How can Java developers best understand their impact on the start-up of the JVM?

[David Holmes]

cc'ing hotspot-jfr-dev

On 26/05/2022 9:45 am, Mat Carter wrote:
> The Microsoft Java Engineering Group (JEG) have been looking at how Java developers can best understand the impact of their decisions on start-up time. Specifically, how developers can identify and mitigate the dominant activities that occur during start-up, through to their app / service reaching a steady state (warmed up).
> 
> Upon investigation into the available tooling (i.e., jstat, jcmd, jmc, VisualVM, MXBeans), it seems that Performance Counters (PCs) and JDK Flight Recorder events (JFREs) are the main source of the available startup metrics that get surfaced.
> 
> Whilst the tooling does provide insight to specific areas (GC, JIT Compilation etc.), there does not seem to be one view to help a developer identify the dominant activities that are impacting start-up.
> 
> Q: Has the above been discussed before and are there any existing plans to address the outcomes that I might have missed?
> 
> Looking at the specific PCs and JFREs there are a few “missing” metrics, for example a timestamp at the very beginning entry point of the JVM. Also, it appears that the JFREs represent a subset of the PCs - does this indicate that either some PCs should be removed (or only enabled for JVM development) or that there remain JFREs to be added?
> 
> Q: Is there a current position on whether PCs and JFREs will continue to co-exist (the former supporting JMX MXBeans, et al.), or is there a plan to phase out PCs in favor of JFREs?
> 
> Q: Assuming PCs should continue to exist, do folks see the value in documenting them and making them readily accessible to the running application? Especially in the deployment scenario where running a second application or tool (to sample the PCs remotely) is not feasible. FYI - during our initial investigations we wrote some scripts to find all possible PCs. We did this because jcmd only lists instantiated PCs and the PC names are procedurally generated, making them non-trivial to search for.
> 
> Q: Do folks see value in documenting the JFREs, especially since the application can access the JFR Event stream? (There is non-official documentation available [1])
> 
> Q: While Project Leyden concerns itself primarily with improving start-up times, are there any other known activities targeting profiling of Java applications (esp. start-up)?
> 
> Answers and discussion of the above is most welcome so that our group moves forward in helping our customers, whilst staying aligned with and contributing to the OpenJDK community!
> 
> Thanks in advance
> Mat Carter
> 
> [1] https://bestsolution-at.github.io/jfr-doc/index.html