Zero-allocation steady state for I/O object and operation churn by sgerbino · Pull Request #389 · cppalliance/corosio

sgerbino · 2026-10-08T16:51:25Z

Replaces per-impl shared_ptr ownership with an intrusive refcount
(object_ref) and a per-service recycling pool (object_pool), then
extends recycling to the remaining per-operation allocations. Once
warmed, construct/destroy, accept, positional file I/O, timer waits,
and tcp_server connection dispatch allocate nothing.

Changes

object_pool<Impl> / object_ref: live/free intrusive lists behind
one mutex, variadic acquire() as the library's only new site for
impls, retire() at the zero-crossing recycles instead of deleting.
All backends converted; IOCP's wrapper/internal split collapsed;
timer unified (keeping its thread-local slot).
uring multishot accept: waiter_node + per-completion op merged into
one pooled accept_node (2 heap nodes per accept -> 0).
Random-access file ops recycle through per-file free lists on posix,
uring, and IOCP (~500 B malloc/free per read/write -> 0).
tcp_server launch: reused worker stop_source, pooled launch
frame, post via frame-embedded continuation.
Fix (pre-existing, found by new tests): completions handed a
deferring executor the op's embedded continuation, which its queue
holds by reference while the op is recycled/reused/freed --
use-after-free or corrupted resume on every backend.
dispatch_coro now lends a self-destroying trampoline frame
instead; the io_context fast path is unchanged. Includes
deterministic multi-executor regression tests.
Zero-alloc gates probe that the counting interposer is live and
loud-skip otherwise (Valgrind replaces the allocator at run time).
Fault harness: chain interposed syscalls through sanitizer
__interceptor_ exports so ASan/TSan thread bookkeeping survives
interception (clang >= 23 link-order change).

Verification

Zero-alloc gates (allocation-counting interposer) pin exact counts:
accept/echo/close, impl reuse, delay(), raf write+read, and
tcp_server churn at 0/cycle on epoll and uring (select keeps its
documented 2/cycle descriptor-map cost).
A/B benchmarks: epoll accept_churn +4.4-8.2% across all rows,
uring +1.2-2.2% measured before the accept-node pooling that removes
its remaining per-accept allocations.
Adversarial review of the pool core and the recycling protocols;
full CI matrix green; ASan/TSan clean.

Resolves #381.

cppalliance-bot · 2026-10-08T16:54:30Z

An automated preview of the documentation is available at https://389.corosio.prtest3.cppalliance.org/index.html

If more commits are pushed to the pull request, the docs will rebuild at the same URL.

2026-10-08 18:09:23 UTC

cppalliance-bot · 2026-10-08T17:06:36Z

GCOVR code coverage report https://389.corosio.prtest3.cppalliance.org/gcovr/index.html
LCOV code coverage report https://389.corosio.prtest3.cppalliance.org/genhtml/index.html
Coverage Diff Report https://389.corosio.prtest3.cppalliance.org/diff-report/index.html

Build time: 2026-10-08 18:21:20 UTC

The multishot acceptor allocated a waiter_node when parking an accept and a fresh uring_accept_op at every completion boundary, copying the request fields across and deleting the waiter -- two heap round trips per accepted connection, plus a ready_fd_node whenever a delivery arrived before an accept was parked. A parked waiter and its completion have strictly sequential lifetimes, so fold them into one accept_node that is both an intrusive list node and the posted op. Nodes recycle through a per-acceptor free list via the op's new dispose hook, ready-fd nodes recycle the same way, and each in-flight node keeps the acceptor alive through the op envelope's object_ref slot. The inherited cancelled flag keeps its claim-token role; claiming paths now restore the real completion status after stop_cb.reset() quiesces any in-flight canceller. The uring zero-alloc gate pins accept churn at 0 allocations per cycle, matching epoll; only select retains its documented 2/cycle registered_descs_ cost.

Every read_some_at/write_some_at heap-allocated its op -- a deliberate choice so concurrent operations at distinct offsets can overlap, but one that cost ~500 bytes of malloc/free per call on the posix, io_uring, and IOCP backends alike. Keep the unlimited-concurrency model and recycle the ops instead: completion pushes the op onto a per-file free list (posix reuses the outstanding-ops hook and mutex; uring wraps the heap op types in a pooled subclass that swaps the completion handler; IOCP recycles raf_concurrent_op under its existing ops_mutex_). Each completion copies out what it needs before the push, since a concurrent initiation may refill the op immediately, and drops its impl keepalive only after leaving the list mutex. The storage belongs to the impl, survives impl recycling, and is freed by the impl destructor. The zero-alloc gate grows a raf churn test pinning a warmed write+read round trip at 0 allocations on every backend.

Launching a worker for an accepted connection paid three heap round trips: a fresh std::stop_source shared state, a launch_wrapper coroutine frame through the global allocator, and a wrapper op for posting the bare coroutine handle. Reuse the worker's stop_source across connections and replace it only once a stop has actually been delivered; route the launch_wrapper frame through capy's frame_alloc_mixin so the thread-local recycling allocator serves it; and post the launch through a frame-embedded capy::continuation, which the Executor concept already guarantees every executor accepts. The zero-alloc gate grows a server churn test pinning a warmed connect/echo/close round trip through tcp_server at 0 allocations (select keeps its documented 2/cycle registered_descs_ cost).

Since clang 23's static ASan runtime, the fault harness's interposed symbols bind ahead of the sanitizer's interceptors (verbosity=1 reports "failed to intercept 'pthread_create'"), so threads created through the raw libc pthread_create bypass ASan's thread registry and the first join aborts with "Joining already joined thread" -- every fault test failed instantly under local ASan builds. Sanitizers export each interceptor as __interceptor_<name> precisely so an interposer can cooperate. Under a sanitizer build, resolve the census's libc half through those exports (weak references, so a name the runtime does not intercept falls back to the RTLD_NEXT lookup unchanged). Injection semantics are untouched: shadows still run first and armed faults short-circuit before the chained call; unsanitized builds compile the mechanism away entirely.

A completion whose awaiter runs on a non-io_context executor takes dispatch_coro's deferring branch, which handed the executor the op's EMBEDDED continuation -- held on its queue by reference until dequeue (a documented stable-address contract). Every completion tail then consumes its op: recycling it into a free list (pooled uring accept/raf nodes), reusing the embedded slot on the next operation (reactor and IOCP ops), or deleting it (heap ops). A concurrent operation could therefore rewrite or free a queued node, and the dequeued resume targeted a dead or rewritten handle. Two new tests reproduce it deterministically (latch-sequenced, no timing): park a single-worker pool behind a latch so the deferred completion queues into it, refill the op from the io_context side, then release the worker. They crashed on every backend -- uring (pooled nodes), epoll/select (embedded accept op via coro_resume), and IOCP in CI (embedded accept op) -- so the fix lives at the one chokepoint every path shares: dispatch_coro's deferring branch now lends the executor a continuation owned by a self-destroying trampoline frame (post_resume; frames served by capy's recycling pool) and no longer retains the caller's. The io_context fast path is unchanged, and callers keep the uniform consume-then-resume shape since nothing they own stays queued. run_async could not serve as the trampoline: its launch wrapper calls ex.on_work_finished() after the task completes, but the op's executor_ref points into the awaiting coroutine's environment, which the resume destroys -- completions must never touch the executor after resuming the handle, and the trampoline doesn't.

Each reactor wake constructed to_add/to_cancel inside the loop, swapped away pending_register_/pending_cancel_'s capacity, and freed it at scope end -- two malloc/free per wake (pollfds beside them was already hoisted). Hoist both and clear per iteration: the capacities then ping-pong between the locals and the members through the under-lock swap, and a warmed reactor wake allocates nothing.

Vale's docstring lint reads the wrapped line's leading identifier as bare prose; rewrapping keeps the reference intact.

github-actions · 2026-10-08T18:02:42Z

Benchmark report

ff2966547b00 (base) vs 23880ff31a65 (head)

✅ linux — clean (epoll, uring)
⚠️ windows — 2 flagged (iocp)
⚠️ macos — 2 flagged (kqueue)

⚠️ Flagged

Platform	Backend	Benchmark	Metric	Base	Head	Δ	Noise
windows	iocp	accept_churn/burst/100	ops_per_sec	149.1 ops/s	155.6 ops/s	🟢▲ +4.40%	0.93%
windows	iocp	accept_churn/burst_lockless/100	ops_per_sec	152.4 ops/s	158.5 ops/s	🟢▲ +3.85%	0.64%
macos	kqueue	accept_churn/burst/100	ops_per_sec	431.9 ops/s	443.3 ops/s	🟢▲ +3.36%	0.81%
macos	kqueue	accept_churn/concurrent/1	ops_per_sec	19.58K ops/s	20.18K ops/s	🟢▲ +3.49%	1.03%

linux — full results (226 benchmarks, 5 iterations, 1.5s each)

accept_churn

Benchmark	Backend	Metric	Base	Head	Δ	Noise
burst/10	epoll	ops_per_sec	1.88K ops/s	1.91K ops/s	🟢▲ +2.01%	1.01%
burst/10	uring	ops_per_sec	1.71K ops/s	1.73K ops/s	🟢▲ +1.32%	0.88%
burst/100	epoll	ops_per_sec	189.3 ops/s	192.6 ops/s	🟢▲ +2.10%	1.11%
burst/100	uring	ops_per_sec	174.3 ops/s	177.2 ops/s	🟢▲ +1.89%	0.57%
burst_lockless/10	epoll	ops_per_sec	1.90K ops/s	1.93K ops/s	🟢▲ +2.09%	1.17%
burst_lockless/10	uring	ops_per_sec	1.83K ops/s	1.85K ops/s	🟢▲ +1.54%	0.99%
burst_lockless/100	epoll	ops_per_sec	191.0 ops/s	194.1 ops/s	🟢▲ +2.06%	1.07%
burst_lockless/100	uring	ops_per_sec	186.3 ops/s	189.6 ops/s	🟢▲ +2.22%	0.99%
concurrent/1	epoll	ops_per_sec	14.98K ops/s	15.29K ops/s	🟢▲ +2.41%	0.94%
concurrent/1	uring	ops_per_sec	12.77K ops/s	12.99K ops/s	🟢▲ +1.76%	0.80%
concurrent/16	epoll	ops_per_sec	14.99K ops/s	15.31K ops/s	🟢▲ +2.53%	1.13%
concurrent/16	uring	ops_per_sec	12.64K ops/s	12.85K ops/s	🟢▲ +1.69%	0.85%
concurrent/4	epoll	ops_per_sec	14.98K ops/s	15.29K ops/s	🟢▲ +2.30%	0.95%
concurrent/4	uring	ops_per_sec	12.75K ops/s	12.91K ops/s	🟢▲ +1.07%	1.11%
sequential	epoll	ops_per_sec	14.99K ops/s	15.31K ops/s	🟢▲ +2.46%	1.26%
sequential	uring	ops_per_sec	12.77K ops/s	12.98K ops/s	🟢▲ +1.54%	1.14%
sequential_lockless	epoll	ops_per_sec	15.05K ops/s	15.36K ops/s	🟢▲ +2.27%	1.13%
sequential_lockless	uring	ops_per_sec	13.66K ops/s	13.90K ops/s	🟢▲ +1.86%	0.98%

fan_out

Benchmark	Backend	Metric	Base	Head	Δ	Noise
concurrent_parents/1	epoll	ops_per_sec	3.71K ops/s	3.69K ops/s	🟢▲ +0.03%	1.13%
concurrent_parents/1	uring	ops_per_sec	3.45K ops/s	3.44K ops/s	🟢▲ +0.09%	1.20%
concurrent_parents/16	epoll	ops_per_sec	3.58K ops/s	3.58K ops/s	🔴▼ -0.20%	0.49%
concurrent_parents/16	uring	ops_per_sec	3.33K ops/s	3.31K ops/s	🔴▼ -0.23%	1.04%
concurrent_parents/4	epoll	ops_per_sec	3.61K ops/s	3.59K ops/s	🔴▼ -0.39%	1.11%
concurrent_parents/4	uring	ops_per_sec	3.36K ops/s	3.34K ops/s	🔴▼ -0.06%	1.13%
concurrent_parents_lockless/1	epoll	ops_per_sec	3.73K ops/s	3.73K ops/s	🟢▲ +0.06%	0.20%
concurrent_parents_lockless/1	uring	ops_per_sec	3.67K ops/s	3.66K ops/s	🔴▼ -0.15%	0.82%
concurrent_parents_lockless/16	epoll	ops_per_sec	3.61K ops/s	3.61K ops/s	🔴▼ -0.18%	0.18%
concurrent_parents_lockless/16	uring	ops_per_sec	3.52K ops/s	3.52K ops/s	🔴▼ -0.28%	0.38%
concurrent_parents_lockless/4	epoll	ops_per_sec	3.63K ops/s	3.63K ops/s	🔴▼ -0.03%	0.15%
concurrent_parents_lockless/4	uring	ops_per_sec	3.55K ops/s	3.55K ops/s	🔴▼ -0.03%	0.49%
fork_join/1	epoll	ops_per_sec	54.94K ops/s	54.75K ops/s	🟢▲ +0.09%	1.04%
fork_join/1	uring	ops_per_sec	53.48K ops/s	53.30K ops/s	🟢▲ +0.01%	1.05%
fork_join/16	epoll	ops_per_sec	3.69K ops/s	3.70K ops/s	🟢▲ +0.46%	0.26%
fork_join/16	uring	ops_per_sec	3.45K ops/s	3.45K ops/s	🔴▼ -0.25%	1.01%
fork_join/4	epoll	ops_per_sec	14.64K ops/s	14.65K ops/s	🟢▲ +0.01%	0.25%
fork_join/4	uring	ops_per_sec	13.79K ops/s	13.71K ops/s	🔴▼ -0.27%	1.30%
fork_join/64	epoll	ops_per_sec	897.4 ops/s	900.8 ops/s	🟢▲ +0.24%	0.19%
fork_join/64	uring	ops_per_sec	813.7 ops/s	817.0 ops/s	🟢▲ +0.02%	1.16%
fork_join_lockless/1	epoll	ops_per_sec	55.59K ops/s	55.51K ops/s	🔴▼ -0.22%	0.18%
fork_join_lockless/1	uring	ops_per_sec	55.95K ops/s	56.13K ops/s	🔴▼ -0.08%	1.11%
fork_join_lockless/16	epoll	ops_per_sec	3.73K ops/s	3.73K ops/s	🟢▲ +0.06%	0.20%
fork_join_lockless/16	uring	ops_per_sec	3.67K ops/s	3.67K ops/s	🟢▲ +0.40%	1.03%
fork_join_lockless/4	epoll	ops_per_sec	14.80K ops/s	14.82K ops/s	🟢▲ +0.06%	0.25%
fork_join_lockless/4	uring	ops_per_sec	14.61K ops/s	14.58K ops/s	🔴▼ -0.18%	1.23%
fork_join_lockless/64	epoll	ops_per_sec	907.8 ops/s	909.3 ops/s	🟢▲ +0.31%	0.26%
fork_join_lockless/64	uring	ops_per_sec	870.1 ops/s	869.5 ops/s	🟢▲ +0.30%	1.05%
nested/16	epoll	ops_per_sec	895.2 ops/s	892.7 ops/s	🟢▲ +0.25%	1.12%
nested/16	uring	ops_per_sec	810.3 ops/s	802.2 ops/s	🔴▼ -0.49%	1.06%
nested/4	epoll	ops_per_sec	3.68K ops/s	3.67K ops/s	🟢▲ +0.04%	1.25%
nested/4	uring	ops_per_sec	3.42K ops/s	3.41K ops/s	🔴▼ -0.12%	1.19%
nested_lockless/16	epoll	ops_per_sec	905.0 ops/s	902.7 ops/s	🟢▲ +0.05%	1.01%
nested_lockless/16	uring	ops_per_sec	861.0 ops/s	856.9 ops/s	🔴▼ -0.27%	0.95%
nested_lockless/4	epoll	ops_per_sec	3.72K ops/s	3.71K ops/s	🟢▲ +0.11%	1.15%
nested_lockless/4	uring	ops_per_sec	3.65K ops/s	3.63K ops/s	🔴▼ -0.49%	1.13%

http_server

Benchmark	Backend	Metric	Base	Head	Δ	Noise
concurrent/1	epoll	ops_per_sec	55.12K ops/s	55.36K ops/s	🟢▲ +0.29%	0.94%
concurrent/1	uring	ops_per_sec	53.47K ops/s	53.84K ops/s	🟢▲ +0.27%	0.86%
concurrent/16	epoll	ops_per_sec	57.82K ops/s	57.84K ops/s	🟢▲ +0.04%	0.15%
concurrent/16	uring	ops_per_sec	52.82K ops/s	53.20K ops/s	🟢▲ +0.14%	1.08%
concurrent/32	epoll	ops_per_sec	57.40K ops/s	57.46K ops/s	🟢▲ +0.19%	0.29%
concurrent/32	uring	ops_per_sec	52.44K ops/s	52.82K ops/s	🟢▲ +0.26%	1.13%
concurrent/4	epoll	ops_per_sec	57.16K ops/s	57.45K ops/s	🟢▲ +0.36%	0.96%
concurrent/4	uring	ops_per_sec	52.82K ops/s	53.34K ops/s	🟢▲ +0.40%	1.21%
multithread/1	epoll	ops_per_sec	57.38K ops/s	57.45K ops/s	🟢▲ +0.08%	0.11%
multithread/1	uring	ops_per_sec	52.43K ops/s	52.82K ops/s	🟢▲ +0.26%	1.13%
multithread/16	epoll	ops_per_sec	162.1K ops/s	162.3K ops/s	🔴▼ -0.09%	0.41%
multithread/16	uring	ops_per_sec	75.25K ops/s	75.47K ops/s	🟢▲ +0.42%	19.53%
multithread/2	epoll	ops_per_sec	87.31K ops/s	86.90K ops/s	🔴▼ -0.73%	0.84%
multithread/2	uring	ops_per_sec	90.43K ops/s	91.36K ops/s	🟢▲ +0.86%	0.62%
multithread/4	epoll	ops_per_sec	159.8K ops/s	159.7K ops/s	🔴▼ -0.26%	0.73%
multithread/4	uring	ops_per_sec	80.15K ops/s	80.32K ops/s	🟢▲ +0.31%	0.27%
multithread/8	epoll	ops_per_sec	162.2K ops/s	161.9K ops/s	🔴▼ -0.29%	0.31%
multithread/8	uring	ops_per_sec	77.50K ops/s	77.83K ops/s	🟢▲ +0.33%	0.26%
single_conn	epoll	ops_per_sec	55.09K ops/s	55.30K ops/s	🟢▲ +0.06%	0.79%
single_conn	uring	ops_per_sec	53.34K ops/s	53.52K ops/s	🟢▲ +0.45%	0.39%
single_conn_lockless	epoll	ops_per_sec	55.74K ops/s	56.11K ops/s	🟢▲ +0.61%	0.99%
single_conn_lockless	uring	ops_per_sec	55.95K ops/s	56.27K ops/s	🟢▲ +0.53%	0.33%

local_socket_latency

Benchmark	Backend	Metric	Base	Head	Δ	Noise
concurrent/1	epoll	latency_mean_ns	7.17 µs	7.23 µs	🟢▲ +0.47%	4.08%
concurrent/1	uring	latency_mean_ns	6.56 µs	6.56 µs	🔴▼ -0.26%	1.16%
concurrent/16	epoll	latency_mean_ns	109.89 µs	109.72 µs	🟢▲ +0.34%	0.91%
concurrent/16	uring	latency_mean_ns	108.11 µs	108.02 µs	🟢▲ +0.19%	1.27%
concurrent/4	epoll	latency_mean_ns	27.74 µs	27.70 µs	🟢▲ +0.16%	1.42%
concurrent/4	uring	latency_mean_ns	27.03 µs	26.67 µs	🟢▲ +2.13%	2.28%
concurrent_lockless/1	epoll	latency_mean_ns	7.25 µs	7.25 µs	🟢▲ +0.31%	5.73%
concurrent_lockless/1	uring	latency_mean_ns	6.55 µs	6.81 µs	🔴▼ -0.03%	6.08%
concurrent_lockless/16	epoll	latency_mean_ns	108.34 µs	108.63 µs	🔴▼ -0.07%	0.91%
concurrent_lockless/16	uring	latency_mean_ns	108.02 µs	107.74 µs	🟢▲ +0.28%	0.70%
concurrent_lockless/4	epoll	latency_mean_ns	27.19 µs	27.48 µs	🟢▲ +0.09%	2.12%
concurrent_lockless/4	uring	latency_mean_ns	26.86 µs	26.68 µs	🟢▲ +0.30%	1.23%
pingpong/1	epoll	latency_mean_ns	7.20 µs	7.18 µs	🟢▲ +0.52%	4.28%
pingpong/1	uring	latency_mean_ns	6.60 µs	6.61 µs	🟢▲ +0.25%	2.82%
pingpong/1024	epoll	latency_mean_ns	8.02 µs	7.59 µs	🟢▲ +8.21%	4.93%
pingpong/1024	uring	latency_mean_ns	7.09 µs	7.22 µs	🟢▲ +0.22%	5.18%
pingpong/64	epoll	latency_mean_ns	7.38 µs	7.08 µs	🟢▲ +0.72%	4.80%
pingpong/64	uring	latency_mean_ns	6.72 µs	6.57 µs	🟢▲ +0.54%	4.35%
pingpong_lockless/1	epoll	latency_mean_ns	6.96 µs	6.97 µs	🔴▼ -0.04%	0.40%
pingpong_lockless/1	uring	latency_mean_ns	6.54 µs	6.53 µs	🟢▲ +0.44%	0.67%
pingpong_lockless/1024	epoll	latency_mean_ns	7.52 µs	7.52 µs	⚪ 0.00%	5.09%
pingpong_lockless/1024	uring	latency_mean_ns	7.26 µs	7.30 µs	🟢▲ +0.17%	4.53%
pingpong_lockless/64	epoll	latency_mean_ns	7.14 µs	7.28 µs	🔴▼ -0.02%	5.21%
pingpong_lockless/64	uring	latency_mean_ns	6.57 µs	6.70 µs	🔴▼ -0.22%	4.31%

local_socket_throughput

Benchmark	Backend	Metric	Base	Head	Δ	Noise
bidirectional/1024	epoll	bytes_per_sec	305.3 MB/s	304.9 MB/s	🔴▼ -0.13%	1.04%
bidirectional/1024	uring	bytes_per_sec	301.9 MB/s	301.7 MB/s	⚪ 0.00%	0.65%
bidirectional/1048576	epoll	bytes_per_sec	3.16 GB/s	3.16 GB/s	🟢▲ +0.29%	2.32%
bidirectional/1048576	uring	bytes_per_sec	3.32 GB/s	3.33 GB/s	🔴▼ -0.35%	1.76%
bidirectional/16384	epoll	bytes_per_sec	3.11 GB/s	3.02 GB/s	🔴▼ -3.23%	1.74%
bidirectional/16384	uring	bytes_per_sec	2.39 GB/s	2.40 GB/s	🔴▼ -0.37%	2.48%
bidirectional/262144	epoll	bytes_per_sec	3.18 GB/s	3.20 GB/s	🟢▲ +0.83%	2.53%
bidirectional/262144	uring	bytes_per_sec	3.31 GB/s	3.31 GB/s	🟢▲ +0.34%	1.24%
bidirectional/4096	epoll	bytes_per_sec	1.01 GB/s	998.8 MB/s	🔴▼ -0.64%	0.78%
bidirectional/4096	uring	bytes_per_sec	995.1 MB/s	989.4 MB/s	🔴▼ -0.53%	0.72%
bidirectional/65536	epoll	bytes_per_sec	2.98 GB/s	2.97 GB/s	🔴▼ -1.53%	1.88%
bidirectional/65536	uring	bytes_per_sec	3.05 GB/s	3.01 GB/s	🔴▼ -0.76%	2.40%
bidirectional_lockless/1024	epoll	bytes_per_sec	308.1 MB/s	308.7 MB/s	🔴▼ -0.33%	1.24%
bidirectional_lockless/1024	uring	bytes_per_sec	304.7 MB/s	304.8 MB/s	🟢▲ +0.90%	0.81%
bidirectional_lockless/1048576	epoll	bytes_per_sec	3.16 GB/s	3.17 GB/s	🔴▼ -0.78%	2.63%
bidirectional_lockless/1048576	uring	bytes_per_sec	3.39 GB/s	3.40 GB/s	🔴▼ -0.97%	2.45%
bidirectional_lockless/16384	epoll	bytes_per_sec	3.08 GB/s	3.09 GB/s	🔴▼ -0.04%	2.45%
bidirectional_lockless/16384	uring	bytes_per_sec	2.59 GB/s	2.62 GB/s	🟢▲ +1.44%	2.94%
bidirectional_lockless/262144	epoll	bytes_per_sec	3.19 GB/s	3.20 GB/s	🟢▲ +0.56%	2.88%
bidirectional_lockless/262144	uring	bytes_per_sec	3.37 GB/s	3.37 GB/s	🟢▲ +1.13%	1.44%
bidirectional_lockless/4096	epoll	bytes_per_sec	1.02 GB/s	1.01 GB/s	🔴▼ -0.51%	0.84%
bidirectional_lockless/4096	uring	bytes_per_sec	1.00 GB/s	991.0 MB/s	🔴▼ -0.79%	1.21%
bidirectional_lockless/65536	epoll	bytes_per_sec	2.98 GB/s	2.95 GB/s	🔴▼ -1.69%	2.02%
bidirectional_lockless/65536	uring	bytes_per_sec	3.22 GB/s	3.18 GB/s	🔴▼ -0.88%	2.28%
unidirectional/1024	epoll	bytes_per_sec	297.5 MB/s	297.1 MB/s	🔴▼ -0.47%	1.07%
unidirectional/1024	uring	bytes_per_sec	303.1 MB/s	302.2 MB/s	🟢▲ +0.10%	1.01%
unidirectional/1048576	epoll	bytes_per_sec	3.30 GB/s	3.29 GB/s	🔴▼ -1.02%	2.24%
unidirectional/1048576	uring	bytes_per_sec	3.46 GB/s	3.46 GB/s	🔴▼ -1.35%	2.40%
unidirectional/16384	epoll	bytes_per_sec	3.04 GB/s	3.05 GB/s	🟢▲ +0.41%	1.90%
unidirectional/16384	uring	bytes_per_sec	2.63 GB/s	2.59 GB/s	🔴▼ -1.50%	3.36%
unidirectional/262144	epoll	bytes_per_sec	3.29 GB/s	3.33 GB/s	🟢▲ +0.26%	2.45%
unidirectional/262144	uring	bytes_per_sec	3.42 GB/s	3.45 GB/s	🟢▲ +1.48%	2.90%
unidirectional/4096	epoll	bytes_per_sec	989.3 MB/s	979.6 MB/s	🔴▼ -0.94%	0.55%
unidirectional/4096	uring	bytes_per_sec	991.0 MB/s	992.2 MB/s	🟢▲ +0.74%	1.87%
unidirectional/65536	epoll	bytes_per_sec	3.27 GB/s	3.25 GB/s	🔴▼ -0.79%	1.47%
unidirectional/65536	uring	bytes_per_sec	3.31 GB/s	3.32 GB/s	🟢▲ +0.80%	2.65%
unidirectional_lockless/1024	epoll	bytes_per_sec	301.9 MB/s	301.6 MB/s	🟢▲ +0.26%	1.16%
unidirectional_lockless/1024	uring	bytes_per_sec	305.1 MB/s	304.5 MB/s	🔴▼ -0.17%	0.95%
unidirectional_lockless/1048576	epoll	bytes_per_sec	3.31 GB/s	3.32 GB/s	🟢▲ +0.61%	3.14%
unidirectional_lockless/1048576	uring	bytes_per_sec	3.39 GB/s	3.40 GB/s	🟢▲ +0.39%	2.23%
unidirectional_lockless/16384	epoll	bytes_per_sec	3.07 GB/s	3.08 GB/s	🟢▲ +1.00%	1.74%
unidirectional_lockless/16384	uring	bytes_per_sec	2.67 GB/s	2.60 GB/s	🔴▼ -3.03%	3.46%
unidirectional_lockless/262144	epoll	bytes_per_sec	3.30 GB/s	3.35 GB/s	🟢▲ +0.72%	1.68%
unidirectional_lockless/262144	uring	bytes_per_sec	3.38 GB/s	3.39 GB/s	🟢▲ +2.44%	4.16%
unidirectional_lockless/4096	epoll	bytes_per_sec	996.7 MB/s	989.6 MB/s	🔴▼ -1.16%	0.97%
unidirectional_lockless/4096	uring	bytes_per_sec	991.4 MB/s	995.2 MB/s	🔴▼ -0.54%	1.15%
unidirectional_lockless/65536	epoll	bytes_per_sec	3.25 GB/s	3.26 GB/s	🟢▲ +0.40%	1.15%
unidirectional_lockless/65536	uring	bytes_per_sec	3.31 GB/s	3.30 GB/s	🔴▼ -0.31%	2.04%

socket_latency

Benchmark	Backend	Metric	Base	Head	Δ	Noise
concurrent/1	epoll	latency_mean_ns	14.79 µs	14.71 µs	🟢▲ +0.28%	1.06%
concurrent/1	uring	latency_mean_ns	14.53 µs	14.55 µs	🔴▼ -0.15%	0.39%
concurrent/16	epoll	latency_mean_ns	233.21 µs	233.11 µs	🟢▲ +0.12%	0.16%
concurrent/16	uring	latency_mean_ns	235.46 µs	235.34 µs	🟢▲ +0.01%	0.33%
concurrent/4	epoll	latency_mean_ns	58.23 µs	58.20 µs	🔴▼ -0.05%	0.18%
concurrent/4	uring	latency_mean_ns	58.40 µs	58.41 µs	🔴▼ -0.03%	0.37%
concurrent_lockless/1	epoll	latency_mean_ns	14.67 µs	14.62 µs	🟢▲ +0.29%	0.86%
concurrent_lockless/1	uring	latency_mean_ns	14.54 µs	14.52 µs	🟢▲ +0.24%	0.49%
concurrent_lockless/16	epoll	latency_mean_ns	231.96 µs	230.50 µs	🟢▲ +0.15%	1.08%
concurrent_lockless/16	uring	latency_mean_ns	235.27 µs	234.75 µs	🟢▲ +0.31%	0.32%
concurrent_lockless/4	epoll	latency_mean_ns	57.88 µs	57.52 µs	🟢▲ +0.15%	1.21%
concurrent_lockless/4	uring	latency_mean_ns	58.41 µs	58.36 µs	🟢▲ +0.27%	0.32%
pingpong/1	epoll	latency_mean_ns	14.63 µs	14.60 µs	🟢▲ +0.14%	1.15%
pingpong/1	uring	latency_mean_ns	14.46 µs	14.50 µs	🔴▼ -0.39%	0.27%
pingpong/1024	epoll	latency_mean_ns	15.03 µs	14.92 µs	🟢▲ +0.22%	1.07%
pingpong/1024	uring	latency_mean_ns	14.81 µs	14.79 µs	🟢▲ +0.24%	0.35%
pingpong/64	epoll	latency_mean_ns	14.73 µs	14.64 µs	🟢▲ +0.04%	1.33%
pingpong/64	uring	latency_mean_ns	14.54 µs	14.55 µs	🟢▲ +0.04%	0.32%
pingpong_lockless/1	epoll	latency_mean_ns	14.61 µs	14.54 µs	🔴▼ -0.03%	1.12%
pingpong_lockless/1	uring	latency_mean_ns	14.45 µs	14.45 µs	🟢▲ +0.02%	0.35%
pingpong_lockless/1024	epoll	latency_mean_ns	14.93 µs	14.87 µs	🟢▲ +0.13%	1.15%
pingpong_lockless/1024	uring	latency_mean_ns	14.76 µs	14.76 µs	🟢▲ +0.13%	0.46%
pingpong_lockless/64	epoll	latency_mean_ns	14.67 µs	14.61 µs	🟢▲ +0.08%	1.16%
pingpong_lockless/64	uring	latency_mean_ns	14.53 µs	14.53 µs	🔴▼ -0.09%	0.40%

socket_throughput

Benchmark	Backend	Metric	Base	Head	Δ	Noise
bidirectional/1024	epoll	bytes_per_sec	135.4 MB/s	134.2 MB/s	🔴▼ -0.32%	1.34%
bidirectional/1024	uring	bytes_per_sec	140.9 MB/s	140.5 MB/s	🔴▼ -0.18%	0.99%
bidirectional/1048576	epoll	bytes_per_sec	3.30 GB/s	3.32 GB/s	🟢▲ +0.15%	2.10%
bidirectional/1048576	uring	bytes_per_sec	3.24 GB/s	3.26 GB/s	🟢▲ +0.42%	0.63%
bidirectional/16384	epoll	bytes_per_sec	1.64 GB/s	1.65 GB/s	🟢▲ +0.65%	0.87%
bidirectional/16384	uring	bytes_per_sec	1.55 GB/s	1.54 GB/s	🟢▲ +0.39%	2.50%
bidirectional/262144	epoll	bytes_per_sec	3.23 GB/s	3.21 GB/s	🔴▼ -0.24%	1.05%
bidirectional/262144	uring	bytes_per_sec	3.21 GB/s	3.15 GB/s	🔴▼ -1.77%	1.19%
bidirectional/4096	epoll	bytes_per_sec	507.3 MB/s	506.8 MB/s	🟢▲ +0.42%	1.35%
bidirectional/4096	uring	bytes_per_sec	492.0 MB/s	491.2 MB/s	🟢▲ +0.28%	1.45%
bidirectional/65536	epoll	bytes_per_sec	2.64 GB/s	2.58 GB/s	🔴▼ -2.02%	2.32%
bidirectional/65536	uring	bytes_per_sec	2.78 GB/s	2.80 GB/s	🟢▲ +0.55%	1.02%
bidirectional_lockless/1024	epoll	bytes_per_sec	135.2 MB/s	134.6 MB/s	🔴▼ -0.36%	0.99%
bidirectional_lockless/1024	uring	bytes_per_sec	140.8 MB/s	140.9 MB/s	🟢▲ +0.05%	0.41%
bidirectional_lockless/1048576	epoll	bytes_per_sec	3.30 GB/s	3.34 GB/s	🟢▲ +0.20%	3.26%
bidirectional_lockless/1048576	uring	bytes_per_sec	3.36 GB/s	3.32 GB/s	🔴▼ -0.98%	0.91%
bidirectional_lockless/16384	epoll	bytes_per_sec	1.64 GB/s	1.63 GB/s	🟢▲ +0.09%	2.23%
bidirectional_lockless/16384	uring	bytes_per_sec	1.54 GB/s	1.54 GB/s	🟢▲ +0.09%	0.39%
bidirectional_lockless/262144	epoll	bytes_per_sec	3.22 GB/s	3.22 GB/s	🟢▲ +0.22%	0.99%
bidirectional_lockless/262144	uring	bytes_per_sec	3.28 GB/s	3.24 GB/s	🔴▼ -1.09%	1.00%
bidirectional_lockless/4096	epoll	bytes_per_sec	509.4 MB/s	506.9 MB/s	🔴▼ -0.73%	0.72%
bidirectional_lockless/4096	uring	bytes_per_sec	493.8 MB/s	494.2 MB/s	🟢▲ +0.17%	0.43%
bidirectional_lockless/65536	epoll	bytes_per_sec	2.58 GB/s	2.58 GB/s	🟢▲ +2.64%	3.33%
bidirectional_lockless/65536	uring	bytes_per_sec	2.76 GB/s	2.78 GB/s	🟢▲ +1.38%	1.69%
multithread/2	epoll	bytes_per_sec	3.41 GB/s	3.39 GB/s	🔴▼ -0.40%	0.50%
multithread/2	uring	bytes_per_sec	3.63 GB/s	3.67 GB/s	🔴▼ -0.17%	1.15%
multithread/4	epoll	bytes_per_sec	4.85 GB/s	4.88 GB/s	🟢▲ +0.35%	0.51%
multithread/4	uring	bytes_per_sec	5.07 GB/s	5.06 GB/s	🔴▼ -0.48%	0.45%
multithread/8	epoll	bytes_per_sec	4.96 GB/s	4.97 GB/s	🟢▲ +0.19%	0.14%
multithread/8	uring	bytes_per_sec	5.13 GB/s	5.14 GB/s	🟢▲ +0.07%	0.23%
unidirectional/1024	epoll	bytes_per_sec	125.5 MB/s	125.6 MB/s	⚪ 0.00%	0.25%
unidirectional/1024	uring	bytes_per_sec	139.5 MB/s	140.3 MB/s	🟢▲ +0.37%	1.21%
unidirectional/1048576	epoll	bytes_per_sec	3.34 GB/s	3.36 GB/s	🟢▲ +0.29%	1.45%
unidirectional/1048576	uring	bytes_per_sec	3.41 GB/s	3.43 GB/s	🟢▲ +0.33%	0.82%
unidirectional/16384	epoll	bytes_per_sec	1.56 GB/s	1.54 GB/s	🔴▼ -1.20%	1.35%
unidirectional/16384	uring	bytes_per_sec	1.55 GB/s	1.58 GB/s	🟢▲ +1.83%	2.10%
unidirectional/262144	epoll	bytes_per_sec	3.34 GB/s	3.33 GB/s	🔴▼ -0.82%	0.71%
unidirectional/262144	uring	bytes_per_sec	3.20 GB/s	3.21 GB/s	🔴▼ -0.20%	0.77%
unidirectional/4096	epoll	bytes_per_sec	470.5 MB/s	472.1 MB/s	🟢▲ +0.14%	0.64%
unidirectional/4096	uring	bytes_per_sec	489.5 MB/s	492.0 MB/s	🟢▲ +0.06%	1.03%
unidirectional/65536	epoll	bytes_per_sec	2.63 GB/s	2.69 GB/s	🟢▲ +0.13%	5.37%
unidirectional/65536	uring	bytes_per_sec	2.71 GB/s	2.70 GB/s	🟢▲ +0.89%	2.93%
unidirectional_lockless/1024	epoll	bytes_per_sec	126.6 MB/s	126.3 MB/s	🔴▼ -0.02%	0.71%
unidirectional_lockless/1024	uring	bytes_per_sec	139.4 MB/s	139.1 MB/s	🔴▼ -0.24%	0.48%
unidirectional_lockless/1048576	epoll	bytes_per_sec	3.39 GB/s	3.37 GB/s	🔴▼ -0.57%	1.67%
unidirectional_lockless/1048576	uring	bytes_per_sec	3.36 GB/s	3.38 GB/s	🟢▲ +0.32%	0.98%
unidirectional_lockless/16384	epoll	bytes_per_sec	1.58 GB/s	1.57 GB/s	🔴▼ -1.38%	1.44%
unidirectional_lockless/16384	uring	bytes_per_sec	1.61 GB/s	1.63 GB/s	🟢▲ +0.89%	1.35%
unidirectional_lockless/262144	epoll	bytes_per_sec	3.38 GB/s	3.35 GB/s	🔴▼ -0.40%	1.82%
unidirectional_lockless/262144	uring	bytes_per_sec	3.26 GB/s	3.25 GB/s	🟢▲ +0.17%	1.20%
unidirectional_lockless/4096	epoll	bytes_per_sec	475.8 MB/s	474.7 MB/s	🟢▲ +0.13%	1.14%
unidirectional_lockless/4096	uring	bytes_per_sec	492.9 MB/s	492.0 MB/s	🟢▲ +0.44%	0.94%
unidirectional_lockless/65536	epoll	bytes_per_sec	2.76 GB/s	2.75 GB/s	🔴▼ -0.22%	2.75%
unidirectional_lockless/65536	uring	bytes_per_sec	2.75 GB/s	2.71 GB/s	🔴▼ -0.12%	3.79%

windows — full results (77 benchmarks, 5 iterations, 1.5s each)

accept_churn

Benchmark	Backend	Metric	Base	Head	Δ	Noise
burst/10	iocp	ops_per_sec	1.53K ops/s	1.55K ops/s	🟢▲ +1.42%	0.25%
burst/100	iocp	ops_per_sec	149.1 ops/s	155.6 ops/s	🟢▲ +4.40%	0.93%
burst_lockless/10	iocp	ops_per_sec	1.53K ops/s	1.56K ops/s	🟢▲ +1.65%	0.18%
burst_lockless/100	iocp	ops_per_sec	152.4 ops/s	158.5 ops/s	🟢▲ +3.85%	0.64%
concurrent/1	iocp	ops_per_sec	12.24K ops/s	12.49K ops/s	🟢▲ +1.84%	1.42%
concurrent/16	iocp	ops_per_sec	13.21K ops/s	13.46K ops/s	🟢▲ +1.79%	0.45%
concurrent/4	iocp	ops_per_sec	12.92K ops/s	13.15K ops/s	🟢▲ +1.88%	0.38%
sequential	iocp	ops_per_sec	12.19K ops/s	12.44K ops/s	🟢▲ +1.62%	1.26%
sequential_lockless	iocp	ops_per_sec	12.28K ops/s	12.51K ops/s	🟢▲ +1.14%	1.33%

fan_out

Benchmark	Backend	Metric	Base	Head	Δ	Noise
concurrent_parents/1	iocp	ops_per_sec	7.20K ops/s	7.23K ops/s	🟢▲ +0.23%	0.36%
concurrent_parents/16	iocp	ops_per_sec	6.46K ops/s	6.47K ops/s	🟢▲ +0.05%	1.41%
concurrent_parents/4	iocp	ops_per_sec	6.91K ops/s	6.95K ops/s	🟢▲ +0.83%	0.79%
concurrent_parents_lockless/1	iocp	ops_per_sec	7.20K ops/s	7.22K ops/s	🟢▲ +0.50%	0.39%
concurrent_parents_lockless/16	iocp	ops_per_sec	6.49K ops/s	6.51K ops/s	🟢▲ +0.32%	1.50%
concurrent_parents_lockless/4	iocp	ops_per_sec	6.93K ops/s	6.98K ops/s	🟢▲ +0.75%	0.64%
fork_join/1	iocp	ops_per_sec	113.5K ops/s	113.6K ops/s	🟢▲ +0.35%	0.46%
fork_join/16	iocp	ops_per_sec	7.17K ops/s	7.24K ops/s	🟢▲ +0.52%	0.69%
fork_join/4	iocp	ops_per_sec	29.37K ops/s	29.41K ops/s	🟢▲ +0.30%	0.50%
fork_join/64	iocp	ops_per_sec	1.72K ops/s	1.74K ops/s	🟢▲ +0.47%	1.25%
fork_join_lockless/1	iocp	ops_per_sec	113.3K ops/s	113.6K ops/s	🟢▲ +0.13%	0.41%
fork_join_lockless/16	iocp	ops_per_sec	7.18K ops/s	7.23K ops/s	🟢▲ +0.30%	0.31%
fork_join_lockless/4	iocp	ops_per_sec	29.31K ops/s	29.42K ops/s	🟢▲ +0.34%	0.24%
fork_join_lockless/64	iocp	ops_per_sec	1.73K ops/s	1.74K ops/s	🟢▲ +0.67%	1.00%
nested/16	iocp	ops_per_sec	1.70K ops/s	1.71K ops/s	🟢▲ +0.72%	1.21%
nested/4	iocp	ops_per_sec	7.04K ops/s	7.08K ops/s	🟢▲ +0.60%	0.25%
nested_lockless/16	iocp	ops_per_sec	1.70K ops/s	1.71K ops/s	🟢▲ +0.20%	1.20%
nested_lockless/4	iocp	ops_per_sec	7.03K ops/s	7.08K ops/s	🟢▲ +0.74%	0.32%

http_server

Benchmark	Backend	Metric	Base	Head	Δ	Noise
concurrent/1	iocp	ops_per_sec	111.6K ops/s	112.0K ops/s	🟢▲ +0.58%	0.49%
concurrent/16	iocp	ops_per_sec	108.3K ops/s	109.2K ops/s	🟢▲ +0.54%	0.44%
concurrent/32	iocp	ops_per_sec	107.7K ops/s	108.8K ops/s	🟢▲ +1.12%	0.31%
concurrent/4	iocp	ops_per_sec	109.2K ops/s	109.7K ops/s	🟢▲ +0.27%	0.48%
multithread/1	iocp	ops_per_sec	107.8K ops/s	108.9K ops/s	🟢▲ +1.15%	0.30%
multithread/16	iocp	ops_per_sec	266.1K ops/s	267.1K ops/s	🟢▲ +0.35%	0.58%
multithread/2	iocp	ops_per_sec	180.9K ops/s	182.3K ops/s	🟢▲ +0.71%	0.96%
multithread/4	iocp	ops_per_sec	224.6K ops/s	225.7K ops/s	🟢▲ +0.53%	0.83%
multithread/8	iocp	ops_per_sec	265.5K ops/s	267.7K ops/s	🟢▲ +0.63%	0.45%
single_conn	iocp	ops_per_sec	111.8K ops/s	111.8K ops/s	🔴▼ -0.17%	0.57%
single_conn_lockless	iocp	ops_per_sec	111.7K ops/s	112.2K ops/s	🟢▲ +0.46%	0.51%

socket_latency

Benchmark	Backend	Metric	Base	Head	Δ	Noise
concurrent/1	iocp	latency_mean_ns	7.59 µs	7.55 µs	🟢▲ +0.36%	0.33%
concurrent/16	iocp	latency_mean_ns	125.94 µs	127.93 µs	🔴▼ -1.01%	3.55%
concurrent/4	iocp	latency_mean_ns	30.39 µs	30.50 µs	🔴▼ -0.34%	0.52%
concurrent_lockless/1	iocp	latency_mean_ns	7.59 µs	7.56 µs	🟢▲ +0.17%	0.39%
concurrent_lockless/16	iocp	latency_mean_ns	121.12 µs	121.06 µs	🟢▲ +0.14%	0.38%
concurrent_lockless/4	iocp	latency_mean_ns	30.32 µs	30.41 µs	🔴▼ -0.42%	0.35%
pingpong/1	iocp	latency_mean_ns	7.57 µs	7.54 µs	🟢▲ +0.38%	0.44%
pingpong/1024	iocp	latency_mean_ns	7.68 µs	7.69 µs	🟢▲ +0.08%	0.38%
pingpong/64	iocp	latency_mean_ns	7.59 µs	7.58 µs	🔴▼ -0.03%	0.24%
pingpong_lockless/1	iocp	latency_mean_ns	7.56 µs	7.53 µs	🟢▲ +0.52%	0.55%
pingpong_lockless/1024	iocp	latency_mean_ns	7.68 µs	7.68 µs	🟢▲ +0.05%	0.26%
pingpong_lockless/64	iocp	latency_mean_ns	7.60 µs	7.56 µs	🟢▲ +0.61%	0.26%

socket_throughput

Benchmark	Backend	Metric	Base	Head	Δ	Noise
bidirectional/1024	iocp	bytes_per_sec	282.3 MB/s	282.8 MB/s	🟢▲ +0.07%	0.64%
bidirectional/1048576	iocp	bytes_per_sec	5.72 GB/s	5.20 GB/s	🔴▼ -13.04%	8.36%
bidirectional/16384	iocp	bytes_per_sec	3.75 GB/s	3.77 GB/s	🟢▲ +0.42%	0.70%
bidirectional/262144	iocp	bytes_per_sec	9.41 GB/s	9.62 GB/s	🟢▲ +2.18%	1.77%
bidirectional/4096	iocp	bytes_per_sec	1.08 GB/s	1.09 GB/s	🟢▲ +0.37%	0.40%
bidirectional/65536	iocp	bytes_per_sec	9.01 GB/s	9.09 GB/s	🟢▲ +0.29%	2.38%
bidirectional_lockless/1024	iocp	bytes_per_sec	281.9 MB/s	282.5 MB/s	🟢▲ +0.26%	0.42%
bidirectional_lockless/1048576	iocp	bytes_per_sec	5.94 GB/s	5.52 GB/s	🔴▼ -8.07%	6.07%
bidirectional_lockless/16384	iocp	bytes_per_sec	3.73 GB/s	3.73 GB/s	🔴▼ -0.05%	0.67%
bidirectional_lockless/262144	iocp	bytes_per_sec	9.59 GB/s	9.69 GB/s	🟢▲ +1.84%	2.56%
bidirectional_lockless/4096	iocp	bytes_per_sec	1.08 GB/s	1.08 GB/s	🟢▲ +0.31%	0.39%
bidirectional_lockless/65536	iocp	bytes_per_sec	9.01 GB/s	8.95 GB/s	🔴▼ -0.59%	1.70%
multithread/2	iocp	bytes_per_sec	12.80 GB/s	12.72 GB/s	🔴▼ -0.18%	1.06%
multithread/4	iocp	bytes_per_sec	16.95 GB/s	17.13 GB/s	🟢▲ +0.27%	1.17%
multithread/8	iocp	bytes_per_sec	18.42 GB/s	18.50 GB/s	🔴▼ -0.17%	1.14%
unidirectional/1024	iocp	bytes_per_sec	285.4 MB/s	286.5 MB/s	🟢▲ +0.06%	0.66%
unidirectional/1048576	iocp	bytes_per_sec	4.35 GB/s	4.39 GB/s	🟢▲ +1.73%	5.29%
unidirectional/16384	iocp	bytes_per_sec	3.80 GB/s	3.80 GB/s	🔴▼ -0.16%	0.41%
unidirectional/262144	iocp	bytes_per_sec	8.62 GB/s	9.12 GB/s	🟢▲ +5.37%	5.29%
unidirectional/4096	iocp	bytes_per_sec	1.10 GB/s	1.10 GB/s	🔴▼ -0.23%	0.57%
unidirectional/65536	iocp	bytes_per_sec	9.27 GB/s	9.31 GB/s	🟢▲ +0.04%	1.73%
unidirectional_lockless/1024	iocp	bytes_per_sec	285.7 MB/s	287.0 MB/s	🟢▲ +0.54%	0.44%
unidirectional_lockless/1048576	iocp	bytes_per_sec	5.19 GB/s	4.85 GB/s	🔴▼ -5.51%	4.53%
unidirectional_lockless/16384	iocp	bytes_per_sec	3.79 GB/s	3.81 GB/s	🟢▲ +0.57%	1.08%
unidirectional_lockless/262144	iocp	bytes_per_sec	9.66 GB/s	9.75 GB/s	🟢▲ +0.22%	2.30%
unidirectional_lockless/4096	iocp	bytes_per_sec	1.10 GB/s	1.10 GB/s	🔴▼ -0.03%	0.54%
unidirectional_lockless/65536	iocp	bytes_per_sec	9.21 GB/s	9.39 GB/s	⚪ 0.00%	3.99%

macos — full results (113 benchmarks, 5 iterations, 1.5s each)

accept_churn

Benchmark	Backend	Metric	Base	Head	Δ	Noise
burst/10	kqueue	ops_per_sec	4.26K ops/s	4.40K ops/s	🟢▲ +2.84%	1.13%
burst/100	kqueue	ops_per_sec	431.9 ops/s	443.3 ops/s	🟢▲ +3.36%	0.81%
burst_lockless/10	kqueue	ops_per_sec	4.28K ops/s	4.40K ops/s	🟢▲ +2.88%	0.98%
burst_lockless/100	kqueue	ops_per_sec	436.2 ops/s	452.3 ops/s	🟢▲ +3.25%	1.35%
concurrent/1	kqueue	ops_per_sec	19.58K ops/s	20.18K ops/s	🟢▲ +3.49%	1.03%
concurrent/16	kqueue	ops_per_sec	35.63K ops/s	36.46K ops/s	🟢▲ +2.65%	1.14%
concurrent/4	kqueue	ops_per_sec	29.10K ops/s	29.24K ops/s	🔴▼ -0.08%	0.86%
sequential	kqueue	ops_per_sec	19.88K ops/s	20.37K ops/s	🟢▲ +2.21%	0.74%
sequential_lockless	kqueue	ops_per_sec	19.80K ops/s	20.14K ops/s	🟢▲ +1.86%	1.20%

fan_out

Benchmark	Backend	Metric	Base	Head	Δ	Noise
concurrent_parents/1	kqueue	ops_per_sec	7.67K ops/s	7.69K ops/s	🔴▼ -0.04%	1.08%
concurrent_parents/16	kqueue	ops_per_sec	7.99K ops/s	8.04K ops/s	🔴▼ -0.23%	14.60%
concurrent_parents/4	kqueue	ops_per_sec	8.13K ops/s	8.06K ops/s	🔴▼ -0.32%	1.82%
concurrent_parents_lockless/1	kqueue	ops_per_sec	7.74K ops/s	7.73K ops/s	🔴▼ -0.11%	1.18%
concurrent_parents_lockless/16	kqueue	ops_per_sec	8.08K ops/s	8.05K ops/s	🟢▲ +0.39%	14.77%
concurrent_parents_lockless/4	kqueue	ops_per_sec	8.15K ops/s	8.14K ops/s	🔴▼ -0.03%	1.58%
fork_join/1	kqueue	ops_per_sec	45.79K ops/s	45.79K ops/s	🟢▲ +0.43%	1.26%
fork_join/16	kqueue	ops_per_sec	7.69K ops/s	7.69K ops/s	🔴▼ -0.06%	1.18%
fork_join/4	kqueue	ops_per_sec	21.71K ops/s	21.81K ops/s	🟢▲ +0.69%	0.51%
fork_join/64	kqueue	ops_per_sec	2.02K ops/s	2.02K ops/s	🟢▲ +0.28%	2.47%
fork_join_lockless/1	kqueue	ops_per_sec	47.00K ops/s	46.91K ops/s	🔴▼ -0.19%	0.08%
fork_join_lockless/16	kqueue	ops_per_sec	7.74K ops/s	7.74K ops/s	🔴▼ -0.24%	1.34%
fork_join_lockless/4	kqueue	ops_per_sec	21.90K ops/s	21.88K ops/s	🟢▲ +0.19%	0.53%
fork_join_lockless/64	kqueue	ops_per_sec	2.03K ops/s	2.02K ops/s	🔴▼ -0.11%	2.48%
nested/16	kqueue	ops_per_sec	2.00K ops/s	1.99K ops/s	🔴▼ -0.08%	1.79%
nested/4	kqueue	ops_per_sec	7.06K ops/s	7.06K ops/s	🔴▼ -0.06%	0.92%
nested_lockless/16	kqueue	ops_per_sec	2.01K ops/s	2.00K ops/s	🟢▲ +0.69%	2.41%
nested_lockless/4	kqueue	ops_per_sec	7.15K ops/s	7.13K ops/s	🔴▼ -0.50%	1.25%

http_server

Benchmark	Backend	Metric	Base	Head	Δ	Noise
concurrent/1	kqueue	ops_per_sec	49.59K ops/s	49.25K ops/s	🔴▼ -0.42%	0.38%
concurrent/16	kqueue	ops_per_sec	127.3K ops/s	129.0K ops/s	🟢▲ +1.19%	0.69%
concurrent/32	kqueue	ops_per_sec	127.0K ops/s	128.8K ops/s	🟢▲ +1.50%	1.41%
concurrent/4	kqueue	ops_per_sec	88.02K ops/s	88.77K ops/s	🟢▲ +0.87%	0.55%
multithread/1	kqueue	ops_per_sec	126.8K ops/s	128.4K ops/s	🟢▲ +0.84%	1.29%
multithread/16	kqueue	ops_per_sec	127.3K ops/s	127.5K ops/s	🔴▼ -0.14%	0.67%
multithread/2	kqueue	ops_per_sec	176.9K ops/s	177.2K ops/s	🟢▲ +0.42%	1.77%
multithread/4	kqueue	ops_per_sec	160.8K ops/s	161.5K ops/s	🟢▲ +0.38%	1.01%
multithread/8	kqueue	ops_per_sec	132.6K ops/s	132.0K ops/s	🔴▼ -1.40%	1.62%
single_conn	kqueue	ops_per_sec	49.65K ops/s	49.34K ops/s	🔴▼ -0.47%	0.27%
single_conn_lockless	kqueue	ops_per_sec	49.80K ops/s	49.51K ops/s	🔴▼ -0.59%	0.34%

local_socket_latency

Benchmark	Backend	Metric	Base	Head	Δ	Noise
concurrent/1	kqueue	latency_mean_ns	2.06 µs	2.06 µs	🔴▼ -0.10%	0.24%
concurrent/16	kqueue	latency_mean_ns	30.64 µs	30.53 µs	🟢▲ +0.16%	0.32%
concurrent/4	kqueue	latency_mean_ns	7.66 µs	7.65 µs	🟢▲ +0.29%	0.37%
concurrent_lockless/1	kqueue	latency_mean_ns	2.02 µs	2.04 µs	🔴▼ -0.89%	0.45%
concurrent_lockless/16	kqueue	latency_mean_ns	30.13 µs	30.11 µs	🟢▲ +0.04%	0.27%
concurrent_lockless/4	kqueue	latency_mean_ns	7.54 µs	7.54 µs	🔴▼ -0.05%	0.37%
pingpong/1	kqueue	latency_mean_ns	2.05 µs	2.06 µs	🔴▼ -0.54%	0.49%
pingpong/1024	kqueue	latency_mean_ns	2.16 µs	2.16 µs	🟢▲ +0.14%	0.34%
pingpong/64	kqueue	latency_mean_ns	2.05 µs	2.06 µs	🔴▼ -0.07%	0.31%
pingpong_lockless/1	kqueue	latency_mean_ns	2.02 µs	2.03 µs	🔴▼ -0.42%	0.64%
pingpong_lockless/1024	kqueue	latency_mean_ns	2.13 µs	2.13 µs	🔴▼ -0.29%	0.14%
pingpong_lockless/64	kqueue	latency_mean_ns	2.02 µs	2.03 µs	🔴▼ -0.23%	0.33%

local_socket_throughput

Benchmark	Backend	Metric	Base	Head	Δ	Noise
bidirectional/1024	kqueue	bytes_per_sec	1.22 GB/s	1.22 GB/s	🔴▼ -0.01%	0.32%
bidirectional/1048576	kqueue	bytes_per_sec	3.85 GB/s	3.86 GB/s	🟢▲ +0.41%	0.38%
bidirectional/16384	kqueue	bytes_per_sec	3.85 GB/s	3.87 GB/s	🔴▼ -0.07%	1.12%
bidirectional/262144	kqueue	bytes_per_sec	3.86 GB/s	3.86 GB/s	🟢▲ +0.44%	0.94%
bidirectional/4096	kqueue	bytes_per_sec	2.91 GB/s	2.92 GB/s	🟢▲ +0.15%	0.65%
bidirectional/65536	kqueue	bytes_per_sec	3.86 GB/s	3.86 GB/s	🔴▼ -0.60%	1.05%
bidirectional_lockless/1024	kqueue	bytes_per_sec	1.23 GB/s	1.23 GB/s	🔴▼ -0.34%	0.44%
bidirectional_lockless/1048576	kqueue	bytes_per_sec	3.90 GB/s	3.92 GB/s	🟢▲ +0.13%	0.83%
bidirectional_lockless/16384	kqueue	bytes_per_sec	3.92 GB/s	3.91 GB/s	🔴▼ -0.11%	0.41%
bidirectional_lockless/262144	kqueue	bytes_per_sec	3.91 GB/s	3.91 GB/s	🟢▲ +0.22%	0.81%
bidirectional_lockless/4096	kqueue	bytes_per_sec	2.96 GB/s	2.94 GB/s	🔴▼ -0.45%	0.38%
bidirectional_lockless/65536	kqueue	bytes_per_sec	3.92 GB/s	3.90 GB/s	🔴▼ -0.42%	0.75%
unidirectional/1024	kqueue	bytes_per_sec	1.19 GB/s	1.19 GB/s	🟢▲ +0.09%	0.44%
unidirectional/1048576	kqueue	bytes_per_sec	3.46 GB/s	3.46 GB/s	🟢▲ +0.01%	0.28%
unidirectional/16384	kqueue	bytes_per_sec	3.46 GB/s	3.47 GB/s	🟢▲ +0.26%	0.36%
unidirectional/262144	kqueue	bytes_per_sec	3.46 GB/s	3.46 GB/s	🔴▼ -0.47%	0.70%
unidirectional/4096	kqueue	bytes_per_sec	2.68 GB/s	2.69 GB/s	🟢▲ +0.40%	0.47%
unidirectional/65536	kqueue	bytes_per_sec	3.46 GB/s	3.47 GB/s	🟢▲ +0.40%	0.34%
unidirectional_lockless/1024	kqueue	bytes_per_sec	1.20 GB/s	1.20 GB/s	🔴▼ -0.18%	0.47%
unidirectional_lockless/1048576	kqueue	bytes_per_sec	3.50 GB/s	3.51 GB/s	🟢▲ +0.32%	0.31%
unidirectional_lockless/16384	kqueue	bytes_per_sec	3.52 GB/s	3.53 GB/s	🟢▲ +0.38%	0.41%
unidirectional_lockless/262144	kqueue	bytes_per_sec	3.51 GB/s	3.51 GB/s	🟢▲ +0.07%	0.34%
unidirectional_lockless/4096	kqueue	bytes_per_sec	2.73 GB/s	2.72 GB/s	🔴▼ -0.40%	0.26%
unidirectional_lockless/65536	kqueue	bytes_per_sec	3.52 GB/s	3.52 GB/s	🟢▲ +0.16%	0.51%

socket_latency

Benchmark	Backend	Metric	Base	Head	Δ	Noise
concurrent/1	kqueue	latency_mean_ns	18.80 µs	18.78 µs	🔴▼ -0.05%	0.66%
concurrent/16	kqueue	latency_mean_ns	120.23 µs	120.56 µs	🔴▼ -0.15%	0.29%
concurrent/4	kqueue	latency_mean_ns	43.35 µs	43.49 µs	🔴▼ -0.54%	1.06%
concurrent_lockless/1	kqueue	latency_mean_ns	18.66 µs	18.76 µs	🔴▼ -0.71%	0.45%
concurrent_lockless/16	kqueue	latency_mean_ns	119.68 µs	120.07 µs	🔴▼ -0.53%	0.69%
concurrent_lockless/4	kqueue	latency_mean_ns	43.07 µs	43.07 µs	🔴▼ -0.08%	0.45%
pingpong/1	kqueue	latency_mean_ns	15.45 µs	15.50 µs	🔴▼ -0.15%	0.80%
pingpong/1024	kqueue	latency_mean_ns	18.90 µs	18.95 µs	🔴▼ -0.12%	0.30%
pingpong/64	kqueue	latency_mean_ns	18.72 µs	18.77 µs	🔴▼ -0.29%	0.50%
pingpong_lockless/1	kqueue	latency_mean_ns	15.40 µs	15.47 µs	🔴▼ -0.22%	0.79%
pingpong_lockless/1024	kqueue	latency_mean_ns	18.93 µs	18.94 µs	🟢▲ +0.03%	0.35%
pingpong_lockless/64	kqueue	latency_mean_ns	18.67 µs	18.74 µs	🔴▼ -0.33%	0.50%

socket_throughput

Benchmark	Backend	Metric	Base	Head	Δ	Noise
bidirectional/1024	kqueue	bytes_per_sec	410.5 MB/s	408.7 MB/s	🟢▲ +0.10%	0.79%
bidirectional/1048576	kqueue	bytes_per_sec	7.78 GB/s	7.54 GB/s	🔴▼ -0.77%	4.87%
bidirectional/16384	kqueue	bytes_per_sec	3.99 GB/s	3.98 GB/s	🔴▼ -0.39%	0.62%
bidirectional/262144	kqueue	bytes_per_sec	8.74 GB/s	8.68 GB/s	🔴▼ -0.63%	1.47%
bidirectional/4096	kqueue	bytes_per_sec	1.49 GB/s	1.49 GB/s	🔴▼ -0.01%	0.65%
bidirectional/65536	kqueue	bytes_per_sec	7.53 GB/s	7.51 GB/s	🟢▲ +0.80%	1.91%
bidirectional_lockless/1024	kqueue	bytes_per_sec	414.1 MB/s	410.3 MB/s	🔴▼ -0.63%	0.79%
bidirectional_lockless/1048576	kqueue	bytes_per_sec	7.79 GB/s	7.68 GB/s	🔴▼ -1.00%	6.56%
bidirectional_lockless/16384	kqueue	bytes_per_sec	4.00 GB/s	4.01 GB/s	🟢▲ +0.82%	0.86%
bidirectional_lockless/262144	kqueue	bytes_per_sec	8.72 GB/s	8.79 GB/s	🟢▲ +1.21%	1.63%
bidirectional_lockless/4096	kqueue	bytes_per_sec	1.50 GB/s	1.50 GB/s	🔴▼ -0.46%	0.75%
bidirectional_lockless/65536	kqueue	bytes_per_sec	7.59 GB/s	7.66 GB/s	🟢▲ +1.05%	1.18%
multithread/2	kqueue	bytes_per_sec	7.68 GB/s	7.57 GB/s	🟢▲ +1.04%	3.08%
multithread/4	kqueue	bytes_per_sec	3.83 GB/s	3.71 GB/s	🔴▼ -1.62%	5.39%
multithread/8	kqueue	bytes_per_sec	3.30 GB/s	3.29 GB/s	🔴▼ -3.22%	5.20%
unidirectional/1024	kqueue	bytes_per_sec	317.7 MB/s	316.7 MB/s	🔴▼ -0.45%	1.05%
unidirectional/1048576	kqueue	bytes_per_sec	9.04 GB/s	8.78 GB/s	🔴▼ -0.33%	5.69%
unidirectional/16384	kqueue	bytes_per_sec	3.71 GB/s	3.70 GB/s	🔴▼ -0.25%	0.50%
unidirectional/262144	kqueue	bytes_per_sec	8.82 GB/s	8.83 GB/s	🔴▼ -0.74%	3.57%
unidirectional/4096	kqueue	bytes_per_sec	1.23 GB/s	1.24 GB/s	🟢▲ +0.08%	1.71%
unidirectional/65536	kqueue	bytes_per_sec	7.26 GB/s	7.23 GB/s	🟢▲ +0.20%	2.01%
unidirectional_lockless/1024	kqueue	bytes_per_sec	327.9 MB/s	327.2 MB/s	🔴▼ -0.02%	0.38%
unidirectional_lockless/1048576	kqueue	bytes_per_sec	9.03 GB/s	8.85 GB/s	🟢▲ +0.14%	5.30%
unidirectional_lockless/16384	kqueue	bytes_per_sec	3.73 GB/s	3.74 GB/s	🟢▲ +0.19%	1.17%
unidirectional_lockless/262144	kqueue	bytes_per_sec	8.85 GB/s	8.79 GB/s	🔴▼ -1.19%	0.98%
unidirectional_lockless/4096	kqueue	bytes_per_sec	1.25 GB/s	1.24 GB/s	🔴▼ -0.57%	1.42%
unidirectional_lockless/65536	kqueue	bytes_per_sec	7.21 GB/s	7.20 GB/s	🔴▼ -0.54%	2.13%

Run & raw JSON artifacts · flag rule: |median Δ| > max(2%, 3×CV of the noisier side) · advisory only

sgerbino added 3 commits October 8, 2026 18:47

feat: add object_ref and object_pool

7f3aadb

refactor: recycle io_object impls through object_pool

c154684

refactor: collapse IOCP wrappers and recycle impls through object_pool

b822510

refactor: unify timer recycling onto object_pool

15ae22d

sgerbino force-pushed the pr/mem-opts branch from a4ca5b3 to 383919e Compare October 8, 2026 17:02

sgerbino added 8 commits October 8, 2026 19:41

test: add recycled-impl poison, churn exercise, and zero-allocation gate

26879b9

docs: keep @ref and its target on one line in ip_address

23880ff

Vale's docstring lint reads the wrapped line's leading identifier as bare prose; rewrapping keeps the reference intact.

sgerbino force-pushed the pr/mem-opts branch from 383919e to 23880ff Compare October 8, 2026 18:06

sgerbino marked this pull request as ready for review October 8, 2026 19:12

Navigation Menu

Repository navigation

Zero-allocation steady state for I/O object and operation churn - #389

Zero-allocation steady state for I/O object and operation churn#389
sgerbino wants to merge 12 commits into
cppalliance:developfrom
sgerbino:pr/mem-opts

sgerbino commented Oct 8, 2026 •

edited

Loading

Uh oh!

cppalliance-bot commented Oct 8, 2026 •

edited

Loading

Uh oh!

cppalliance-bot commented Oct 8, 2026 •

edited

Loading

Uh oh!

github-actions Bot commented Oct 8, 2026 •

edited

Loading

Uh oh!

This branch has not been deployed

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

2 participants

Conversation

sgerbino commented Oct 8, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Changes

Verification

Uh oh!

cppalliance-bot commented Oct 8, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

cppalliance-bot commented Oct 8, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

github-actions Bot commented Oct 8, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Benchmark report

⚠️ Flagged

Uh oh!

This branch has not been deployed

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

2 participants

sgerbino commented Oct 8, 2026 •

edited

Loading

cppalliance-bot commented Oct 8, 2026 •

edited

Loading

cppalliance-bot commented Oct 8, 2026 •

edited

Loading

github-actions Bot commented Oct 8, 2026 •

edited

Loading