Switches - Fabric Capacity/Forwarding Bandwidth/Forwarding Rate

[cscott]

What is the difference between Forwarding Bandwidth and Switch Fabric Capacity? I'm comparing the Cisco 2970G-24T to the Dell PowerConnect 5224. The Cisco has a "Forwarding Bandwidth" of 24gbps. The Dell has a "Switch Fabric Capacity" of 48gbps. What is the difference between the two and how are the numbers calculated. I realize it probably has more to do with than just being able to handle 1gbps per-port(although technically it would need to be 2gbps per port, 1gbps up and down). So how are these numbers computed?

Next question: I'm assuming that when the forwarding rate is listed as 38Mpps that means Million Packets Per Second? Is that for one port, or through the whole switch? Does this number correlate directly to Switch Fabric/Forwarding Bandwidth?

I'm just trying to make sense of all these numbers. Different manufacturers call them different things, it'd be nice if there was an industry standard to use in describing switch specs. I realize the comparison of the Cisco to a Dell is probably not a fair one considering the quality/target/price point. But all management cares about are numbers..hehe.

They dont care about the people that have to support these things =).

By the way, anyone have any experience with the Dell 5224? I've used the 5212. Performs ok, I hate the web interface though and the CLI looks and feels almost exactly like IOS...but it's stripped down, big time. I'd much rather have Cisco equipment in my racks, but managers dont see much past a $2000 price gap.

[konigl · Accepted solution]

Take the maximum port speed on each Cisco switch port, add them all up. If it is equal to or less than the stated forwarding bandwidth, then the switch is considered non-blocking. This appears to be how Cisco is doing the math these days. Old Cisco math would have said 1000*2*24=48Gbps switch fabric capacity for the 2970G-24T, and 1000*2*28=56Gbps for the 2970G-24TS.

Not sure about the Dell, it sounds like a 24-port switch from its model number. Their "switch fabric capacity" sounds like they were using old Cisco math to arrive at that figure. Assuming it has 24 ports that are 10/100/1000, that would be top speed of each port, times 2 for full duplex, summed up for all ports: 1000*2*24 = 48,000Mbps or 48Gbps.

Millions of packets per second is the aggregate routing performance for the whole switch. To see how that matches up against the switch, it's time to do some more math again.

Maximum frames per second (Layer 2) on an Ethernet port is 14,880. That's a minimum-sized 64-byte frame. For Fast Ethernet, it's 14,880*10=148.8kfps, and for Gigabit Ethernet it's 14,880*100=1488kfps or 1.488Mfps.

If that frame happens to be an IP packet, then you're talking about 1.488Mpps per port on a Gigabit Ethernet port. Times 24 ports equals 35.712Mpps for the 2970G-24T. (Cisco quotes 35.7Mpps for that model, and 38.7Mpps for the 2970G-24TS; sounds like they're only expecting to actively use two of the four SFPs on that model, not sure why, will have to look into that -- because 38.688Mpps is what you get for 26 out of 28 ports running full GbE speed.) So you can effectively route IP at Layer 2 switching speeds or line-speed of the interfaces, according to these figures.

An IOS router won't even come close to that level of performance, unless you spend a lot more money on it. Then again, you can't terminate a serial T1 Frame Relay connection or a point-to-point T3 line on a switch, so I guess routers still have their uses.

Comparing Cisco to Dell is fair. Cisco is more expensive, because they incorporate a lot of management features that the cheaper competitors leave out. Pure and simple. You need fast, raw speed, cheap? Buy something else. You need to be able to do lots of cool things with that switch, that you can't do with anybody else's? Buy Cisco. If you need the features, it's a no-brainer. Otherwise, you're buying commodity products, and any switch will do. (Yea, right. Until the day when you need those features and they aren't there.)

Does the Dell do Layer 3 switching also? Basic, or full? That could account for a big part of the price gap between the two. On some Cisco switch models, for example the 24-port 3550's, you can get them with basic of full L3 software. Price difference is about $2,000 list. (Basic is static and RIP IP routing; full is the works, all the other cool protocols like OSPF, etc.)

Hope this helps.

[cscott]

Thanks for the reply! I understand it now. That's what I thought initially, but I wasnt sure if anything else was factored in or not. And hey, now I know what non-blocking means too.

I knew the Dell's were stripped down, thats what I meant when I was talking about how their CLI is *remarkably* similar to IOS, but very, very stripped down...and very clumsy to work with, it looks just like IOS and everything though. Like I said I've used their 5212, the 12 port version (you are correct in your assumption that the 5224 is a 24 port), it doesnt have any of the mac address port-security stuff on it, anything like Cisco's CMS, no port-protection or any of the bells and whistles that make my life easier. It's just a question I guess of whether or not those bells and whistles justify $2000 extra...but moreso whether I can convince my boss of that. Heh.

Anyhow, thanks for your answer! You answered my question!

-Chris

[mcardell]

Hi,

Your post is really interesting.

I have got few quetions about that.

You were speaking of Forwarding rate to define non-blocking switch. Is the same of the current link ?: http://www.cisco.com/warp/public/765/tools/quickreference/switchperformance.pdf

And in the same link Cisco use Switch Fabric to indicate the Backplane bandwidth.Why?

What is SFP?

Following your indication it seems that 2970G-24T is non-blocking and what about 2950 - 24 ports?

On the table is reported 66000000 pps per second.

Calculation: 24 ports * 148.8 kpfs * 2 (full duplex)= 7142 Kpps . Does it means that it is a blocking switches when it is configured with all ports 100Mbps full duplex?

Thanks

Matteo

[konigl]

Non-blocking switch: one that is capable of transferring the aggregate traffic load from all ports while they are running at wire speed (the actual maximum transmission rate of the link).

Switch fabric: the internal interconnection architecture that supports connectivity among the various ports of the switch. Sometimes called the "backplane".

Switch fabric (or backplane) bandwidth must be equal to or greater than the aggregate bandwidth of all ports on the switch, for a switch to be called "non-blocking".

I'm certainly no expert on this, but here's one person's perspective (mine) on the history of these figures.

Manufacturers used to quote switch fabric bandwidth as the number of ports times the bandwidth of each port. Early on, Cisco switches were at a performance disadvantage compared to competitors' products: they switched in software, while the competition did it in hardware. Hardware throughput was much faster.

To counter customer's perceptions that the Cisco switches were slower, the Cisco marketing types hit on the idea that the max. bandwidth was the same in each direction, so they doubled their figures and published that. If it was in print, it must be true; many customers bought into the new "Cisco math".

Other vendors, not to be outdone, started doing the same thing and doubled their figures. Whatever; as long as you knew how they were reaching their numbers, you could compare them fairly.

Somewhere along the way, Cisco seems to have stopped doing the "Cisco math", reverting back to the earlier way of calculating. Maybe it's because they now have switches that CAN do all their switching in hardware.

I'm not slamming Cisco, because I love their products, and make a living installing and maintaining them; just the marketing confusion they've caused.

The PDF chart that you posted a link to has some misleading figures in it. Because some seem to be calculated with the old "new math" (i.e. x2), others with the new "old math" (i.e. not x2). So I don't think it provides fair comparisons, even among Cisco's own switches. Furthermore, in some cases I think it oversimplifies, grouping whole sets of some switches under one line (e.g. 3500 XL for the 3512, 3524, 3548, and 3508G) and giving the max. Gbps of the fastest one (3508G). So I don't put too much confidence in it.

Maybe it's about time Cisco explained their math, and put together a table that specified the exact pps, Gbps, and blocking/non-blocking status of each and every switch model. (HINT, HINT.)

Let's run the numbers for one of the switches you mentioned, see how it compares with the PDF.

2970G-24T has 24 10/100/1000 ports.

24 ports * 1 Gbps per port = 24 Gbps (new "old math")

which is what the PDF quotes; I would say it's non-blocking.

Its switching performance is

24 ports * 1,488,000 pps max. per Gig port = 35,712,000 pps

which is close enough to the 35,700,000 the PDF quotes; I would also say that it's non-blocking from this figure.

OK so far. Let's try another one that you mentioned.

2950-24 has 24 10/100 ports.

24 ports * 100 Mbps per port = 2400 Mbps or 2.4 Gbps(new "old math");

even doubling it (old "new math") only gets you 4.8 Gbps, and neither of these figures shows up in the PDF, so I couldn't guess whether this switch was non-blocking or blocking without digging elsewhere on Cisco's web site for more information.

Its switching performance is

24 ports * 148,800 pps max. per Fast Ethernet port = 3,571,200 pps

which isn't even close to the 6,600,000 the PDF quotes; so again, I'd have to dig deeper for more accurate information on this switch than what's in the PDF table.

Not sure where the PDF's figures came from for this switch. Maybe it's a typo.

Moving on, let's look at a switch almost the same as the last one, but with a "G" in its model number.

2950G-24 has 24 10/100 and two Gigabit Ethernet GBIC ports.

24 ports * 100 Mbps per FE port = 2.4 Gbps (new "old math"); plus

2 ports * 1 Gbps per GE port = 2.0 Gbps (new "old math"); equals

4.4 Gbps (new "old math"); x2 (for old "new math") equals 8.8 Gbps,

which is the same figure quoted for the 2950"no G"-24 switch, but not quite the 13.6 Gbps the PDF equates with this switch. Hmmm.....

Its switching performance is

24 ports * 148,800 pps max. per FE port = 3,571,200 pps; plus

2 ports * 1,488,000 pps max. per GE port = 2,976,000 pps; equals

6,547,200 pps,

which is pretty close to the 6,600,000 the PDF quotes for the 2950"no G"-24 switch, but substantially less than the 10,100,000 pps listed for this switch. Hmmm.....

Looks like maybe they put the right numbers next to the wrong switch.

Without showing all the gory arithmetic here, it turns out that those figures shown in the PDF for the 2950G-24 are accurate...for the 2950G-48!

My point being, this PDF file may be a useful reference for someone in marketing who may not understand all the technical details, but it's so full of inaccuracies that it borders on false advertising from a technical perspective.

And to answer your question about SFP, that is short for Small Format Pluggable, and refers to the new miniature Gigabit transceivers that some Cisco switches now support instead of GBICs. The SFP modules use LC connectors instead of SC connectors, and are a little smaller than GBICs, so you can fit them into smaller spaces than a GBIC (or maybe fit more of them into the same space as a GBIC). SFPs and GBICs can communicate over fiber, but they are not interchangeable in a switch: that is, you cannot put an SFP in a GBIC opening, or a GBIC into an SFP opening.

Oh, and that x2 stuff only applies to figures calculating bits per second on the switch fabric/backplane, not packets per second of switching throughput. (Technically, it's frames per second when you're talking Layer 2 switching, and packets per second when it's Layer 3 switching. Just more confusion from that PDF.)

A blocking switch is one that cannot handle the max. traffic load of all its interfaces across its switching fabric/backplane at the same time. The 3508G was such a switch.

Hope this hasn't added to the confusion...