Resident/Shared memory size is showing 3 times the given heap size

[Alex Elent]

Hi Sundara,

I had a similar question in January
http://mail.openjdk.java.net/pipermail/zgc-dev/2019-January/000570.html.

@Erik Österlund <erik.osterlund at oracle.com> recommended Native Memory
Tracking (NMT) and I've been using it with success.

On Fri, Sep 20, 2019 at 12:44 AM Per Liden <per.liden at oracle.com> wrote:

> Hi,
>
> On 9/20/19 3:54 AM, Sundara Mohan M wrote:
> > Hi,
> >     We are running our server with ZGC and seeing the resident memory
> size
> > is approximately 3 times the given heap size. I am a newbie still trying
> to
> > understand basic concepts. Can someone help me understand this better
> >
> > 1. Is this expected?
>
> Yes
>
> > 2. At high level i know ZGC is having multiple view of memory region. Can
> > you explain how this can happen?
>
> Exactly. RSS (Resident set size) basically just says how much memory the
> process has mapped at the moment. It doesn't take into account that some
> of these mappings might be backed by the same memory, or might be are
> shared with other process, etc. A more interesting number, to get a
> better view of how much memory a process is actually using is PSS
> (Proportional set size). There are various tools that display PSS (smem,
> procrank, ps_mem.py, etc), but the raw data is available in
> /proc/<pid>/smaps_rollup.
>
> For example, a JVM running ZGC with a 4G heap looks like this:
>
> $ cat /proc/4509/smaps_rollup
> 00400000-7ffd472e2000 ---p 00000000 00:00 0
> [rollup]
> Rss:            12843604 kB
> Pss:             4451825 kB
> Shared_Clean:       3232 kB
> Shared_Dirty:   12582912 kB
> Private_Clean:     14380 kB
> Private_Dirty:    243080 kB
> Referenced:     12843604 kB
> Anonymous:        243048 kB
> LazyFree:              0 kB
> AnonHugePages:         0 kB
> ShmemPmdMapped:        0 kB
> Shared_Hugetlb:        0 kB
> Private_Hugetlb:       0 kB
> Swap:                  0 kB
> SwapPss:               0 kB
> Locked:                0 kB
>
> Here the multi-mapping causes the RSS to be inflated by 3x, while PSS
> shows a more accurate number reflecting the fact that the three
> different heap mappings are backed by the same memory.
>
> > 3. Also i see something like this in /proc/<pid>/maps
> >
> > ...
> > 13fff9a00000-13fff9c00000 rw-s 5f8e00000 00:12 1488924473
> >   /mnt/tmpfs/java_heap.10087 (deleted)
> > 13fff9c00000-13fff9e00000 rw-s 5fb000000 00:12 1488924473
> >   /mnt/tmpfs/java_heap.10087 (deleted)
> > 13fff9e00000-13fffa000000 rw-s 603200000 00:12 1488924473
> >   /mnt/tmpfs/java_heap.10087 (deleted)
> > 13fffa000000-13fffa200000 rw-s 612c00000 00:12 1488924473
> >   /mnt/tmpfs/java_heap.10087 (deleted)
> > 13fffa200000-13fffa400000 rw-s 61a800000 00:12 1488924473
> >   /mnt/tmpfs/java_heap.10087 (deleted)
> > 13fffa400000-13fffa600000 rw-s 326000000 00:12 1488924473
> >   /mnt/tmpfs/java_heap.10087 (deleted)
> > 13fffa600000-13fffa800000 rw-s 5d3600000 00:12 1488924473
> >   /mnt/tmpfs/java_heap.10087 (deleted)
> > ...
> > Why is this still in deleted state?
>
> The "(deleted)" state just means that the file has been unlinked. I.e.
> the file is still in use, but the directory entry
> /mnt/tmpfs/java_heap.10087 has been deleted (as it should be).
>
> > 4. Trying to get heap dump with following command
> >   jmap -heap <pid> (tried with same user as well root but it is not
> printing
> > heap)
> > Is there something changed regarding dumping heap with ZGC?
>
> IIRC, jmap -heap was removed in JDK 9, use jcmd instead:
>
> $ jcmd <pid> GC.heap_dump <filename>
>
> cheers,
> Per
>
> >
> > TIA
> > Sundar
> >
>