ZGC heap size and RSS counters
per.liden at oracle.com
Mon Dec 11 11:19:17 UTC 2017
On 2017-12-11 11:12, Aleksey Shipilev wrote:
> On 12/11/2017 10:59 AM, Per Liden wrote:
>> On 2017-12-11 10:36, Aleksey Shipilev wrote:
>>> On 12/11/2017 10:14 AM, Per Liden wrote:
>>>> On 2017-12-11 09:55, Aleksey Shipilev wrote:
>>>>> Hi there,
>>>>> I'm trying ZGC on trivial workloads, and I have a question
>>>>> about footprint. The workload runs with -Xms16g -Xms16g, but
>>>>> the RSS figures are at least 3x larger:
>>>>> VmPeak: 18256721392 kB VmSize: 18256721392 kB VmLck:
>>>>> 0 kB VmPin: 0 kB VmHWM: 50729036 kB VmRSS:
>>>>> 50729036 kB RssAnon: 369700 kB RssFile: 27688 kB
>>>>> RssShmem: 50331648 kB
>>>>> Is this because ZGC maps the same physical space with
>>>>> multiple virtual mappings? Or is it a bug?
>>>> The kernel's RSS accounting is flaky at best, and varies
>>>> depending on if you're using small or large pages (and it can
>>>> also vary depending on which kernel version you're using).
>>>> On Linux/x86_64, we map the heap in three different locations.
>>>> When using small pages, you'll typically see that the same
>>>> physical page will incorrectly be accounted for three times
>>>> instead of once. On the other hand, when using large pages,
>>>> you'll typically see a different behavior, as it's accounted to
>>>> the hugetlbfs inode and not the process.
>>>> In summary, it's not a bug in ZGC, but more a limitation in
>>>> Linux's accounting.
>>> Understood, that's what I thought. Do you think that is the
>>> problem in lieu of pervasive use of containers that
>>> allocate/limit resources based on RSS?
>> If RSS limits are used in a container, then I'd argue that the
>> kernel better get the accounting right, otherwise those limits is
>> fairly useless wouldn't you say? In the kernel's defense, it is
>> gradually getting better in this area.
> I agree that's kernel's job to account this properly. But I am also
> concerned about the practicalities with real deployments on current
> kernels :( Shenandoah is also about do to double-mapping for related
Interesting. Do you also plan to do some kind of colored pointers or do
you have some other use case for multi-mapping in Shenandoah?
> reasons, and it would gradually come to the same trouble. I was
> wondering if you have observed problems with ZGC running in
> containers that shed more light on this concern.
We haven't observed problems so far, but in all honesty ZGC has had
limited exposure to deployments outside of Oracle and Intel,
so I'm not sure to what degree we've just been lucky so far.
Should it be a real problem, there are various solutions and workarounds
to pick from. For example, in such environments ZGC could run in what we
call "colorless roots"-mode (not yet implemented), which would remove
the need for multi-mapping all together at the expense of an extra "and
reg,imm8" instruction in the load barrier.
> Thanks, -Aleksey
More information about the zgc-dev