cancel
Showing results for 
Search instead for 
Did you mean: 
cancel
2532
Views
5
Helpful
13
Replies

HW label limit on ASR9K platform

Adam Vitkovsky
Level 3
Level 3

Does anyone know exactly what memory is the L2-overwrite including MPLS labels stored on please?

And what is the max MPLS label limit for ASR9K's Typhoon and Tomahawk LCs please? (might RSP or a version of code be a limiting factor in available label space)

Is “show controllers np struct Label-UFIB summary” the right command to check please?

Is the Trident/Typhoo/Tomahawk available label space split into locally allocated labels and received labels, or is it just one shared space please?

What happens if one runs out of available label space –any early warnings please?

 

Thank you very much

 

 

 

adam

adam
13 Replies 13

Aleksandar Vidakovic
Cisco Employee
Cisco Employee

Information required for HW forwarding is stored in the RAM used by the Network Processor (NP). You can see the utilisation of this address space by running:

sh cef platform resource summary location <location>

You can find the Typhoon and Tomahawk scale limits in any of the recent BRKARC-2003 slides from Cisco Live.

Thank you very much Aleksandar,

It’s been quite a while since we last talked about NPU internals :)

 

Is the 1M label space divided into space available for locally allocated labels and space available for received labels, or it is indeed just one common space please?

 

Regarding the memory type where lookup information is stored,

 

Does ASR9k NPU use the RLDRAM for route/mac lookups please?

 

Does CRSX and NCS6000 nPowerX1e use both DPE(nDRAM Processing Engine) and EPE (Embedded DPE) to store FIB forwarding tables please?

DPE is 8xATRIS memory for External Resource 1GB

EPE is SDRAM memory inside the nPower X1e NPU itself 8MB

 

 

Thank you

 

adam

 

netconsultings.com

::carrier-class solutions for the telecommunications industry::

 

adam

hi Adam,

indeed. ;)

The 1M label space is not divided into local and received. It's all one space.

Search memory on Typhoon and Tomahawk is DDR3. I can't speak for CRS and NCS.

Are you asking because of performance expectations or is there some other reason. If we understand what are you looking for, maybe we can help you better.

regards,

/Aleksandar

Hi Aleksandar,

 

I have just realized that it’s not mentioned in the scaling slides whether these scaling numbers 1M-labels, 2M-macs, 4/2M-v4/v6-routes are per NPU or per LC (I assume per NPU) –or do the Tomahawk slices share some common memories (or memory spaces for that matter)?

 

 

I’m interested in the nominal scalability and performance of especially the Tomahawk cards, (and how they compare to nPowerX1e based cards and competition)

Also I’d like to confirm the nominal pps rate budget for Tomahawk NPU please –is it 150Mpps(per direction) please?

 

Thank you very much

 

adam

adam

hi Adam,

are you looking into the recent BRKSPG-2003 from Cisco Live? This should be all explained there.

The hard limits (1M labels, 2M MAC, 4M/2M v4/v6 routes) are per NP. However, since the content of respective tables is synchronised between NPs, the same scale limit applies at LC level. If selective VRF download is disabled (and it is disabled by default), this is also a per-chassis scale.

Tomahawk NPU can process 150Mpps with fast track features. Packets pass through the same pipeline regardless of whether they come from fabric or network interface. So if you have bidir traffic this will be 75Mpps in each direction.

/Aleksandar

Hi Aleksandar,

 

Sorry for the late reply,

Yeah excellent point –the routing info is synced between the NPs by default so it’s per chassis limit :)

 

 

 

Regarding the pps rate budget on Tomahawk,

If it is 150Mpps indeed, then I think the budget actually has to be at least 300Mpps per NPU

i.e. 150Mpps per direction- if the ratio of traffic coming from WAN and Fabric to the NPU is 1:1.

 

My understanding is that these are per direction:

Typhoon:                             44Mpps

nPower X1e G1:               300Mpps

 

 

my ruler for comparing the NPU capabilities is: 29.761Mpps (two directions)

-of course one doesn’t really need the NPU to provide line-rate for all packet sizes but the closer it gets the more headroom one has for enabling features.

 

Line-rate 10Gbps@L2(64B)@L1(84B) is 29.761Mpps (two directions) <--this is essentially the fastest rate at which the packets can arrive at the PFE, cause the packets have to be spaced 20B apart on the wire. (but still bit of an assumption, because packets from Fabric may arrive at faster/slower rate –depending on many factors (the FIA and NP overhead imposed on the frame, backpressure state, the FIA to Fabric BW, …)

Thank you 

 

adam

 

netconsultings.com

::carrier-class solutions for the telecommunications industry::

 

adam

Typhoon's theoretical limit is 45Mpps per direction

Tomahawk is 200 (or something) Mpps per direction

once you add features you drop incrementally.

see feature order of operation to get some ballpark numbers of what feature (combo) will cost you what.

xander

That’s more like it :)

That would be 20Mpps budget for both directions per 10GE entity,

 

I got worried because BRKARC-2003 page 61 say 150Mpps (maybe that was meat to be 250mpps?)

 

Yes features drain the available pps budget, that’s why I’d like the NPU to be as close as possible to the theoretical line-rate max of ~30Mpps. Of course would not need that to forward 200Gbps of 64B framse back-to-back, but rather use it as budget for features while trying to get as close as possible to linerate for my backbone’s ussual packetsize mix and peak traffic rates.

 

 

adam

adam

Hi Aleksander,

but what is it the max label stack depth on these machines? In particular do you have documentation about the ASR9k , CRS-X and the GSR12k? HOw many labels are supported in a stack (bottom to top) ?

 

Many thanks

Alessio

hi Alessio,

the depth of the MPLS stack on transient nodes doesn't really mater. Forwarding decision depends only on the top MPLS label. For load-balancing on asr9k we'll stop looking for IP header if there are more than 8 labels in the stack.

If you're asking how many labels can we push in Segment Routing solution, I think it was 8 but will be increased to 12. This is for asr9k.

/Aleks

Can you please share any documentation about this? Also, what about without SR?

If you're asking how many labels can we push in Segment Routing solution, I think it was 8 but will be increased to 12. This is for asr9k.


Can you share any document about this?

Getting Started

Find answers to your questions by entering keywords or phrases in the Search bar above. New here? Use these resources to familiarize yourself with the community: