Add OpenTelemetry performance benchmark spec (#748)

Documented the performance benchmark requirement for language libraries. The most important metrics like throughput/CPU/memory are included.

CHANGELOG.md

@@ -9,6 +9,8 @@ release.
 
 New:
 
+- Add performance benchmark specification
+  ([#748](https://github.com/open-telemetry/opentelemetry-specification/pull/748))
 - Enforce that the Baggage API must be fully functional, even without an installed SDK.
   ([#1103](https://github.com/open-telemetry/opentelemetry-specification/pull/1103))
 - Rename "Canonical status code" to "Status code"

specification/performance-benchmark.md

@@ -0,0 +1,70 @@
+# Performance Benchmark of OpenTelemetry API
+
+This document describes common performance benchmark guidelines on how to
+measure and report the performance of OpenTelemetry SDKs.
+
+The goal of this benchmark is to provide a tool to get the basic performance
+overhead of the OpenTelemetry SDK for given events throughput on the target
+platform.
+
+## Benchmark Configuration
+
+### Span Configuration
+
+- No parent `Span` and `SpanContext`.
+- Default Span [Kind](./trace/api.md#spankind) and
+  [Status](./trace/api.md#set-status).
+- Associated to a [resource](overview.md#resources) with attributes
+  `service.name`, `service.version` and 10 characters string value for each
+  attribute, and attribute `service.instance.id` with a unique UUID. See
+  [Service](./resource/semantic_conventions/README.md#service) for details.
+- 1 [attribute](./common/common.md#attributes) with a signed 64-bit integer
+  value.
+- 1 [event](./trace/api.md#add-events) without any attributes.
+- The `AlwaysOn` sampler should be enabled.
+- Each `Span` is created and immediately ended.
+
+### Measurement Configuration
+
+For the languages with bootstrap cost like JIT compilation, a warm-up phase is
+recommended to take place before the measurement, which runs under the same
+`Span` [configuration](#span-configuration).
+
+## Throughput Measurement
+
+### Create Spans
+
+Number of spans which could be created and exported via OTLP exporter in 1
+second per logical core and average number over all logical cores, with each
+span containing 10 attributes, and each attribute containing two 20 characters
+strings, one as attribute name the other as value.
+
+## Instrumentation Cost
+
+### CPU Usage Measurement
+
+With given number of span throughput specified by user, or 10,000 spans per
+second as default if user does not input the number, measure and report the CPU
+usage for SDK with both default configured simple and batching span processors
+together with OTLP exporter. The benchmark should create an out-of-process OTLP
+receiver which listens on the exporting target or adopts existing OTLP exporter
+which runs out-of-process, responds with success status immediately and drops
+the data. The collector should not add significant CPU overhead to the
+measurement. Because the benchmark does not include user processing logic, the
+total CPU consumption of benchmark program could be considered as approximation
+of SDK's CPU consumption.
+
+The total running time for one test iteration is suggested to be at least 15
+seconds. The average and peak CPU usage should be reported.
+
+### Memory Usage Measurement
+
+Measure dynamic memory consumption, e.g. heap, for the same scenario as above
+CPU Usage section with 15 seconds duration.
+
+## Report
+
+### Report Format
+
+All the numbers above should be measured multiple times (suggest 10 times at
+least) and reported.