Concurrent performance improvements; add performance settings field; support IgnoreCollisions

[jmacd]

Description: This streamlines the concurrent access path.

Link to tracking Issue:

Fixes #377.

Documentation: New section in lightstep/sdk/metric/README.md.

Testing: New test added. New benchmarks added, see below. Results are no change for standard configuration, potential 10% improvement with new option.

Performance claim evaluated by raising the Concurrent test load factor:

diff --git a/lightstep/sdk/metric/internal/syncstate/sync_test.go b/lightstep/sdk/metric/internal/syncstate/sync_test.go
index 47abff3..da42016 100644
--- a/lightstep/sdk/metric/internal/syncstate/sync_test.go
+++ b/lightstep/sdk/metric/internal/syncstate/sync_test.go
@@ -111,9 +111,9 @@ func testSyncStateConcurrency[N number.Any, Traits number.Traits[N]](t *testing.
 	// tests are needed.
 	const (
 		numReaders  = 2
-		numRoutines = 10
-		numAttrs    = 10
-		numUpdates  = 1e5
+		numRoutines = 20
+		numAttrs    = 30
+		numUpdates  = 5e6
 	)
 
 	var traits Traits

Performance after:

=== RUN   TestSyncStateDeltaConcurrencyInt
--- PASS: TestSyncStateDeltaConcurrencyInt (2.16s)
=== RUN   TestSyncStateCumulativeConcurrencyInt
--- PASS: TestSyncStateCumulativeConcurrencyInt (3.66s)
=== RUN   TestSyncStateCumulativeConcurrencyIntFiltered
--- PASS: TestSyncStateCumulativeConcurrencyIntFiltered (2.88s)
=== RUN   TestSyncStateDeltaConcurrencyFloat
--- PASS: TestSyncStateDeltaConcurrencyFloat (2.09s)
=== RUN   TestSyncStateCumulativeConcurrencyFloat
--- PASS: TestSyncStateCumulativeConcurrencyFloat (3.68s)
=== RUN   TestSyncStateCumulativeConcurrencyFloatFiltered
--- PASS: TestSyncStateCumulativeConcurrencyFloatFiltered (2.89s)

Performance before shows each test was substantially slower:

=== RUN   TestSyncStateDeltaConcurrencyInt
--- PASS: TestSyncStateDeltaConcurrencyInt (2.87s)
=== RUN   TestSyncStateCumulativeConcurrencyInt
--- PASS: TestSyncStateCumulativeConcurrencyInt (5.86s)
=== RUN   TestSyncStateCumulativeConcurrencyIntFiltered
--- PASS: TestSyncStateCumulativeConcurrencyIntFiltered (4.03s)
=== RUN   TestSyncStateDeltaConcurrencyFloat
--- PASS: TestSyncStateDeltaConcurrencyFloat (3.09s)
=== RUN   TestSyncStateCumulativeConcurrencyFloat
--- PASS: TestSyncStateCumulativeConcurrencyFloat (5.91s)
=== RUN   TestSyncStateCumulativeConcurrencyFloatFiltered
--- PASS: TestSyncStateCumulativeConcurrencyFloatFiltered (3.84s)

Benchmarks after are about the same as below, using the default, safe configuration.

BenchmarkCounterAddNoAttrs-10                   	34563432	        34.50 ns/op	       0 B/op	       0 allocs/op
BenchmarkCounterAddOneAttr-10                   	13396021	        88.49 ns/op	      64 B/op	       1 allocs/op
BenchmarkCounterAddOneInvalidAttr-10            	 8608786	       139.1 ns/op	     128 B/op	       1 allocs/op
BenchmarkCounterAddManyAttrs-10                 	 1000000	      1084 ns/op	     569 B/op	       6 allocs/op
BenchmarkCounterAddManyInvalidAttrs-10          	  770696	      1758 ns/op	    1088 B/op	      10 allocs/op
BenchmarkCounterAddManyFilteredAttrs-10         	 1000000	      1409 ns/op	     954 B/op	       8 allocs/op
BenchmarkCounterCollectOneAttrNoReuse-10        	 2520536	       470.0 ns/op	     400 B/op	       7 allocs/op
BenchmarkCounterCollectOneAttrWithReuse-10      	 3608900	       331.6 ns/op	     136 B/op	       3 allocs/op
BenchmarkCounterCollectTenAttrs-10              	  691329	      1702 ns/op	     712 B/op	      12 allocs/op
BenchmarkCounterCollectTenAttrsTenTimes-10      	   70843	     16864 ns/op	    7128 B/op	     120 allocs/op

New benchmarks use the unsafe code path, indicating about a 10% time improvement vs the corresponding benchmark above. Also note one less allocation which is because when ignoring collisions there is no need to copy the original attribute slice.

BenchmarkCounterAddManyAttrsUnsafe-10           	 1000000	      1003 ns/op	     505 B/op	       5 allocs/op
BenchmarkCounterAddOneAttrUnsafe-10             	15834658	        75.12 ns/op	      64 B/op	       1 allocs/op
BenchmarkCounterAddManyInvalidAttrsUnsafe-10    	  794605	      1719 ns/op	     957 B/op	       9 allocs/op

Benchmarks before:

BenchmarkCounterAddNoAttrs-10                 	35009677	        34.38 ns/op	       0 B/op	       0 allocs/op
BenchmarkCounterAddOneAttr-10                 	13377559	        89.41 ns/op	      64 B/op	       1 allocs/op
BenchmarkCounterAddOneInvalidAttr-10          	 8423766	       142.4 ns/op	     128 B/op	       1 allocs/op
BenchmarkCounterAddManyAttrs-10               	 1000000	      1116 ns/op	     569 B/op	       6 allocs/op
BenchmarkCounterAddManyInvalidAttrs-10        	  779258	      1763 ns/op	    1087 B/op	      10 allocs/op
BenchmarkCounterAddManyFilteredAttrs-10       	 1000000	      1367 ns/op	     953 B/op	       8 allocs/op
BenchmarkCounterCollectOneAttrNoReuse-10      	 2531304	       471.5 ns/op	     400 B/op	       7 allocs/op
BenchmarkCounterCollectOneAttrWithReuse-10    	 3621260	       330.8 ns/op	     136 B/op	       3 allocs/op
BenchmarkCounterCollectTenAttrs-10            	  681168	      1686 ns/op	     712 B/op	      12 allocs/op
BenchmarkCounterCollectTenAttrsTenTimes-10    	   69726	     16827 ns/op	    7128 B/op	     120 allocs/op

[jmacd]

(of the 32-bit build)

panic: unaligned 64-bit atomic operation
	panic: unaligned 64-bit atomic operation

[jaronoff97]

One nit, one question. Very excited for this!! 🚀