[dpdk-dev] drops while transmitting to the kni using rte_kni_tx_burst()

[Ferruh Yigit]

On 1/17/2017 12:34 PM, Shirley Avishour wrote:
> Hi,
> can the KNI_KTHREAD_RESCHEDULE_INTERVAL decrease to lower values than 5
> usecs?? is it effective at all?

You are right, it may not be since metric is usecs, and
usecs_to_jiffies() most probably will give same value for 5 and lower.

Removing that sleep completely (CONFIG_RTE_KNI_PREEMPT_DEFAULT=n) helps
a little.

> and what is the purpose of KNI_RX_LOOP_NUM? I am not sure I understand
> what it does and what should I do with it to improve the performance.

Increasing the loop number, also increases the work/sleep ratio, and
increasing the performance a little.

> 
> thanks!
> 
> On Mon, Jan 16, 2017 at 5:43 PM, Shirley Avishour
> <shirley at imvisiontech.com <mailto:shirley at imvisiontech.com>> wrote:
> 
>     Hi,
>     KNI_KTHREAD_RESCHEDULE_INTERVAL is currently set to 5usec. how
>     should I tweak this value to get better performance?
>     and can you explain the use of KNI_RX_LOOP_NUM and what can I
>     possibly modify it to ?
> 
>     Thanks!
> 
> 
>     On Mon, Jan 16, 2017 at 4:55 PM, Ferruh Yigit
>     <ferruh.yigit at intel.com <mailto:ferruh.yigit at intel.com>> wrote:
> 
>         On 1/16/2017 2:47 PM, Shirley Avishour wrote:
>         > Hi,
>         > As I wrote the kernel thread runs on a dedicated lcore.
>         > running top while my application is running I see kni_single and the cpu
>         > usage is really low...
>         > Is there any rate limitation for transmitting to the kernel interface
>         > (since packets are being copied in the kernel).
> 
>         Yes, kind of, kernel thread sleeps periodically, with a value
>         defined by
>         KNI_KTHREAD_RESCHEDULE_INTERVAL. You can try tweaking this
>         value, if you
>         want thread do more work, less sleep J
> 
>         Also KNI_RX_LOOP_NUM can be updated for same purpose.
> 
>         >
>         >
>         > On Mon, Jan 16, 2017 at 4:42 PM, Ferruh Yigit <ferruh.yigit at intel.com <mailto:ferruh.yigit at intel.com>
>         > <mailto:ferruh.yigit at intel.com
>         <mailto:ferruh.yigit at intel.com>>> wrote:
>         >
>         >     On 1/16/2017 12:20 PM, Shirley Avishour wrote:
>         >     > Hi,
>         >     > I have an application over dpdk which is consisted of
>         the following threads
>         >     > each running on a separate core:
>         >     > 1) rx thread which listens on in a poll mode for traffic
>         >     > 2) 2 packet processing threads (for load balancing)
>         >     > 3) kni thread (which also runs on a separate core).
>         >
>         >     This is kernel thread, right? Is it bind to any specific core?
>         >     Is it possible that this thread shares the core with 2nd
>         processing
>         >     thread when enabled?
>         >
>         >     >
>         >     > the rx thread receives packets and clones them and
>         transmit a copy
>         >     to the
>         >     > kni and the other packet is sent to the packet
>         processing unit
>         >     (hashing
>         >     > over 2 threads).
>         >     > the receive traffic rate is 100Mbps.
>         >     > When working with single packet processing thread I am
>         able to get
>         >     all the
>         >     > 100Mbps towards the kni with no drops.
>         >     > but when I activate my application with 2 packet processing
>         >     threads I start
>         >     > facing drops towards the kni.
>         >     > the way I see it the only difference now is that I have
>         another
>         >     threads
>         >     > which handles an mbuf and frees it once processing is
>         completed.
>         >     > Can anyone assist with this case please?
>         >     >
>         >     > Thanks!
>         >     >
>         >
>         >
> 
> 
>