RFR: JDK-8290870: NMT: Increase MallocSiteTable size and allocate it only when needed

[Zhengyu Gu]

On Fri, 22 Jul 2022 12:05:03 GMT, Thomas Stuefe <stuefe at openjdk.org> wrote:

> MST size is too small. It was created with a very strict footprint limit, since it needed to be created statically and unconditionally.
> 
> However, [JDK-8256844](https://bugs.openjdk.org/browse/JDK-8256844) reworked NMT initialization and now we can move MST initialization to after argument parsing. Since it is only needed with -XX:NativeMemoryTracking=detail, it makes sense to only conditionally allocate it.
> 
> And therefore we can afford to make it larger.
> 
> At the moment, it is 511 entries wide. A typical VM run (release VM, reasonably complex scenario) comes to ~3000-6000 entries. Atm the median bucket chain length is 6-8, which is a lot.
> 
> Increasing table size to ~4000 drops median bucket chain length to 1, which is nice.
> 
> Example, spring petclinic boot up:
> 
> before:
> 
> Bucket chain length distribution:
> unused: 1
> longest: 14
> median: 7
> 
> 
> with table size 4099:
> 
> Bucket chain length distribution:
> unused: 1751
> longest: 5
> median: 1

LGTM

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

Marked as reviewed by zgu (Reviewer).

PR: https://git.openjdk.org/jdk/pull/9613