[j-nsp] Distributing OSPF load on MX80
As mentioned in the thread on OSPF packet drops, I have an MX80 dropping OSPF packets during every commit, after adding ~1500 VLAN interfaces. The major load seems to be ppmd, not rpd. On larger MX's, it is apparently possible to distribute ppmd processing to the line cards. Does that work on the MX80? Alternative, is BFD cheap on an MX80? If I turn on BFD, I could set the OSPF hello timers longer than the current 10 seconds. Of course that is no good if BFD just makes even more work for the already-busy routing engine. /Benny ___ juniper-nsp mailing list juniper-nsp@puck.nether.net https://puck.nether.net/mailman/listinfo/juniper-nsp
Re: [j-nsp] Distributing OSPF load on MX80
29.11.2012, Benny Amorsen wrote: Alternative, is BFD cheap on an MX80? If I turn on BFD, I could set the OSPF hello timers longer than the current 10 seconds. Of course that is no good if BFD just makes even more work for the already-busy routing engine. AFAIK, at least as of 11.something, BFD was handled by RE on MX80, not the host-CPU like it is on the big MXes. Looks like it's because the host-CPU on MX80 is quite less quick (marketing way of reading this is it's more power and heat efficient thus more suitable for mobile applications, which is what MX80 was first primary designed for :). The issue has been discussed here about a year ago: https://puck.nether.net/pipermail/juniper-nsp/2011-March/019245.html I think it's worth to check if any hellows are dropped by the control plane protection policers on the way to the RE. ___ juniper-nsp mailing list juniper-nsp@puck.nether.net https://puck.nether.net/mailman/listinfo/juniper-nsp
Re: [j-nsp] Distributing OSPF load on MX80
On (2012-11-29 20:34 +0400), Pavel Lunin wrote: AFAIK, at least as of 11.something, BFD was handled by RE on MX80, not the host-CPU like it is on the big MXes. Looks like it's because the host-CPU on MX80 is quite less quick (marketing way of reading this is I suppose host-CPU means PFE/LC CPU? MX80 has pq3 8544, MPC2 has pq3 8548. MPC2 needs to talk to two trios, MX80 only one. So it does not feel like MX80 PFE CPU is underpowered compared to big brothers. -- ++ytti ___ juniper-nsp mailing list juniper-nsp@puck.nether.net https://puck.nether.net/mailman/listinfo/juniper-nsp
Re: [j-nsp] Distributing OSPF load on MX80
Pavel Lunin plu...@senetsy.ru writes: AFAIK, at least as of 11.something, BFD was handled by RE on MX80, not the host-CPU like it is on the big MXes. Looks like it's because the host-CPU on MX80 is quite less quick (marketing way of reading this is it's more power and heat efficient thus more suitable for mobile applications, which is what MX80 was first primary designed for :). Fair enough, it seems that moving to BFD would not help. Unless BFD is significantly cheaper for the CPU than full-blown OSPF hellos, of course. The issue has been discussed here about a year ago: https://puck.nether.net/pipermail/juniper-nsp/2011-March/019245.html I think it's worth to check if any hellows are dropped by the control plane protection policers on the way to the RE. Thank you, I will check that. /Benny ___ juniper-nsp mailing list juniper-nsp@puck.nether.net https://puck.nether.net/mailman/listinfo/juniper-nsp
Re: [j-nsp] Distributing OSPF load on MX80
2012/11/29 Saku Ytti s...@ytti.fi On (2012-11-29 20:34 +0400), Pavel Lunin wrote: AFAIK, at least as of 11.something, BFD was handled by RE on MX80, not the host-CPU like it is on the big MXes. Looks like it's because the host-CPU on MX80 is quite less quick (marketing way of reading this is I suppose host-CPU means PFE/LC CPU? MX80 has pq3 8544, MPC2 has pq3 8548. MPC2 needs to talk to two trios, MX80 only one. So it does not feel like MX80 PFE CPU is underpowered compared to big brothers. From the frequency point of view — yes. Not sure my knowlege is enough for a full-scale Steve Jobs' style megaherz-myth holywar, but looks like the 8548's built-in ethernet can be the key: http://www.freescale.com/webapp/sps/site/taxonomy.jsp?code=PCPPCMPC85XX Well' don't get me wrong, I don't know the right anwer, I only heard some rumors that MX80 linecard host subsystem is weeker that that of the MPCs. It might be totally wrong though. IMHO, a good field way to check whether it's of the same capability would be to test inline Trio IPFIX performance on MX80. Flow export is done by this CPU. I've been wondering how powerful MX80 is in inline JFlow since Juniper announced this feature, but didn't have an opportunity to test it. -- Regards, Pavel ___ juniper-nsp mailing list juniper-nsp@puck.nether.net https://puck.nether.net/mailman/listinfo/juniper-nsp
Re: [j-nsp] Distributing OSPF load on MX80
I've been using IPFIX on a few MX80's for a while now, the only impact I've seen on the RE CPU is that it can spike to 100% during a commit, if the router also has a full BGP table. Otherwise the RE sits at 6%. Using the default Jflow on the MX80's was horrible, the RE CPU would sit around 70% most of the time, and if doing a 1:1 sampling rate around 25% of the flows were lost. -- Dicko. On 30 November 2012 06:45, Pavel Lunin plu...@senetsy.ru wrote: 2012/11/29 Saku Ytti s...@ytti.fi On (2012-11-29 20:34 +0400), Pavel Lunin wrote: AFAIK, at least as of 11.something, BFD was handled by RE on MX80, not the host-CPU like it is on the big MXes. Looks like it's because the host-CPU on MX80 is quite less quick (marketing way of reading this is I suppose host-CPU means PFE/LC CPU? MX80 has pq3 8544, MPC2 has pq3 8548. MPC2 needs to talk to two trios, MX80 only one. So it does not feel like MX80 PFE CPU is underpowered compared to big brothers. From the frequency point of view — yes. Not sure my knowlege is enough for a full-scale Steve Jobs' style megaherz-myth holywar, but looks like the 8548's built-in ethernet can be the key: http://www.freescale.com/webapp/sps/site/taxonomy.jsp?code=PCPPCMPC85XX Well' don't get me wrong, I don't know the right anwer, I only heard some rumors that MX80 linecard host subsystem is weeker that that of the MPCs. It might be totally wrong though. IMHO, a good field way to check whether it's of the same capability would be to test inline Trio IPFIX performance on MX80. Flow export is done by this CPU. I've been wondering how powerful MX80 is in inline JFlow since Juniper announced this feature, but didn't have an opportunity to test it. -- Regards, Pavel ___ juniper-nsp mailing list juniper-nsp@puck.nether.net https://puck.nether.net/mailman/listinfo/juniper-nsp ___ juniper-nsp mailing list juniper-nsp@puck.nether.net https://puck.nether.net/mailman/listinfo/juniper-nsp
Re: [j-nsp] Distributing OSPF load on MX80
On (2012-11-30 08:08 +0800), Simon Dixon wrote: I've been using IPFIX on a few MX80's for a while now, the only impact I've seen on the RE CPU is that it can spike to 100% during a commit, if the router also has a full BGP table. If you use inline IPFIX export, it's in trio, and should not tax PFE CPU or RE CPU pretty much at all. -- ++ytti ___ juniper-nsp mailing list juniper-nsp@puck.nether.net https://puck.nether.net/mailman/listinfo/juniper-nsp