extreme network latency
rnyberg at murmeldjur.se
Mon Oct 17 13:17:31 PDT 2016
Yes, that was it. Many thanks!
Should I just use polling, which works fine, or is there something one
can do about the interrupt issue?
On 17 October 2016 at 22:05, Matthew Dillon <dillon at backplane.com> wrote:
> That kinda sounds like an interrupt issue, in which case I suggest turning
> polling on for both interfaces. ifconfig <blah> polling ought to do it. If
> that fixes the problem, then it is definitely interrupt-related.
> On Mon, Oct 17, 2016 at 12:52 PM, Richard Nyberg <rnyberg at murmeldjur.se>
>> Thanks again for your suggestions.
>> Actually it's much stranger than I thought. While troubleshooting I
>> had this configuration:
>> df (em0) -> switch <- desktop
>> No other devices or network interfaces were connected. In this
>> configuration there was no problem at all with latency. I then plugged
>> in the cable with internet acces like below:
>> internet <- (re0) df (em0) -> switch <- desktop
>> In this configuration the latency problems immediately showed. The fun
>> thing is that when I unplugged the re0 interface again the em0
>> interface stopped responding at all, until I put the cable back to
>> re0. Then em0 was back but with latency problems.
>> Another data point is that while I downloaded a large file at speed
>> from the internet via df to my desktop in the above configuration and
>> pinged from the desktop to df at the same time, the latency problems
>> were gone. Until the download was finished and they started again.
>> On 16 October 2016 at 19:14, Matthew Dillon <dillon at backplane.com> wrote:
>> > Look for a packet loop on the interface. Use tcpdump on the interface
>> > to
>> > see if there are excess packets being generated from somewhere. There
>> > are
>> > numerous things that can blow up a LAN. The most common being that a
>> > switch
>> > port is wired to loop back into the LAN.
>> > -Matt
>> > On Sun, Oct 16, 2016 at 9:17 AM, Justin Sherrill
>> > <justin at shiningsilence.com>
>> > wrote:
>> >> On Sun, Oct 16, 2016 at 11:49 AM, Richard Nyberg
>> >> <rnyberg at murmeldjur.se>
>> >> wrote:
>> >> > Thanks!
>> >> >
>> >> > Here are some more datapoints.
>> >> I think the only constant at this point is the internal interface on
>> >> the DragonFly system. If you hook the em0 interface that's currently
>> >> internal on the DragonFly machine up to your Internet link (i.e.
>> >> reverse which interface is internal or external), does it still
>> >> perform badly?
>> >> If it doesn't work well, then that interface is bad. I'd be
>> >> surprised, cause I've seen network ports go bad very rarely, but it's
>> >> possible. Plus, I don't have any other ideas.
More information about the Users