Re: [dpdk-dev] very high VIRT memory usage

["Burakov, Anatoly" <anatoly.burakov@intel.com>]

On 10-Jun-20 4:28 PM, Stephen Hemminger wrote:
> On Wed, 10 Jun 2020 10:24:44 +0100
> "Burakov, Anatoly" <anatoly.burakov@intel.com> wrote:
> 
>> On 09-Jun-20 4:35 PM, Stephen Hemminger wrote:
>>> On Tue, 9 Jun 2020 14:39:54 +0100
>>> "Burakov, Anatoly" <anatoly.burakov@intel.com> wrote:
>>>    
>>>> On 09-Jun-20 2:13 PM, Ferruh Yigit wrote:
>>>>> On 6/9/2020 1:46 PM, Burakov, Anatoly wrote:
>>>>>> On 08-Jun-20 12:03 PM, Francesco wrote:
>>>>>>> Hi all,
>>>>>>> I upgraded an old DPDK-based app which was using DPDK 17.11 to latest DPDK
>>>>>>> 20.05 and I noticed that if I look  at "top" I see that the VIRT memory
>>>>>>> taken by my application is now 256.1GB while before it was <1GB.
>>>>>>>
>>>>>>> I've seen this same behavior with also "testpmd" example... is this a known
>>>>>>> issue with latest DPDK versions?
>>>>>>> Can I tweak some setting to have VIRT memory usage more or less similar to
>>>>>>> RSS ?
>>>>>>>
>>>>>>> I forgot to add I'm working on Linux, Centos7
>>>>>>>
>>>>>>> Thanks,
>>>>>>> Francesco Montorsi
>>>>>>>      
>>>>>>
>>>>>> There was a discussion on this not too long ago, but i can't seem to
>>>>>> find it for some reason.
>>>>>
>>>>> Can it be "Big spike in DPDK VSZ" ?
>>>>> http://inbox.dpdk.org/dev/CAGxAMwD6Wtfi=C2Txwjfk0zhFvRzeqBu7mFfE8ayh=EJi2aU-A@mail.gmail.com/#t
>>>>>       
>>>>
>>>> Yes, that's the one, thanks Ferruh :)
>>>>   
>>>>>> Anyway, long story short, that's not a bug,
>>>>>> that's by design.
>>>>>>
>>>>>> Since 18.11 (or 18.05 to be precise), there is a new memory subsystem in
>>>>>> DPDK that allows growing and shrinking DPDK memory usage at runtime.
>>>>>> That means, you can start with zero hugepages preallocated, and then
>>>>>> allocate as you go, letting the memory subsystem decide how much memory
>>>>>> you need.
>>>>>>
>>>>>> The catch is that all of this hugepage memory is allocated into
>>>>>> somewhere, some virtual address space. And *that* address space is
>>>>>> preallocated at startup, to allow for secondary processes to duplicate
>>>>>> primary process's address space exactly, and allow dynamic allocation of
>>>>>> *shared* memory at runtime.
>>>>>>
>>>>>> This memory will show up in top et al. but the truth is, it's zero cost,
>>>>>> because it's anonymous memory. It isn't actually taking up any RAM. It
>>>>>> will show up in dumps (20.05 has already fixed that issue, and the fixes
>>>>>> will probably be backported to stable, including 18.11), so unless you
>>>>>> have a very specific problem, i don't think that's anything you should
>>>>>> be concerned about.
>>>
>>> The one concern is for cases like cgroup memory accounting thinking
>>> the process is huge and OOM killing it.
>>>    
>>
>> Is there any way to know the *actual* memory usage of the process (i.e.
>> not including anonymous memory)?
>>
> 
> Huge pages do not count against the normal memory in cgroup.
> There is a separate hugeTLB controller that limits that.
> 

But hugepages only get mapped when they're required - the rest of the 
memory is mapped anonymously with PROT_NONE. Would that count against 
the limits?

-- 
Thanks,
Anatoly