asmhatre
c7eccd1db5
Custom OrderProcessed event, CPU/allocation/lock workloads, a diagnostic endpoint over the JVM's live FlightRecorder state, a JUnit test using the jdk.jfr.Recording API, and nine captured recordings (flagship, mid-flight snapshot, isolated vs concurrent threshold comparisons, and a real dump+stop duplication bug) with their docs/output/*.txt transcripts.
33 lines
2.0 KiB
Plaintext
33 lines
2.0 KiB
Plaintext
The same 40-round, ~12 ms-wait workload as 07-threshold-isolated-comparison.txt,
|
|
but this time two recordings are started AT THE SAME TIME on the same JVM: one
|
|
with settings=default (JavaMonitorEnter threshold 20 ms) and one with
|
|
settings=profile (threshold 10 ms).
|
|
|
|
$ jcmd 5287 JFR.start name=cmpA settings=default
|
|
$ jcmd 5287 JFR.start name=cmpB settings=profile
|
|
$ for i in $(seq 1 40); do curl -s "localhost:8080/api/lock/contend?workers=2&holdMillis=12" -o /dev/null; done
|
|
$ jcmd 5287 JFR.stop name=cmpA filename=recordings/compare-default.jfr
|
|
$ jcmd 5287 JFR.stop name=cmpB filename=recordings/compare-profile.jfr
|
|
|
|
$ jfr summary recordings/compare-default.jfr | grep JavaMonitorEnter
|
|
jdk.JavaMonitorEnter 40 1080
|
|
$ jfr summary recordings/compare-profile.jfr | grep JavaMonitorEnter
|
|
jdk.JavaMonitorEnter 40 1080
|
|
|
|
Both recordings show all 40 events -- including in "compare-default.jfr", the one
|
|
configured with a 20 ms threshold, for waits that are only ~12 ms long. Compare
|
|
this to 07-threshold-isolated-comparison.txt, where the exact same default
|
|
settings, run ALONE with the exact same workload, captured zero of these waits.
|
|
|
|
The difference is that the two recordings were active at the same time. Duration
|
|
thresholds for events like jdk.JavaMonitorEnter are implemented as a single
|
|
instrumentation check shared by the JVM across every currently-active recording,
|
|
not a private filter per recording -- the effective threshold the running code
|
|
uses is the MINIMUM threshold requested by any active recording. A second,
|
|
unrelated recording (a colleague's ad hoc `jcmd JFR.start`, an APM agent's own
|
|
continuous recording) with a lower threshold for an event type silently widens
|
|
what every other concurrently running recording captures for that same event
|
|
type. "My recording uses the default profile" is not, by itself, enough to know
|
|
what your recording actually contains on a box where something else might also be
|
|
recording.
|