[c-nsp] Understanding ASR1k / ESP40 capacity

Pete Lumbis alumbis at gmail.com
Mon Oct 6 12:47:16 EDT 2014


SIP40 supports 46Gbps through the backplane to the ESP. This is through 2x
23Gbps channels. What I don't know is how the channels connect to the
SIP/SPA. I don't think it's one channel per SPA slot, so there has to be
some sort of hashing internally. Perhaps a single channel is being
oversubscribed (this will probably be tough to determine)?

Off the top of my head the only other possibility I can think of is a
feature issue. The QFP is a processor at the end of the day, so it's
possible that the features you have enabled are taking enough to add extra
overhead, maybe due to small packet sizes (more lookups to achieve the same
throughput)? Another option is that the QFP is multicore and maybe it's
dispatching to the cores unequally? All of these are kind of out there
ideas, but just some other things you could bring up.

-Pete

On Mon, Oct 6, 2014 at 11:24 AM, Simon Lockhart <simon at slimey.org> wrote:

> Pete,
>
> Thanks for this - I'll watch that preso and see if it adds anything useful.
>
> You seem to be supporting my viewpoint, and I've also had an off-list reply
> supporting TAC's viewpoint - so I'm not sure I'm any further forwards.
>
> I'm currently working on a plan to replace the ESP40 with an ESP100 - but
> as
> the ESP100 isn't supported in the ASR1004, I'll also have to do a chassis
> swap
> to an ASR1006. My only remaining concern with this plan is whether the
> SIP40
> can really do 40Gbps. If I stick 4 * 10G SPA's into a SIP40, can I run
> those
> 10G ports at line-rate (assuming sufficient ESP capacity)?
>
> Many thanks,
>
> Simon
>
>
>
> On Sat Oct 04, 2014 at 11:56:45AM -0400, Pete Lumbis wrote:
> > It would be a single pass through the QFP. The SIP could also be a
> limiting
> > factor, but since you are split between SIPs that shouldn't be an issue.
> > The SIP 40 has 2x 40Gig lanes on the backplane. Are you doing crypto or
> > anything like that which would impact performance?
> >
> > There is a great Cisco Live preso on the ASR1k architecture that might
> help
> > you get some ammo to go back to TAC with.
> > http://d2zmdbbm9feqrf.cloudfront.net/2014/usa/pdf/BRKARC-2001.pdf
> >
> > -Pete
> >
> > On Sat, Oct 4, 2014 at 4:56 AM, Simon Lockhart <simon at slimey.org> wrote:
> >
> > > All,
> > >
> > > I'm banging my head against a brick wall trying to get sensible answers
> > > from
> > > Cisco TAC, so thought I'd ask the educated masses who may have come
> across
> > > this before...
> > >
> > > I've got a Cisco ASR1004 with RP2, ESP40, 2 * SIP40's, and 8 * 10GE
> ports.
> > >
> > > A snapshot of usage on these ports at peak is:
> > >
> > > Interface         RxBps     RxPps          TxBps     TxPps
> > > Te0/0/0   4,385,563,000   515,508    906,118,000   339,997
> > > Te0/1/0   3,942,338,000   419,696    984,150,000   358,436
> > > Te0/2/0   3,949,993,000   425,192    933,257,000   349,145
> > > Te0/3/0   4,375,526,000   512,858    873,284,000   334,751
> > > Te1/0/0   1,186,440,000   454,714  5,474,029,000   630,916
> > > Te1/1/0     622,154,000   244,056  3,181,689,000   338,190
> > > Te1/2/0     711,493,000   253,275  3,211,560,000   340,950
> > > Te1/3/0   1,218,873,000   437,195  4,831,708,000   568,488
> > >
> > > TOTAL    20,392,380,000 3,262,494 20,395,795,000 3,260,873
> > >
> > > I'm seeing throughput issues on a portchannel consisting of Te0/0/0 and
> > > Te0/3/0
> > > (it won't go over 10Gbps aggregate)
> > >
> > > Cisco TAC are telling me if I add TxBps and RxBps totals together, I
> get
> > > 40Gbps,
> > > so I've reached capacity of the QFP (i.e. ESP40).
> > >
> > > My arguement against this is that a packet which enters the router on
> > > Te0/0/0,
> > > goes through the SIP40 in slot 0, through the ESP40, through the SIP40
> in
> > > slot
> > > 1, and out through Te1/0/0 is still just one packet, so should only
> need
> > > to be
> > > counted once through the ESP, and once for each SIP. Hence, the
> throughput
> > > on
> > > the ESP is only 20.3Gbps on those numbers above.
> > >
> > > If I poll ceqfpUtilProcessingLoad by SNMP, I see peaks of around 65%,
> which
> > > would correlate with this level of throughput.
> > >
> > > I'm assuming there are others of you using this platform. What sort of
> > > throughput are you seeing? Am I right, or is the Cisco TAC engineer?
> > >
> > > TIA,
> > >
> > > Simon
> > > _______________________________________________
> > > cisco-nsp mailing list  cisco-nsp at puck.nether.net
> > > https://puck.nether.net/mailman/listinfo/cisco-nsp
> > > archive at http://puck.nether.net/pipermail/cisco-nsp/
> > >
>


More information about the cisco-nsp mailing list