> From: Pavel Lunin [mailto:plu...@gmail.com] > Sent: Wednesday, November 01, 2017 4:08 PM > > > > There were two versions of MPC3: > > 1. MPC3 non-NG, which has a single XM buffer manager and four LU chips > (the old good ~65 Mpps LUs as in "classic" MPC1/2/16XGE old trio PFEs). > Yeah stay away from those, have fundamental design flaws apart from being hugely underpowered (especially with queuing chip enabled).
> 2. MPC3-NG which is based on exactly the same chipset as MPC5, based on > XM+XL. > Yes the building blocks are the same but the architecture of 5 is completely different, it's two XMs talking to common XL. So is it two PFEs or just one PFE? If it's two I'd suspect there to be a separate set of VOQs per XM so that each XM can backpressure to ingress LC independently. Also if the common XL gets oversubscribed does it backpressure to both XMs or just the one responsible for most of its cycles? Two discrete arbiters feeding one common XL is just asking for trouble. > > MPC4 is much like MPC3 non-NG though it has two LUs instead of four with a > new more "performant" microcode. > 4 is the other way around to 5 it has two gen1 LUs hooked up to one gen2 XM (mixing old and new hence Trio 1.5) -again some experiment or need to keep up with the competition. > > XL chip (extended LU), which is present in MPC5/6 and 2-NG/3-NG has also > multiple ALU cores (four, IIRC) but in contrast to MPC3 non-NG and MPC4 > these cores have a shared memory, so they don't suffer from some > limitations (like not very precise policers) which you can face with multi-LU > PFE architectures. > > MPC7 has a completely new single core 400G chip (also present in the > recently announced MX204 and MX10003). > This said, I find MPC4 quite not bad in most scenarios. Never had any issues, > specific to its architecture. > > P. S. Finally this choice is all about money/performance. > As I mentioned in the other discussion most of the folks don't really need to dig down to how routers actually work cause either the design does not allow the router to get clogged or in cases where ASICs get overloaded customers don’t care and understand (...oh I see you had a DDoS situation, yeah that’s understandable then that my VPN traffic experienced drops, thank you bye). In these cases there's really no point in spending more on the premium kit from cisco. adam _______________________________________________ juniper-nsp mailing list juniper-nsp@puck.nether.net https://puck.nether.net/mailman/listinfo/juniper-nsp
