0

Our trading network is very latency sensitive. We use Arista switches for the latency sensitive parts, and we know that the latency there is 600 ns.

However, we have a Cisco 3560 in part of the network and we need a rough idea of that switch's latency. 5 us, 50 us?

Ted Graham
  • 171
  • 1
  • 2
  • 5
  • 3
    Have you considered asking Cisco? – James Yale Jan 05 '12 at 15:42
  • Unlike with Arista, I don't have a good technical contact at Cisco. What do you think is the best way to ask them? – Ted Graham Jan 05 '12 at 15:43
  • @TedGraham do you have a SmartNet support contract with Cisco? (If you're trading, i'd be surprised if you didn't.) – Tom O'Connor Jan 05 '12 at 15:53
  • We don't have a support contract with Cisco. We use Aristas for the stuff that matters, but it turns out I have some data flowing through the Cisco where it would be helpful to know the latency. I wrote to the tac@cisco support address, we'll see if I get a response. – Ted Graham Jan 05 '12 at 15:59
  • 2
    If some trading data is flowing through the Ciscos, you really should have a smartnet contract. – petrus Jan 05 '12 at 21:37
  • Emails to tac@cisco are answered with a request for support contract numbers and the like. We could get a SmartNet contract, but it looks like we will be replacing the switch with an Arista instead. – Ted Graham Jan 05 '12 at 21:39
  • I assume you actually mean the Cisco 3560E and not the 3560 which is a 100Mbit/s switch. – pfo Jan 05 '12 at 22:02

2 Answers2

3

Well a 3650 seems to offer a fairly linear amount of forwarding bandwidth based on the number of ports (i.e. 2.7Mpps for an 8 port, 3.2 for a 12 port, 6.4 for a 24 port etc.) based on 64k packets. So this lets us work out the following;

3.2Mpps / 12 ports = 266k 64k pps or 133k 128k pps (roughly).

1 second / 133k = ~7 microseconds per 128k packet.

Does that sound about right?

Chopper3
  • 101,299
  • 9
  • 108
  • 239
  • That sounds low to me, but I don't really know. I've heard 40 microseconds thrown around as a Cisco switch latency number. Does it affect things that my connections are all 1 gig? – Ted Graham Jan 05 '12 at 17:52
  • Yeah, but latency comparisons without including message size is pretty meaningless. – pfo Jan 05 '12 at 22:00
  • 1
    40us sounds right for older gen Cisco stuff, 20 for current Nexus based hardware for forwarding only. Talking about 1 byte ping pong type RTT test done using MPI. – pfo Jan 05 '12 at 22:05
  • @Chopper3, You mention that "forwarding bandwidth based on the number of ports (i.e. 2.7Mpps for an 8 port, 3.2 for a 12 port, 6.4 for a 24 port etc.) based on 64k packets." Can you please specify what does 64K means . Is it a packet size ? – user1389735 Apr 06 '17 at 07:54
  • No sorry, this was five years ago, I'll have taken that from some Cisco documentation. – Chopper3 Apr 06 '17 at 09:48
3

In this thread over at cisco-nsp mailing-list, there was a talk about using 3560 with financial flows that you could find useful.

It doesn't answer your question about latency, but explains why using this switch with financial flows is not such a good idea.

With bursty traffic like financial data, you'll probably see output drops, meaning that the output buffers of the switch are too small. You could set up some CoS, but CoS is only the way to tell the switch which packet to drop, but that is not a solution with critical data that has to be delivered and not dropped.

I'd like also to quote Joshua's answer:

The buffer sizes on the small form-factor cisco switches are too small for anything but a quiet office lan. The exception is the 4948.

As others have mentioned, you can play with mls qos/buffer tuning, but, i personally would recommend planning an upgrade to something that is more capable.

There is also two cisco papers regarding switch performances, but they don't talk about latency:

Cisco switching performance product sheet.

Cisco Switch Guide.

And the router equivalent.

petrus
  • 5,297
  • 26
  • 42