Hi Saku, thanks for reply!
> > This is well know behavior and documented in several KB articles. > > However, what exactly causes this? > > I think just CPU doing something else before given time to do the ICMP > packets. Like busy running some RPD task. I also thought that it has something to do with control-plane process/thread scheduling at first. However, I would expect the RTT of ping to become even more inconsistent when less CPU-time is available for sending the ICMP "echo request" messages, but this is not the case. For example, if I pin the vCPU of virtual control plane to physical CPU core 0: $ sudo virsh vcpuinfo vcp-vmx1 VCPU: 0 CPU: 0 State: running CPU time: 142291.0s CPU Affinity: y------- $ ..and at the same time run stress-ng on the core 0: $ sudo taskset 0x00000001 stress-ng --cpu 1 --timeout 600s ..and "while :; do /usr/bin/nice -n -20 sha256 /var/tmp/2GB & /usr/bin/nice -n -20 sha256 /var/tmp/2GB & /usr/bin/nice -n -20 sha256 /var/tmp/2GB; done" on the RE, then it has no affect to RTT. Physical core CPU utilization was ~100% and RE CPU utilization seen with "show chassis routing-engine" was >80%: 1 min CPU utilization: User 83 percent Background 0 percent Kernel 16 percent Interrupt 1 percent Idle 0 percent In addition, the KB28157 says that "An ICMP ping shows intermittent latency on a directly connected link with the default Junos Class of Service (CoS) configuration" which gives an impression that there is a non-default CoS configuration where there is steady latency. This makes me think that the this occasional abnormal latency is not caused by RE. However, I could be easily wrong. > > Based on this information I changed the CoS configuration in two MX > > series routers. First I tried with vMX(runs in performance mode) and > > configured RE-sent ICMP "echo request" messages into forwarding-class > > named "ef" and associated this class with low-latency queue: > > This won't do anything, this is only applicable for forwarding-plane > traffic. For control-plane you want 'set class-of-service > host-outbound-traffic forwarding-class ef'. However this won't do > anything to give CPU scheduling priority on sending and receiving ICMP > messages. I classified control-plane ICMP traffic into "ef" forwarding-class with multifield classifier on lo0. It has the same effect as "host-outbound-traffic". Martin _______________________________________________ juniper-nsp mailing list juniper-nsp@puck.nether.net https://puck.nether.net/mailman/listinfo/juniper-nsp