From f8dbec0abbcab8058c2a0a8884195ea02a54d53f Mon Sep 17 00:00:00 2001 From: Anatol Belski Date: Thu, 26 Mar 2026 23:26:35 +0100 Subject: [PATCH] performance-metrics: Add QCOW2 async read micro benchmark Add micro_bench_qcow_async_read which reads clusters through the QcowDiskAsync io_uring backend. Single allocated cluster reads go through io_uring for true asynchronous completion, unlike the sync benchmarks which use QcowDiskSync with blocking I/O. Workloads: 128 and 256 clusters. Signed-off-by: Anatol Belski --- performance-metrics/src/main.rs | 26 +++++++++++++++++- performance-metrics/src/micro_bench_block.rs | 29 +++++++++++++++++++- 2 files changed, 53 insertions(+), 2 deletions(-) diff --git a/performance-metrics/src/main.rs b/performance-metrics/src/main.rs index 2c33bf9a8..1177b54f4 100644 --- a/performance-metrics/src/main.rs +++ b/performance-metrics/src/main.rs @@ -378,7 +378,7 @@ mod adjuster { } } -const TEST_LIST: [PerformanceTest; 82] = [ +const TEST_LIST: [PerformanceTest; 84] = [ PerformanceTest { name: "boot_time_ms", func_ptr: performance_boot_time, @@ -1493,6 +1493,30 @@ const TEST_LIST: [PerformanceTest; 82] = [ }, unit_adjuster: adjuster::s_to_us, }, + PerformanceTest { + name: "micro_block_qcow_async_read_128_us", + func_ptr: micro_bench_block::micro_bench_qcow_async_read, + control: PerformanceTestControl { + test_timeout: 10, + test_iterations: 20, + warmup_iterations: 5, + num_ops: Some(128), + ..PerformanceTestControl::default() + }, + unit_adjuster: adjuster::s_to_us, + }, + PerformanceTest { + name: "micro_block_qcow_async_read_256_us", + func_ptr: micro_bench_block::micro_bench_qcow_async_read, + control: PerformanceTestControl { + test_timeout: 10, + test_iterations: 20, + warmup_iterations: 5, + num_ops: Some(256), + ..PerformanceTestControl::default() + }, + unit_adjuster: adjuster::s_to_us, + }, ]; fn run_test_with_timeout( diff --git a/performance-metrics/src/micro_bench_block.rs b/performance-metrics/src/micro_bench_block.rs index 2b83ab495..5ab907a87 100644 --- a/performance-metrics/src/micro_bench_block.rs +++ b/performance-metrics/src/micro_bench_block.rs @@ -17,7 +17,8 @@ use block::raw_async_aio::RawFileAsyncAio; use crate::PerformanceTestControl; use crate::util::{ self, BLOCK_SIZE, L2_ENTRIES_PER_TABLE, QCOW_CLUSTER_SIZE, deterministic_permutation, - drain_completions, read_iovec, submit_reads, submit_writes, write_iovec, + drain_async_completions, drain_completions, read_iovec, submit_reads, submit_writes, + write_iovec, }; /// Submit num_ops AIO writes, wait for them all to land, then time @@ -328,3 +329,29 @@ pub fn micro_bench_qcow_l2_cache_miss(control: &PerformanceTestControl) -> f64 { elapsed } + +/// Read num_ops clusters from a prepopulated qcow2 image through the +/// QcowAsync io_uring path and time the total wall clock. +/// +/// Unlike micro_bench_qcow_read which uses QcowDiskSync (blocking), +/// this uses QcowDiskAsync where single-allocated-cluster reads go +/// through io_uring for true asynchronous completion. +/// +/// Returns the total read wall clock time in seconds. +pub fn micro_bench_qcow_async_read(control: &PerformanceTestControl) -> f64 { + let num_ops = control.num_ops.expect("num_ops required") as usize; + let (_tmp, disk) = util::qcow_async_tempfile(num_ops); + let mut async_io = disk + .new_async_io(num_ops as u32) + .expect("new_async_io failed"); + + let mut buf = vec![0u8; QCOW_CLUSTER_SIZE as usize]; + let iovec = read_iovec(&mut buf); + + let start = Instant::now(); + submit_reads(async_io.as_mut(), num_ops, QCOW_CLUSTER_SIZE, &[iovec]); + + // Drain all io_uring completions before stopping the clock. + drain_async_completions(async_io.as_mut(), num_ops); + start.elapsed().as_secs_f64() +}