Docs Home

Wine-NSPA – Validation Baselines and Comparison

This page preserves the archived RT-suite snapshots and the subsystem deltas that are still useful when reading older Wine-NSPA results.

Table of Contents

  1. Overview
  2. Archived suite snapshots
  3. Current archived boundary
  4. Historical suite deltas worth keeping
  5. Targeted subsystem A/Bs after the archive boundary
  6. Methodology boundaries
  7. References

1. Overview

The public test story has two layers:

This page keeps both. It does not try to turn every later A/B into a fake matrix version, but it also does not throw away the concrete per-test and before/after data that explains why individual carries mattered.

Public validation data has three useful shapes archived full suite versioned snapshot same runner for every default test headline totals stay meaningful per-test detail rapidmutex, philosophers, socket-io, dispatcher-burst, condvar-pi, etc. explains what changed inside the total targeted A/B used after the archive boundary preserves feature-specific evidence without minting fake suite versions Rule keep archived suite totals for boundary-to-boundary comparison, and keep the concrete per-test / per-feature numbers that explain architectural change

2. Archived suite snapshots

These are the public snapshots worth keeping as suite-level boundaries.

Snapshot Date Suite shape Result Why it matters
v4 2026-04-15 PE-only matrix, 10 tests x baseline + RT 20/20 PASS first public socket io_uring numbers and post-PI-v2 kernel fixes
v5 2026-04-15/16 PE-only matrix, 10 tests x baseline + RT 20/20 PASS SIMD / SRW-spin era with detailed per-test deltas
v7 2026-04-28 early two-layer Layer 1 native suite added; Layer 2 22/22 PASS native ntsync Layer 1 becomes part of the public suite
v8 2026-04-30 early two-layer Layer 1 3 PASS / 0 FAIL; Layer 2 24 PASS / 0 FAIL / 0 TIMEOUT dispatcher-burst enters the default PE matrix
v9-validation-default 2026-05-03 current archived two-layer default Layer 1 3 PASS / 0 FAIL / 0 SKIP; Layer 2 32 PASS / 0 FAIL / 0 TIMEOUT current archived baseline used by the public docs

Later carries are documented through targeted validators and workload A/Bs unless another full-suite archive supersedes v9-validation-default.

The last older-method whole-suite reference run before the methodology shift was v18. Its useful late-suite numbers are carried in 5.4 ntsync allocator and wait-queue follow-ons rather than being presented as a fake new archived boundary.

3. Current archived boundary

The current archived full-suite boundary is v9-validation-default (2026-05-03).

Layer Result Notes
Layer 1 native suite 3 PASS / 0 FAIL / 0 SKIP test-event-set-pi, test-channel-recv-exclusive, test-aggregate-wait
Layer 2 PE matrix 32 PASS / 0 FAIL / 0 TIMEOUT 16 default tests x baseline + rt

The archived default PE test set is:

Sync and contention I/O and bypass Timers and UI
rapidmutex socket-io nt-timer
philosophers rpc-bypass wm-timer
fork-mutex irot-bypass dispatcher-burst
cs-contention large-pages signal-recursion
ntsync-d4
ntsync-d8
ntsync-d12
condvar-pi

This is the baseline to use when a public page says “archived matrix” or “current archived validation boundary.”

4. Historical suite deltas worth keeping

4.1 v3 -> v4

The v4 archive is still useful because it was the first full public matrix after the PI-v2 ntsync fixes and the socket io_uring overlapped path landed.

Metric Earlier run v4 Read
philosophers RT max wait 1620 us 601 us -63%; PI-v2 bugfix removes boost/unboost thrash
socket-io B avg latency — 113.2 us first public overlapped io_uring result
socket-io B throughput — 8837 msg/s first public overlapped io_uring result
rapidmutex RT throughput ~262K-301K ops/s 312K ops/s stable CS-PI hot path after PI-v2 cleanup

4.2 v4 -> v5

v5 is still the densest archived example of how a micro-optimization bundle shows up in RT tests.

Metric v4 v5 Read
rapidmutex RT throughput 312K ops/s 327K ops/s +4.7%; lower CS-path copy overhead
rapidmutex RT max wait 46 us 36 us lower transition overhead
fork-mutex RT elapsed 1021 ms 948 ms -7.1%; process-start path trimmed
ntsync d4 rapid throughput 232K ops/s 259K ops/s +11.6%; lower lock transition cost
ntsync d8 RT PI avg 419 ms 201 ms -52%; the earlier reversal disappears
baseline socket-io B avg 133.2 us 104.5 us -21%; overlapped path copy cost drops
baseline socket-io B throughput 7506 msg/s 9568 msg/s +27%; same socket path improvement

4.3 Representative per-test numbers

Some individual test numbers are still worth keeping because later pages cite them directly.

Test Metric Value
condvar-pi with PI max wait 152 us
condvar-pi without PI max wait 263 us
rapidmutex RT throughput (v5) 327K ops/s
rapidmutex RT max wait (v5) 36 us
philosophers meals served 250/250
socket-io B overlapped avg (v5 baseline) 104.5 us
socket-io B overlapped throughput (v5 baseline) 9568 msg/s

5. Targeted subsystem A/Bs after the archive boundary

The public docs kept moving after v9-validation-default, but not every carry was followed by a new archived suite snapshot. These are the concrete deltas worth keeping.

5.1 Dispatcher and message-path work

Feature Before After Read
dispatcher-burst burst ops/sec 555,567 841,765 +34% / 1.5x with TRY_RECV2 burst drain
dispatcher-burst burst worst max ns 31,843,082 23,014,325 -28%
channel_dispatcher samples 14.51% 0.70% -13.81pp / -95%
main_loop_epoll samples 7.24% 2.68% -4.56pp
get_message calls / 60 s 3,880 866 -78% with empty-poll cache
get_message direct handler time 16.5 ms 2.2 ms -87%
get_message total handler time 46.8 ms 36.9 ms -21%

5.2 Shared-state and zero-time waits

Feature Before After Read
zero-time process wait ~10000 ns/poll ~144 ns/poll process_shm short-circuit
zero-time thread wait ~11940 ns/poll ~164 ns/poll thread_shm short-circuit
local sections nspa_create_mapping_from_unix_fd count 2,664 ~800 -70%; mapping work stays local
local sections total handler time 1,991 ms 1,077 ms lighter mapping path

5.3 x86_64 hot-path bundle

Metric Before After Read
user-mode samples 97K 86K -11.3%
iTLB misses 229.7M 180.8M -21.30%
dTLB misses 51.5M 42.4M -17.69%
branch misses 348.3M 308.4M -11.45%
page faults 130,349 71,754 -44.95%
NtGetTickCount dispatcher entries 3,081,551 0 inline path confirmed end-to-end

5.4 ntsync allocator and wait-queue follow-ons

The later ntsync cache-isolation and wait-queue work was validated more through targeted stress and tail numbers than through a new public full-suite archive. The best late post-1011 whole-suite reference in that family was v18, so it is worth keeping its numbers explicit instead of collapsing them into prose.

Metric post-1011 v17 post-1011 v18 later cache-isolated sample Read
philosophers RT max wait 4154 us 3039 us 857 us v18 was the tighter post-1011 run; later cache isolation tightens the tail again
philosophers worst max wait 13097 us 9963 us 1424 us same tail-tightening story
rapidmutex ops/sec 862K 254K 270K no collapse after the later allocator/cache changes
dispatcher p99 22.8M ns 17.0M ns 16.9M ns no regression on the dispatcher hot path
socket-io p99 238 us 161 us 234 us later sample remains inside the earlier variance band

These numbers are useful as a no-regression and tail-tightening read, not as a replacement for a new archived suite boundary.

6. Methodology boundaries

The methodology note should be short but explicit.

Family Snapshots What changed Comparison rule
PE-only matrix v3 through v6 one PE layer, smaller default set compare within the family
Early two-layer v7, v8 native Layer 1 added; dispatcher-burst joins the PE set compare within the family
Current archived default v9-validation-default 16-test default PE matrix with Layer 1 intact current public baseline

So:

For the current runner shape and default test list, see RT Test Harness.

7. References