Re: `page allocation failure: order:0` with ixgbe under high load

[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index]

 



On Mon, Aug 14, 2017 at 7:18 AM, Paul Menzel <pmenzel@xxxxxxxxxxxxx> wrote:
> Dear Cong,
>
>
> Thank you for the response.
>
>
> On 08/11/17 19:51, Cong Wang wrote:
>
>> On Fri, Aug 11, 2017 at 8:36 AM, Paul Menzel <pmenzel@xxxxxxxxxxxxx>
>> wrote:
>>> Or should some parameters be tuned?
>>>
>>> ```
>>> $ more /proc/sys/vm/min*
>>> ::::::::::::::
>>> /proc/sys/vm/min_free_kbytes
>>> ::::::::::::::
>>> 39726
>>
>>
>>
>> Can you try to increase this? Although it depends on your workload,
>> 38M seems too small for a host with 96+G memory.
>
>
> Increasing the value to 128 MB did not get rid of the warning. With 256 MB
> we were unable to reproduce the warning.


Interesting. I wonder if we should just increase the hard-coded cap
(64M) for the default min_free_kbytes, or make it configurable at
compile-time.


>
>>> ::::::::::::::
>>> /proc/sys/vm/min_slab_ratio
>>> ::::::::::::::
>>> 5
>>> ::::::::::::::
>>> /proc/sys/vm/min_unmapped_ratio
>>> ::::::::::::::
>>> 1
>>> ```
>>>
>>> There is quite some information about this on the WWW [1], but some
>>> suggest
>>> that with recent Linux kernels, this shouldn’t happen, as memory get
>>> defragmented.
>>
>>
>>
>> On the other hand, the allocation order is 0 anyway. ;)
>
>
> Right. Coherent(?) memory is not needed for an order of 0.
>
> In our case the memory is mainly occupied by the disk(?) buffer/cache, and
> not the real program. So there is plenty available. Shouldn’t the Linux
> kernel be able to deal with such situations, or is this exactly the use
> case, which the parameter `min_free_kbytes` is for?

Well, for atomic memory allocations, we can't wait for these memory
to drain or reclaim, I think this is why min_free_kbytes exits.

Atomic allocations are heavily used by networking, so the 64M cap
is really not enough for a heavily loaded network server with a
fast NIC.

But I am not at all a MM expert. ;)

--
To unsubscribe, send a message with 'unsubscribe linux-mm' in
the body to majordomo@xxxxxxxxx.  For more info on Linux MM,
see: http://www.linux-mm.org/ .
Don't email: <a href



[Index of Archives]     [Linux ARM Kernel]     [Linux ARM]     [Linux Omap]     [Fedora ARM]     [IETF Annouce]     [Bugtraq]     [Linux OMAP]     [Linux MIPS]     [eCos]     [Asterisk Internet PBX]     [Linux API]
  Powered by Linux