branchrust_reboot/main/serverprofiler_codeapicancel
17 Commits over 0 Days - ∞cph!
Update: update ServerProfiler.Core bins to Release
- built on 2a311df
Tests: ran all server profiler unit tests
Update: add profile.ImmediateModeEnabled feature flag
- codegen + unit test
Turns off all managed-side logic for new API
Tests: ran unit tests
Update: introduce export interval (profile.ExportIntervalS, defaults to 30m) + ability to reset the interval (profile.ResetExportInterval)
- codegen and extra unit tests
Tests: unit tests
Bugfix: ProfileExporter.JSON can now export 0-frame main thread profiles
Test: ran previously failing unit tests, checked their exported files - all's gud
Update: immediate mode API improvements
- debug windows binary built from 2a311dfb
- ScopeRecorder automatically exports to json and cleans up recorder state
- added RecordScopeIfSlow(..., TimeSpan, ...) API, same as above except exports only if there was a delay
- updated unit tests since some scenarios are now impossible
Need to fix export next, wrap it with a couple server vars and update to release bins - then it's done
Tests: ran most of the unit tests (stress tests skipped as they would overflow with export tasks)
Update: ServerProfiler.Core - various improvements and fixes
- debug windows binary from f50b4fc9
- change internal constants to be more sensible (assumed worker thread count 4 -> 32, max recorders 64 -> 16, max alloc 1GB -> 512MB)
- bugfix for not cleaning up dead thread state when running immediate mode recording
- MemoryPool no longer allocates from heap as a fallback when it's over capacity
Think core lib is done enough for now, gonna move to finishing rust side
Tests: ran unit tests
Update: add TextContextExhaustionTest
- reduce TestDeferCleanup internal loop count to 8 from 16 (as was still possible to starve the pool)
Tests: ran unit tests, pass (got local unsubmitte fixes)
Update: add TestDeferCleanup test
Works, but discovered that I forgot to clean up threads in ServerProfiler.Core, so I'm starving out the pool
Tests: ran new test
Update: minor changes
- MakeScopeRecording -> RecordScope
- fail starting to record if profiler isn't initialized
Tests: unit tests
Update: ServerProfiler.Core - MemoryReadings are now implemented via MemoryPool
- debug windows bins from 47635f61
- ABI break for MemoryData
Tests: unit tests + 10x of StressTestImmediateCaptureMT
Update: ServerProfiler.Core - use memory pooling
- debug windows binary built from af80ca2c
- this fixes/reduces occurance of the MT race
- also reduces capture overhead (at least in debug, 2.2s -> 0.75ms)
- added MPMCQueue license file
Need to revive support for MemoryReadings, will do that next.
Tests: unit tests + StressTestImmediateCaptureMT 10 times
Update: ServerProfiler.Core - replaced my own MPSC queue with a third-party MPMC queue
- debug windows binary from 268ce0c3
Needed to add memory pooling, my own version couldn't handle non-integral types
Tests: unit tests
Update: add StressTestImmediateCaptureMT test
It smashes the profiler from all 20 threads doing allocations and calling methods, while main tries to record just 1 method in a loop. This triggers heap corruption - think allocation pooling would solve this.
Tests: ran extra unit test - it failed drastically
Bugfix: kind-of-fix the thread race with Immediate Mode API (late profiler callback might be in progress as we're releasing resources, leading to invalid write)
- built debug binaries from a3312fa9
- Added a mini stress test for main thread only, needs multithreading to fully validate
I need to optimize internals a bit, to avoid allocation overhead
Tests: ran unit tests on repeat 10 times - no issues
Update: first working version of immediate capture API
- binaries built from b3a39bd2 commit
Has a bug with a race, will fix next
Tests: passes unit tests
Update: blockout Immediate-Record API
- Added unit tests to validate usage
Tests: ran unit tests, has expected failures
