[OneCollector] Add benchmark project

[CodeBlanch]

Changes

• Adds a benchmark project for OneCollectorExporter and benchmarks for LogRecord export.

Benchmarks

Method	NumberOfBatches	NumberOfLogRecordsPerBatch	EnableCompression	Mean	Error	StdDev	Allocated
Export	1	1000	False	474.8 us	6.18 us	5.16 us	857 B
Export	1	1000	True	1,683.1 us	7.87 us	6.98 us	1066 B
Export	1	10000	False	748.2 us	14.62 us	16.83 us	857 B
Export	1	10000	True	2,558.6 us	33.41 us	31.25 us	1068 B
Export	10	1000	False	4,796.1 us	74.93 us	70.09 us	8568 B
Export	10	1000	True	16,911.2 us	218.06 us	193.31 us	10674 B
Export	10	10000	False	7,151.9 us	72.75 us	68.05 us	8574 B
Export	10	10000	True	25,736.1 us	420.56 us	393.39 us	10674 B
Export	100	1000	False	48,253.9 us	672.36 us	628.93 us	85698 B
Export	100	1000	True	165,695.3 us	2,522.35 us	2,235.99 us	106760 B
Export	100	10000	False	70,449.0 us	1,261.47 us	1,179.98 us	85754 B
Export	100	10000	True	251,097.2 us	2,099.32 us	1,860.99 us	106760 B

Analysis

The chart above looks intimidating so I'll try to break it down 😄

TL;DR: The perf is where I would expect it to be and it is really good.

Here is the key bit:

NumberOfBatches	NumberOfLogRecordsPerBatch	EnableCompression	Mean	Allocated
1	1000	False	474.8 us	857 B
1	10000	False	748.2 us	857 B

The number of records in a batch does NOT increase the amount of memory allocated. That means converting a LogRecord to JSON we are not allocating any strings, boxing any value types, or anything. That is great! The 857 bytes are mostly HttpClient artifacts. A given export needs to allocate a bunch of stuff for the HttpClient API. HttpRequestMessage, HttpResponseMessage, stuff for headers. We also have some allocations for SuppressInstrumentationScope (should improve with open-telemetry/opentelemetry-dotnet#4304). These allocations don't concern me much. Exports trigger when either a batch is full or a time limit is reached. It is an occasional thing and a few allocations shouldn't add much pressure to the GC for the process. What we don't want to see is allocations growing based on the number of items.

NumberOfBatches	NumberOfLogRecordsPerBatch	EnableCompression	Mean	Allocated
1	1000	True	1,683.1 us	1066 B
1	10000	True	2,558.6 us	1068 B

When compression is used (enabled by default) it costs more CPU time and more memory is allocated but we have the same flat amount regardless of the number of records.

[codecov]

Codecov Report

Merging #1088 (d5820e6) into main (5f587d4) will not change coverage.
The diff coverage is n/a.

Additional details and impacted files

@@           Coverage Diff           @@
##             main    #1088   +/-   ##
=======================================
  Coverage   69.29%   69.29%           
=======================================
  Files         214      214           
  Lines        8358     8358           
=======================================
  Hits         5792     5792           
  Misses       2566     2566           

[Kielek]

Project scaffolding LGTM