[dpdk-users] occasionally ~300us delay
stephen at networkplumber.org
Wed Feb 28 16:15:20 CET 2018
On Wed, 28 Feb 2018 10:27:28 +0100
Gaëtan Rivet <gaetan.rivet at 6wind.com> wrote:
> On Wed, Feb 28, 2018 at 12:03:30AM -0500, Yuliang Li wrote:
> > Also, such pause happens regularly--every 0.5s.
> > On Wed, Feb 28, 2018 at 12:01 AM, Yuliang Li <yuliang.li at yale.edu> wrote:
> > > Hi all,
> > >
> > > I am using dpdk to generate packets. Specifically, I have a list of
> > > <packet Pi, time Ti>, meaning I want to send Pi at Ti. I wait for Ti by while
> > > (rte_get_tsc_cycles()<Ti). However, sometimes I see a delay of around
> > > 300us of sending packets, i.e., the sending time of some Pi is around
> > > Ti+300us. It looks like the program just pauses by 300us, and then resumes.
> > > Does anyone know why?
> > >
> > > Thanks,
> I would reduce entropy and gather more data.
> You can isolate your process (i.e. numactl), if possible not on core 0, disable
> hyper-threading, disable CPU frequency scaling.
> Then, running your app:
> # grep ctxt_sw /proc/$(pidof your-dpdk-app)/status
> If you have context switches, then you have other tasks throwing off your measures.
> The rte-intr-thread is notorious for having bad affinities and triggering context
> switches, and can be used for alarms (thus, possible regular occurence).
Sounds like SMI
More information about the users