How much overhead does the oneagent take?

I'll paste the responses with information from dev directly form a support case where this was asked:

The overhead of the Agent (regarding response time, CPU, memory & network traffic) depends on so many factors and their correlations that it's just impossible to estimate it upfront.
Such factors include, but are not limited to e.g.:

• from the application perspective:
  • used technologies (Java, .NET, PHP, Apache, node.js, ...)
  • involved frameworks
  • requests per second (or some other load indicator)
  • general CPU consumption
  • number of concurrently executed threads
  • amount of remote communication and IO interaction
  • error handling, e.g. number of thrown exceptions
  • complexity of the code in general
• from the environment perspective:
  • operating system and version
  • hardware, e.g. CPU
  • any potential virtualization in place
  • disk & network throughput
  • inter-process mutexes and other dependencies
• from the Agent & configuration perspective:
  • type of Agent (OS Agent, Java Agent, Network Agent, Log Agent, ...)
  • number of placed sensors & instrumented methods
  • level of detail of captured information

Measuring the overhead is also non-trivial, as the Agent itself is part of the whole system and can as such not easily measure itself. The only really reliable way to measure the overhead is and always will be to make reasonable load tests with an external tool including reproducible load and warmup/cooldown phases executed multiple times with and without the Agents.
However, over the years we investigated multiple approaches to try to estimate and limit the Agent overhead. Some of those turned out to be effective (e.g. around Auto-Sensor overhead limitations) and are implemented in the Agent code itself. Also, we're always trying to make the Agents even more robust also in terms of overhead.
In addition to that we're always selecting our default configurations to be as low-overhead and as production-ready as possible while still providing valuable information.
To ensure this we execute automated performance tests on a daily basis which will instantly show us if something got better or worse.

___

so here's what we do in terms of overhead measurement and reduction:

• in Dynatrace & AppMon Special Agents (OneAgent and Classic Agent)
  • check overhead of Auto-Sensors in the process and adapt it to not exceed fixed 1-3% (in Dynatrace, depending on Agent type) or configurable 1%-20% (in AppMon)
  • if network is too slow for the captured data, creation of new PurePaths will be reduced and finally data will be discarded
  • also some Sensors will reduce data capturing when certain limits are reached (e.g. Exception Sensor -> limit stacktrace capturing)
• in Dynatrace Network Agent
  • turning off the network data collection as soon as CPU reaches certain limit (5% or so)

I can't give a solid number, but the overhead is kept so that there is no impact on the response time of the application. The Watchdog is also Monitoring the overhead of the One Agent and will for example minimize Network Monitoring if the Overhead goes above 5% CPU.