached in memory.
#[derive(Clone)]
pub struct ReadRandomCold {
    op_size: usize,
    op_count: usize,
}

impl ReadRandomCold {
    pub fn new(op_size: usize, op_count: usize) -> Self {
        Self { op_size, op_count }
    }
}

#[async_trait]
impl<T: CacheClearableFilesystem> Benchmark<T> for ReadRandomCold {
    async fn run(&self, fs: &mut T) -> Vec<OperationDuration> {
        storage_trace::duration!(
            c"benchmark",
            c"ReadRandomCold",
            "op_size" => self.op_size,
            "op_count" => self.op_count
        );
        let file_path = fs.benchmark_dir().join("file");

        // Setup
        let mut file = OpenOptions::new().write(true).create_new(true).open(&file_path).unwrap();
        write_file(&mut file, self.op_size, self.op_count);
        std::mem::drop(file);
        fs.clear_cache().await;

        // Benchmark
        let mut rng = XorShiftRng::seed_from_u64(RNG_SEED);
        let mut file = OpenOptions::new().read(true).open(&file_path).unwrap();
        read_random(&mut file, self.op_size, self.op_count, &mut rng)
    }

    fn name(&self) -> String {
        format!("ReadRandomCold/{}", self.op_size)
    }
}

/// A benchmark that measures how long `pread` calls take to a sparse file that should not already
/// be cached in memory. The file's sparseness is designed to defeat the gains of readahead without
/// making an extremely large file, but will also result in testing access to many small extents.
#[derive(Clone)]
pub struct ReadSparseCold {
    op_size: usize,
    op_count: usize,
}

impl ReadSparseCold {
    pub fn new(op_size: usize, op_count: usize) -> Self {
        Self { op_size, op_count }
    }
}

#[async_trait]
impl<T: CacheClearableFilesystem> Benchmark<T> for ReadSparseCold {
    async fn run(&self, fs: &mut T) -> Vec<OperationDuration> {
        storage_trace::duration!(
            c"benchmark",
            c"ReadSparseCold",
            "op_size" => self.op_size,
            "op_count" => self.op_count
        );
        let file_path = fs.benchmark_dir().join("file");

        // Setup
        let mut file = OpenOptions::new().write(true).create_new(true).open(&file_path).unwrap();
        write_sparse_file(&mut file, self.op_size, self.op_count, BLOCK_SKIP);
        std::mem::drop(file);
        fs.clear_cache().await;

        // Benchmark
        let mut file = OpenOptions::new().read(true).open(&file_path).unwrap();
        read_sparse(&mut file, self.op_size, self.op_count, BLOCK_SKIP)
    }

    fn name(&self) -> String {
        format!("ReadSparseCold/{}", self.op_size)
    }
}

/// A benchmark that measures how long random `pread` calls take to a file that should already be
/// cached in memory.
#[derive(Clone)]
pub struct ReadRandomWarm {
    op_size: usize,
    op_count: usize,
}

impl ReadRandomWarm {
    pub fn new(op_size: usize, op_count: usize) -> Self {
        Self { op_size, op_count }
    }
}

#[async_trait]
impl<T: Filesystem> Benchmark<T> for ReadRandomWarm {
    async fn run(&self, fs: &mut T) -> Vec<OperationDuration> {
        storage_trace::duration!(
            c"benchmark",
            c"ReadRandomWarm",
            "op_size" => self.op_size,
            "op_count" => self.op_count
        );
        let file_path = fs.benchmark_dir().join("file");

        // Setup
        let mut file =
            OpenOptions::new().write(true).read(true).create_new(true).open(&file_path).unwrap();
        write_file(&mut file, self.op_size, self.op_count);

        // Benchmark
        let mut rng = XorShiftRng::seed_from_u64(RNG_SEED);
        read_random(&mut file, self.op_size, self.op_count, &mut rng)
    }

    fn name(&self) -> String {
        format!("ReadRandomWarm/{}", self.op_size)
    }
}

/// A benchmark that measures how long `write` calls take to a new file.
#[derive(Clone)]
pub struct WriteSequentialCold {
    op_size: usize,
    op_count: usize,
}

impl WriteSequentialCold {
    pub fn new(op_size: usize, op_count: usize) -> Self {
        Self { op_size, op_count }
    }
}

#[async_trait]
impl<T: Filesystem> Benchmark<T> for WriteSequentialCold {
    async fn run(&self, fs: &mut T) -> Vec<OperationDuration> {
        storage_trace::duration!(
            c"benchmark",
            c"WriteSequentialCold",
            "op_size" => self.op_size,
            "op_count" => self.op_count
        );
        let file_path = fs.benchmark_dir().join("file");
        let mut file = OpenOptions::new().write(true).create_new(true).open(&file_path).unwrap();
        write_sequential(&mut file, self.op_size, self.op_count)
    }

    fn name(&self) -> String {
        format!("WriteSequentialCold/{}", self.op_size)
    }
}

/// A benchmark that measures how long `write` calls take when overwriting a file that should
/// already be in memory.
#[derive(Clone)]
pub struct WriteSequentialWarm {
    op_size: usize,
    op_count: usize,
}

impl WriteSequentialWarm {
    pub fn new(op_size: usize, op_count: usize) -> Self {
        Self { op_size, op_count }
    }
}

#[async_trait]
impl<T: Filesystem> Benchmark<T> for WriteSequentialWarm {
    async fn run(&self, fs: &mut T) -> Vec<OperationDuration> {
        storage_trace::duration!(
            c"benchmark",
            c"WriteSequentialWarm",
            "op_size" => self.op_size,
            "op_count" => self.op_count
        );
        let file_path = fs.benchmark_dir().join("file");

        // Setup
        let mut file = OpenOptions::new().write(true).create_new(true).open(&file_path).unwrap();
        write_file(&mut file, self.op_size, self.op_count);
        file.seek(SeekFrom::Start(0)).unwrap();

        // Benchmark
        write_sequential(&mut file, self.op_size, self.op_count)
    }

    fn name(&self) -> String {
        format!("WriteSequentialWarm/{}", self.op_size)
    }
}

/// A benchmark that measures how long random `pwrite` calls take to a new file.
#[derive(Clone)]
pub struct WriteRandomCold {
    op_size: usize,
    op_count: usize,
}

impl WriteRandomCold {
    pub fn new(op_size: usize, op_count: usize) -> Self {
        Self { op_size, op_count }
    }
}

#[async_trait]
impl<T: Filesystem> Benchmark<T> for WriteRandomCold {
    async fn run(&self, fs: &mut T) -> Vec<OperationDuration> {
        storage_trace::duration!(
            c"benchmark",
            c"WriteRandomCold",
            "op_size" => self.op_size,
            "op_count" => self.op_count
        );
        let file_path = fs.benchmark_dir().join("file");
        let mut file = OpenOptions::new().write(true).create_new(true).open(&file_path).unwrap();
        let mut rng = XorShiftRng::seed_from_u64(RNG_SEED);
        write_random(&mut file, self.op_size, self.op_count, &mut rng)
    }

    fn name(&self) -> String {
        format!("WriteRandomCold/{}", self.op_size)
    }
}

/// A benchmark that measures how long random `pwrite` calls take when overwriting a file that
/// should already be in memory.
#[derive(Clone)]
pub struct WriteRandomWarm {
    op_size: usize,
    op_count: usize,
}

impl WriteRandomWarm {
    pub fn new(op_size: usize, op_count: usize) -> Self {
        Self { op_size, op_count }
    }
}

#[async_trait]
impl<T: Filesystem> Benchmark<T> for WriteRandomWarm {
    async fn run(&self, fs: &mut T) -> Vec<OperationDuration> {
        storage_trace::duration!(
            c"benchmark",
            c"WriteRandomWarm",
            "op_size" => self.op_size,
            "op_count" => self.op_count
        );
        let file_path = fs.benchmark_dir().join("file");

        // Setup
        let mut file = OpenOptions::new().write(true).create_new(true).open(&file_path).unwrap();
        write_sequential(&mut file, self.op_size, self.op_count);

        // Benchmark
        let mut rng = XorShiftRng::seed_from_u64(RNG_SEED);
        write_random(&mut file, self.op_size, self.op_count, &mut rng)
    }

    fn name(&self) -> String {
        format!("WriteRandomWarm/{}", self.op_size)
    }
}

/// A benchmark that measures how long 'write` and `fsync` calls take to a new file.
#[derive(Clone)]
pub struct WriteSequentialFsyncCold {
    op_size: usize,
    op_count: usize,
}

impl WriteSequentialFsyncCold {
    pub fn new(op_size: usize, op_count: usize) -> Self {
        Self { op_size, op_count }
    }
}

#[async_trait]
impl<T: Filesystem> Benchmark<T> for WriteSequentialFsyncCold {
    async fn run(&self, fs: &mut T) -> Vec<OperationDuration> {
        storage_trace::duration!(
            c"benchmark",
            c"WriteSequentialFsyncCold",
            "op_size" => self.op_size,
            "op_count" => self.op_count
        );
        let file_path = fs.benchmark_dir().join("file");
        let mut file = OpenOptions::new().write(true).create_new(true).open(&file_path).unwrap();
        write_sequential_fsync(&mut file, self.op_size, self.op_count)
    }

    fn name(&self) -> String {
        format!("WriteSequentialFsyncCold/{}", self.op_size)
    }
}

/// A benchmark that measures how long `write` and `fsync` calls take when overwriting a
/// file that should already be in memory.
#[derive(Clone)]
pub struct WriteSequentialFsyncWarm {
    op_size: usize,
    op_count: usize,
}

impl WriteSequentialFsyncWarm {
    pub fn new(op_size: usize, op_count: usize) -> Self {
        Self { op_size, op_count }
    }
}

#[async_trait]
impl<T: Filesystem> Benchmark<T> for WriteSequentialFsyncWarm {
    async fn run(&self, fs: &mut T) -> Vec<OperationDuration> {
        storage_trace::duration!(
            c"benchmark",
            c"WriteSequentialFsyncWarm",
            "op_size" => self.op_size,
            "op_count" => self.op_count
        );
        let file_path = fs.benchmark_dir().join("file");

        // Setup
        let mut file = OpenOptions::new().write(true).create_new(true).open(&file_path).unwrap();
        write_sequential(&mut file, self.op_size, self.op_count);
        // To measure the exact performance of fsync, the previous written data must be synchronized.
        assert_eq!(unsafe { libc::fsync(file.as_raw_fd()) }, 0);
        file.seek(SeekFrom::Start(0)).unwrap();

        // Benchmark
        write_sequential_fsync(&mut file, self.op_size, self.op_count)
    }

    fn name(&self) -> String {
        format!("WriteSequentialFsyncWarm/{}", self.op_size)
    }
}

fn write_file<F: Write + FileExt>(file: &mut F, op_size: usize, op_count: usize) {
    write_sparse_file(file, op_size, op_count, 0);
}

fn write_sparse_file<F: Write + FileExt>(
    file: &mut F,
    op_size: usize,
    op_count: usize,
    block_skip: usize,
) {
    let data = vec![0xAB; op_size];
    let mut offset: u64 = 0;
    for _ in 0..op_count {
        assert_eq!(file.write_at(&data, offset).unwrap(), op_size);
        offset += (op_size * (block_skip + 1)) as u64;
    }
}

/// Makes `op_count` `read` calls to `file`, each for `op_size` bytes.
fn read_sequential<F: AsRawFd>(
    file: &mut F,
    op_size: usize,
    op_count: usize,
) -> Vec<OperationDuration> {
    let mut data = vec![0; op_size];
    let mut durations = Vec::new();
    let fd = file.as_raw_fd();
    for i in 0..op_count {
        storage_trace::duration!(c"benchmark", c"read", "op_number" => i);
        let timer = OperationTimer::start();
        let result = unsafe { libc::read(fd, data.as_mut_ptr() as *mut libc::c_void, data.len()) };
        durations.push(timer.stop());
        assert_eq!(result, op_size as isize);
    }
    durations
}

/// Makes `op_count` `pread` calls to `file`, each for `op_size` bytes `block_skip` * `op_size`
/// bytes apart.
fn read_sparse<F: AsRawFd>(
    file: &mut F,
    op_size: usize,
    op_count: usize,
    block_skip: usize,
) -> Vec<OperationDuration> {
    let mut data = vec![0; op_size];
    let mut durations = Vec::new();
    let fd = file.as_raw_fd();
    let sparse_offset = ((1 + block_skip) * op_size) as i64;
    for i in 0..op_count as i64 {
        storage_trace::duration!(c"benchmark", c"pread", "op_number" => i);
        let timer = OperationTimer::start();
        let result = unsafe {
            libc::pread(fd, data.as_mut_ptr() as *mut libc::c_void, data.len(), i * sparse_offset)
        };
        durations.push(timer.stop());
        assert_eq!(result, op_size as isize);
    }
    durations
}

/// Makes `op_count` `write` calls to `file`, each containing `op_size` bytes.
fn write_sequential<F: AsRawFd>(
    file: &mut F,
    op_size: usize,
    op_count: usize,
) -> Vec<OperationDuration> {
    let data = vec![0xAB; op_size];
    let mut durations = Vec::new();
    let fd = file.as_raw_fd();
    for i in 0..op_count {
        storage_trace::duration!(c"benchmark", c"write", "op_number" => i);
        let timer = OperationTimer::start();
        let result = unsafe { libc::write(fd, data.as_ptr() as *const libc::c_void, data.len()) };
        durations.push(timer.stop());
        assert_eq!(result, op_size as isize);
    }
    durations
}

/// Makes `op_count` `write` calls to `file`, each containing `op_size` bytes.
/// After write, call `fsync` and sync with disk(non-volatile medium).
fn write_sequential_fsync<F: AsRawFd>(
    file: &mut F,
    op_size: usize,
    op_count: usize,
) -> Vec<OperationDuration> {
    let data = vec![0xAB; op_size];
    let mut durations = Vec::new();
    let fd = file.as_raw_fd();
    for i in 0..op_count {
        storage_trace::duration!(c"benchmark", c"write", "op_number" => i);
        let timer = OperationTimer::start();
        let write_result =
            unsafe { libc::write(fd, data.as_ptr() as *const libc::c_void, data.len()) };
        let fsync_result = unsafe { libc::fsync(fd) };
        durations.push(timer.stop());
        assert_eq!(write_result, op_size as isize);
        assert_eq!(fsync_result, 0);
    }
    durations
}

fn create_random_offsets<R: Rng>(op_size: usize, op_count: usize, rng: &mut R) -> Vec<libc::off_t> {
    let op_count = op_count as libc::off_t;
    let op_size = op_size as libc::off_t;
    let mut offsets: Vec<libc::off_t> = (0..op_count).map(|offset| offset * op_size).collect();
    offsets.shuffle(rng);
    offsets
}

/// Reads the first `op_size * op_count` bytes in `file` by making `op_count` `pread` calls, each
/// `pread` call reads `op_size` bytes. The offset order for the `pread` calls is randomized using
/// `rng`.
fn read_random<F: AsRawFd, R: Rng>(
    file: &mut F,
    op_size: usize,
    op_count: usize,
    rng: &mut R,
) -> Vec<OperationDuration> {
    let offsets = create_random_offsets(op_size, op_count, rng);
    let mut data = vec![0xAB; op_size];
    let mut durations = Vec::new();
    let fd = file.as_raw_fd();
    for (i, offset) in offsets.iter().enumerate() {
        storage_trace::duration!(c"benchmark", c"pread", "op_number" => i, "offset" => *offset);
        let timer = OperationTimer::start();
        let result =
            unsafe { libc::pread(fd, data.as_mut_ptr() as *mut libc::c_void, data.len(), *offset) };
        durations.push(timer.stop());
        assert_eq!(result, op_size as isize);
    }
    durations
}

/// Overwrites the first `op_size * op_count` bytes in `file` by making `op_count` `pwrite` calls,
/// each `pwrite` call writes `op_size` bytes. The offset order for the `pwrite` calls is
/// randomized using `rng`.
fn write_random<F: AsRawFd, R: Rng>(
    file: &mut F,
    op_size: usize,
    op_count: usize,
    rng: &mut R,
) -> Vec<OperationDuration> {
    let offsets = create_random_offsets(op_size, op_count, rng);
    let data = vec![0xAB; op_size];
    let mut durations = Vec::new();
    let fd = file.as_raw_fd();
    for (i, offset) in offsets.iter().enumerate() {
        storage_trace::duration!(c"benchmark", c"pwrite", "op_number" => i, "offset" => *offset);
        let timer = OperationTimer::start();
        let result =
            unsafe { libc::pwrite(fd, data.as_ptr() as *const libc::c_void, data.len(), *offset) };
        durations.push(timer.stop());
        assert_eq!(result, op_size as isize);
    }
    durations
}

#[cfg(test)]
mod tests {
    use super::*;
    use crate::testing::TestFilesystem;

    const OP_SIZE: usize = 8;
    const OP_COUNT: usize = 2;

    async fn check_benchmark<T>(benchmark: T, op_count: usize, clear_cache_count: u64)
    where
        T: Benchmark<TestFilesystem>,
    {
        let mut test_fs = Box::new(TestFilesystem::new());
        let results = benchmark.run(test_fs.as_mut()).await;

        assert_eq!(results.len(), op_count);
        assert_eq!(test_fs.clear_cache_count().await, clear_cache_count);
        test_fs.shutdown().await;
    }

    #[fuchsia::test]
    async fn read_sequential_cold_test() {
        check_benchmark(
            ReadSequentialCold::new(OP_SIZE, OP_COUNT),
            OP_COUNT,
            /*clear_cache_count=*/ 1,
        )
        .await;
    }

    #[fuchsia::test]
    async fn read_sequential_warm_test() {
        check_benchmark(
            ReadSequentialWarm::new(OP_SIZE, OP_COUNT),
            OP_COUNT,
            /*clear_cache_count=*/ 0,
        )
        .await;
    }

    #[fuchsia::test]
    async fn read_random_cold_test() {
        check_benchmark(
            ReadRandomCold::new(OP_SIZE, OP_COUNT),
            OP_COUNT,
            /*clear_cache_count=*/ 1,
        )
        .await;
    }

    #[fuchsia::test]
    async fn read_sparse_cold_test() {
        check_benchmark(
            ReadSparseCold::new(OP_SIZE, OP_COUNT),
            OP_COUNT,
            /*clear_cache_count=*/ 1,
        )
        .await;
    }

    #[fuchsia::test]
    async fn read_random_warm_test() {
        check_benchmark(
            ReadRandomWarm::new(OP_SIZE, OP_COUNT),
            OP_COUNT,
            /*clear_cache_count=*/ 0,
        )
        .await;
    }

    #[fuchsia::test]
    async fn write_sequential_cold_test() {
        check_benchmark(
            WriteSequentialCold::new(OP_SIZE, OP_COUNT),
            OP_COUNT,
            /*clear_cache_count=*/ 0,
        )
        .await;
    }

    #[fuchsia::test]
    async fn write_sequential_warm_test() {
        check_benchmark(
            WriteSequentialWarm::new(OP_SIZE, OP_COUNT),
            OP_COUNT,
            /*clear_cache_count=*/ 0,
        )
        .await;
    }

    #[fuchsia::test]
    async fn write_random_cold_test() {
        check_benchmark(
            WriteRandomCold::new(OP_SIZE, OP_COUNT),
            OP_COUNT,
            /*clear_cache_count=*/ 0,
        )
        .await;
    }

    #[fuchsia::test]
    async fn write_random_warm_test() {
        check_benchmark(
            WriteRandomWarm::new(OP_SIZE, OP_COUNT),
            OP_COUNT,
            /*clear_cache_count=*/ 0,
        )
        .await;
    }

    #[fuchsia::test]
    async fn write_sequential_fsync_cold_test() {
        check_benchmark(
            WriteSequentialFsyncCold::new(OP_SIZE, OP_COUNT),
            OP_COUNT,
            /*clear_cache_count=*/ 0,
        )
        .await;
    }

    #[fuchsia::test]
    async fn write_sequential_fsync_warm_test() {
        check_benchmark(
            WriteSequentialFsyncWarm::new(OP_SIZE, OP_COUNT),
            OP_COUNT,
            /*clear_cache_count=*/ 0,
        )
        .await;
    }
}
storage_benchmarks/io_benchmarks.rs

storage_benchmarks/
io_benchmarks.rs
