[dpdk-users] rte_segments: hugepages are not in contiguous memory

[tom.barbette at ulg.ac.be]

Hi,

I had strange change in performances running the same test multiple time, using the system for other things in between (among them, loading and unloading Netmap, which scatters the memory pretty well). It was a very simple forwarding test taking packets on 4 interface and sending them back on the opposite interface. The only difference between each run I could find was the increasing memory scattering. Other strange performance issues included a shift in the performance curve of throughput according to packet size.

No explanation though...

Tom

----- Mail original -----
De: "Sergio Gonzalez Monroy" <sergio.gonzalez.monroy at intel.com>
À: "tom barbette" <tom.barbette at ulg.ac.be>, "Andriy Berestovskyy" <aber at semihalf.com>
Cc: "Renata Saiakhova" <Renata.Saiakhova at oneaccess-net.com>, "users" <users at dpdk.org>
Envoyé: Mardi 4 Octobre 2016 16:09:29
Objet: Re: [dpdk-users] rte_segments: hugepages are not in contiguous memory

Hi folks,

In theory, there shouldn't be any performance difference between having 
a mempool allocated from a single memseg (given the use the same number 
of hugepages) versus multiple memsegs
as it is all done on mempool creation/setup and each mbuf has its own 
phys address.

Tom, I cannot think of a reason why you would have higher memory access 
for having scatter hugapages vs contig hugepages.
Any details on the test you were running?

Sergio

On 04/10/2016 13:02, tom.barbette at ulg.ac.be wrote:
> There is a noticeable performance drop with more scattering of the huge pages.
>
> I did not measure any difference accurately but I ended up rebooting my DUT between each performance test because the pages get scattered with time and re-launch of the DPDK application instead of the whole machine, because the tests showed higher memory access cost each time I re-launched the application.
>
> Tom
>
> ----- Mail original -----
> De: "Andriy Berestovskyy"<aber at semihalf.com>
> À: "Renata Saiakhova"<Renata.Saiakhova at oneaccess-net.com>
> Cc: "Sergio Gonzalez Monroy"<sergio.gonzalez.monroy at intel.com>, "users"<users at dpdk.org>
> Envoyé: Mardi 4 Octobre 2016 13:27:23
> Objet: Re: [dpdk-users] rte_segments: hugepages are not in contiguous	memory
>
> Renata,
> In theory 512 contiguous 2MB huge pages might get transparently
> promoted to a single 1GB "superpage" and single TLB entry, but I am
> not even sure if it is implemented in Linux...
>
> So, I do not think there will be any noticeable performance difference
> between contiguous and non-contiguous 2MB huge pages. But you better
> measure it to make sure ;)
>
> Regards,
> Andriy
>
> On Tue, Oct 4, 2016 at 12:48 PM, Renata Saiakhova
> <Renata.Saiakhova at oneaccess-net.com>  wrote:
>> Hi Andriy,
>>
>> thanks for your reply. I guess that contiguous memory is requested because
>> of the performance reasons. Do you know if I can expect a noticeable
>> performance drop using non-contiguous memory?
>>
>> Renata
>>
>>
>> On 10/04/2016 12:13 PM, Andriy Berestovskyy wrote:
>>> Hi Renata,
>>> DPDK supports non-contiguous memory pools, but
>>> rte_pktmbuf_pool_create() uses rte_mempool_create_empty() with flags
>>> set to zero, i.e. requests contiguous memory.
>>>
>>> As a workaround, in rte_pktmbuf_pool_create() try to pass
>>> MEMPOOL_F_NO_PHYS_CONTIG flag as the last argument to
>>> rte_mempool_create_empty().
>>>
>>> Note that KNI and some PMDs in 16.07 still require contiguous memory
>>> pools, so the trick might not work for your setup. For the KNI try the
>>> DPDK's master branch which includes the commit by Ferruh Yigit:
>>>
>>> 8451269 kni: remove continuous memory restriction
>>>
>>> Regards,
>>> Andriy
>>>
>>>
>>> On Tue, Oct 4, 2016 at 11:38 AM, Renata Saiakhova
>>> <Renata.Saiakhova at oneaccess-net.com>  wrote:
>>>> Hi Sergio,
>>>>
>>>> thank you for your quick answer. I also tried to allocate 1GB hugepage,
>>>> but
>>>> seems kernel fails to allocate it: previously I've seen that
>>>> HugePages_Total
>>>> in /proc/meminfo is set to 0, now - kernel hangs at boot time (don't know
>>>> why).
>>>> But anyway, if there is no way to control hugepage allocation in the
>>>> sense
>>>> they are in contiguous memory there is only way to accept it and adapt
>>>> the
>>>> code that it creates several pools which in total satisfy the requested
>>>> size.
>>>>
>>>> Renata
>>>>
>>>>
>>>> On 10/04/2016 10:27 AM, Sergio Gonzalez Monroy wrote:
>>>>> On 04/10/2016 09:00, Renata Saiakhova wrote:
>>>>>> Hi all,
>>>>>>
>>>>>> I'm using dpdk 16.04 (I tried 16.07 with the same results) and linux
>>>>>> kernel 4.4.20 in a virtual machine (I'm using libvirt framework). I
>>>>>> pass a
>>>>>> parameter in kernel command line to allocate 512 hugepages of 2 MB at
>>>>>> boot
>>>>>> time. They are successfully allocated. When an application with dpdk
>>>>>> starts
>>>>>> it calls rte_pktmbuf_pool_create() which in turns requests internally
>>>>>> 649363712 bytes.  Those bytes should be allocated from one of
>>>>>> rte_memseg.
>>>>>> rte_memsegs describes contiguous portions of memory (both physical and
>>>>>> virtual) built on hugepages. This allocation fails, because there are
>>>>>> no
>>>>>> rte_memsegs of this size (or bigger). Further debugging shows that
>>>>>> hugepages
>>>>>> are allocated in non-contiguous physical memory and therefore
>>>>>> rte_memsegs
>>>>>> are built respecting gaps in physical memory.
>>>>>> Below are the sizes of segments built on hugepages (in bytes)
>>>>>> 2097152
>>>>>> 6291456
>>>>>> 2097152
>>>>>> 524288000
>>>>>> 2097152
>>>>>> 532676608
>>>>>> 2097152
>>>>>> 2097152
>>>>>> So there are 5 segments which includes only one hugepage!
>>>>>> This behavior is completely different to what I observe with linux
>>>>>> kernel
>>>>>> 3.8 (used with the same application with dpdk) - where all hugepages
>>>>>> are
>>>>>> allocated in contiguous memory.
>>>>>> Does anyone experience the same issue? Could it be some kernel option
>>>>>> which can do the magic? If not, and kernel can allocated hugepages in
>>>>>> non-contiguous memory how dpdk is going to resolve it?
>>>>>>
>>>>> I don't think there is anything we can do to force the kernel to
>>>>> pre-allocate contig hugepages on boot. If there was, we wouldn't need to
>>>>> do
>>>>> all this mapping sorting and grouping we do on DPDK
>>>>> as we would rely on the kernel giving us pre-allocated contig hugepages.
>>>>>
>>>>> If you have plenty of memory one possible work around would be to
>>>>> increase
>>>>> the number of default hugepages so we are likely to find more contiguous
>>>>> ones.
>>>>>
>>>>> Is using 1GB hugepages a possibility in your case?
>>>>>
>>>>> Sergio
>>>>>
>>>>>> Thanks in advance,
>>>>>> Renata
>>>>>>
>>>>> .
>>>>>
>