[jdk21u-dev] RFR: 8293850: need a largest_committed metric for each category of NMT's output

[Severin Gehwolf]

On Wed, 3 Apr 2024 11:16:23 GMT, María Arias de Reyna <duke at openjdk.org> wrote:

> This is a backport for https://bugs.openjdk.org/browse/JDK-8293850 (need a largest_committed metric for each category of NMT's output) needed in preparation to backport https://bugs.openjdk.org/browse/JDK-8320061 ( [nmt] Multiple issues with peak accounting )
> 
> This backport adds peak usage of the memory to the NMT output to be able to detect spikes.
> 
> Example of usage before the fix:
> 
> 
> $ jcmd 160812 VM.native_memory detail | grep "mmap: reserved="
>                             (mmap: reserved=8388608KB, committed=1277952KB) 
>                             (mmap: reserved=528392KB, committed=55348KB) 
>                             (mmap: reserved=197140KB, committed=58260KB) 
>                             (mmap: reserved=40KB, committed=36KB) 
>                             (mmap: reserved=8KB, committed=8KB) 
>                             (mmap: reserved=262144KB, committed=244864KB) 
>                             (mmap: reserved=1048576KB, committed=34816KB) 
> 
> 
> After the fix (I just started the app, I wasn't really expecting any peak at this point):
> 
> 
> $ jcmd 161688 VM.native_memory detail | grep "mmap: reserved="
>                             (mmap: reserved=8388608KB, committed=507904KB, largest_committed=0KB) 
>                             (mmap: reserved=1048576KB, committed=31232KB, largest_committed=0KB) 
>                             (mmap: reserved=528392KB, committed=32836KB, largest_committed=0KB) 
>                             (mmap: reserved=197140KB, committed=43220KB, largest_committed=0KB) 
>                             (mmap: reserved=40KB, committed=36KB, largest_committed=0KB) 
>                             (mmap: reserved=8KB, committed=8KB, largest_committed=0KB) 
>                             (mmap: reserved=262144KB, committed=216640KB, largest_committed=0KB)

I think there is a problem with the original patch, which we should fix first.

src/hotspot/share/services/memReporter.cpp line 59:

> 57:   if (peak != 0) {
> 58:     output()->print(", largest_committed=" SIZE_FORMAT "%s", amount_in_current_scale(peak), scale);
> 59:   }

This seems odd. Not a problem of the backport, but of the original change. For any value lower than `scale` we'll get `0` as a result if I'm reading this right.

Your example output showed:

(mmap: reserved=8388608KB, committed=507904KB, largest_committed=0KB) 


Yet, it shouldn't be there (due to `peak != 0`). We should get this fixed before we backport.

-------------

PR Review: https://git.openjdk.org/jdk21u-dev/pull/440#pullrequestreview-1982397697
PR Review Comment: https://git.openjdk.org/jdk21u-dev/pull/440#discussion_r1553213442