[dpdk-dev] DPDK hugepage memory fragmentation

Burakov, Anatoly anatoly.burakov at intel.com
Mon Jul 13 11:27:52 CEST 2020


On 11-Jul-20 8:51 AM, Kamaraj P wrote:
> Hello Anatoly/Bruce,
> 
> We are using the 18_11 version of DPDK and we are using igb_uio.
> The way we observe an issue here is that, after we tried multiple 
> iterations of start/stop of container application(which has DPDK),
> we were not able to allocate the memory for port during the init.
> We thought that it could be an issue of not getting continuous 
> allocation hence it fails.
> 
> Is there an API where I can check if the memory is fragmented before we 
> invoke an allocation ?
> Or do we have any such mechanism to defragment the memory allocation 
> once we exist from the application ?
> Please advise.
> 

This is unlikely due to fragmentation, because the only way for 18.11 to 
be affected my memory fragmentation is 1) if you're using legacy mem 
mode, or 2) you're using IOVA as PA mode and you need huge amounts of 
contiguous memory. (you are using igb_uio, so you would be in IOVA as PA 
mode)

NICs very rarely, if ever, allocate more than a 2M-page worth of 
contiguous memory, because their descriptor rings aren't that big, and 
they'll usually get all the IOVA-contiguous space they need even in the 
face of heavily fragmented memory. Similarly, while 18.11 mempools will 
request IOVA-contiguous memory first, they have a fallback to using 
non-contiguous memory and thus too work just fine in the face of high 
memory fragmentation.

The nature of the 18.11 memory subsystem is such that IOVA layout is 
decoupled from VA layout, so fragmentation does not affect DPDK as much 
as it has for previous versions. The first thing i'd suggest is using 
VFIO as opposed to igb_uio, as it's safer to use in a container 
environment, and it's less susceptible to memory fragmentation issues 
because it can remap memory to appear IOVA-contiguous.

Could you please provide detailed logs of the init process? You can add 
'--log-level=eal,8' to the EAL command-line to enable debug logging in 
the EAL.

> Thanks,
> Kamaraj
> 
> 
> 
> On Fri, Jul 10, 2020 at 9:14 PM Burakov, Anatoly 
> <anatoly.burakov at intel.com <mailto:anatoly.burakov at intel.com>> wrote:
> 
>     On 10-Jul-20 11:28 AM, Bruce Richardson wrote:
>      > On Fri, Jul 10, 2020 at 02:52:16PM +0530, Kamaraj P wrote:
>      >> Hello All,
>      >>
>      >> We are running to run DPDK based application in a container mode,
>      >> When we do multiple start/stop of our container application, the
>     DPDK
>      >> initialization seems to be failing.
>      >> This is because the hugepage memory fragementated and is not
>     able to find
>      >> the continuous allocation of the memory to initialize the buffer
>     in the
>      >> dpdk init.
>      >>
>      >> As part of the cleanup of the container, we do call
>     rte_eal_cleanup() to
>      >> cleanup the memory w.r.t our application. However after
>     iterations we still
>      >> see the memory allocation failure due to the fragmentation issue.
>      >>
>      >> We also tried to set the "--huge-unlink" as an argument before
>     when we
>      >> called the rte_eal_init() and it did not help.
>      >>
>      >> Could you please suggest if there is an option or any existing
>     patches
>      >> available to clean up the memory to avoid fragmentation issues
>     in the
>      >> future.
>      >>
>      >> Please advise.
>      >>
>      > What version of DPDK are you using, and what kernel driver for NIC
>      > interfacing are you using?
>      > DPDK versions since 18.05 should be more forgiving of fragmented
>     memory,
>      > especially if using the vfio-pci kernel driver.
>      >
> 
>     This sounds odd, to be honest.
> 
>     Unless you're allocating huge chunks of IOVA-contiguous memory,
>     fragmentation shouldn't be an issue. How did you determine that this
>     was
>     in fact due to fragmentation?
> 
>      > Regards,
>      > /Bruce
>      >
> 
> 
>     -- 
>     Thanks,
>     Anatoly
> 


-- 
Thanks,
Anatoly


More information about the dev mailing list