About the observability of memory usage in Java

[hhunng]

Hi everyone,

I’m profiling a Java app running in Kubernetes and want to see current memory usage per function.

Right now, I only get allocation metrics such as:

• memory-alloc_in_new_tlab_bytes
• memory-alloc_in_new_tlab_objects

These metrics seem to show allocation rate, not current memory consumption. Meanwhile, the pod’s actual process memory usage (RSS / container memory) is much higher.

Questions:

• Can Java profilers show current memory consumption per function, not just recent allocations?
• If not, how should allocation metrics be correlated with real process memory usage?
• Is this a limitation of the JVM or the profiler?

Thanks!

[hhunng]

here is my observe

[jake-kramer]

Hi @hhunng, note that this question is geared toward individual profilers capabilities. The profilers Pyroscope integrates with do not provide "current" memory consumption per function. As I understand it, this metric is too expensive to compute for any continuous profiler.

The alloc profiling metrics that are currently collected answer the question "where is my memory being allocated?" at the sampled frequency. This correlates to the more general question "which functions are currently using my memory?".

Note that, depending on how your JVM is configured, RSS will show a high watermark since the JVM doesn't always return memory back to the OS.

[hhunng]

i understood, thank you