Cut-Through Switching Isn’t A Thing Anymore

So, cut-through switching isn’t a thing anymore. It hasn’t been for a while really, though in the age of VXLAN, it’s really not a thing. And of course with all things IT, there are exceptions. But by and large, Cut-through switching just isn’t a thing.

And it doesn’t matter.

Cut-through versus store-and-forward was a preference years ago. The idea is that cut-through switching had less latency than store and forward (it does, to a certain extent). It was also the preferred method, and purchasing decisions may have been made (and sometimes still are, mostly erroneously) on whether a switch is cut-through or store-and-forward.

In this article I’m going to cover two things:

  • Why you can’t really do cut-through switching
  • Why it doesn’t matter that you can’t do cut-through switching

Why You Can’t Do Cut-Through Switching (Mostly)

You can’t do cut-through switching when you change speeds. If the bits in a frame are sent at 10 Gigabits, they need to go into a buffer before they’re sent over a 100 Gigabit uplink. The reverse is also true. You can’t stuff a frame that’s piling into an interface 10 times faster than it’s sending (though it’s not slowed down).

So any switch (which is most of them) that uses a higher speed uplink than host facing port is store-and-forward.

Just about every chassis switch involves speed changes. Even if you’re going from a 10 Gigabit port on one line card to a 10 Gigabit port on another line card, there’s a speed change involved. The line card is connected to another line card via a fabric module (typically), and that connection from line card to fabric module is via a higher speed link (typically 100 Gigabit).

There’s also often a speed change when going from one module to another, even if say the line cards were 100 Gigabit and the fabric module were 100 Gigabit, the link between them is usually a slightly higher speed in order to account for internal encapsulations. That’s right, there’s often an internal encapsulation (such as Broadcom’s HiGig2) that slightly enlarges the frames bouncing around inside of a chassis. You never see it, because the encap is added when the packet enters the switch and removed before it leaves the switch. The speed is slightly bumped to account for this, hence a slight speed change. That would necessitate store-and-forward.

As Ivan Pepelnjak noted, I got this part wrong (about Layer 3 and probably VXLAN, the other reasons stand, however).

You can’t do cut-through switching when doing Layer 3. Any Layer 3 operation involves re-writing part of the header (decrementing the TTL) and as such a new CRC for the frame that packet is encapsulated into is needed. This requires storing the entire packet (for a very, very brief amount of time).

So any Layer 3 operation is inherently store-and-forward.

Any VXLAN is store-and-forward. See above about Layer 3, as VXLAN is Layer 3 by nature.

Any time a buffer is utilized. Anytime two frames are destined for the same interface at the same time, one of them has to wait in a buffer. Any time a buffer is utilized, it’s store-and-forward. That one is hopefully obvious.

So any switch with a higher-speed uplink, or any Layer 3 operations, or when buffers are utilized, and of course when VXLAN is used, it’s automatically store-and-forward. So that covers about 99.9% of use cases in the data center. Even if your switch is capable of cut-through, you’re probably not using it.

It Doesn’t Matter That Everything Is (Mostly) Store-and-Forward

Network engineers/architects/whathaveyou of a certain age probably have it engrained that “cut-through: good” and “store-and-forward: bad”. It’s one of those persistent notions, that may have been true at one time (though I’m not sure cut-through was ever that advantageous in most cases), but no longer is. The notion that Hardware RAID is better than software RAID (isn’t not anymore), LAGs should be powers of 2 (not a requirement on most gear), Jumbo frames increase performance (miniscule to no performance benefit today in most cases), MPLS is faster (it hasn’t been for about 20 years) are just a few that come to mind.

“Cut-through switching is faster” is technically true, and still is, but it’s important to define what you mean by “faster”. Cut-through switching doesn’t increase throughput. It doesn’t make a 10 Gigabit link a 25 Gigabit link, or a 25 Gigabit link a 100 Gigabit link, etc. So when we talk about “faster”, we don’t mean throughput.

What it does is cut the amount of time a frame spends in a single switch.

With 10 Gigabit Ethernet a common speed, and most switches these days supporting 25 Gigabit, the serialization delay (the amount of time it takes to transmit or receive a frame) is miniscule. The port-to-port latency of most DC swtiches is 1 or 2 microseconds at this point. Compared to other latencies (app latency, OS network stack latency, etc.) this is imperceptible. If you halved the latency or even doubled the latency, most applications wouldn’t be able to tell the difference. Even benchmarks wouldn’t be able to tell the difference.

Cutting down the port-to-port latency was the selling point of cut-through switching. A frame’s header could be leaving the egress interface while it’s tail-end was still coming in on the ingress interface. But since the speeds are so fast, it’s not really a significant cause of communication latency. Storing the frame/packet just long enough to get the entire frame and then forward it doesn’t cause any significant delay.

From iSCSI to VMotion to SQL to whatever, the difference between cut-through and store-and-forward is unmeasurable.

Where Cut-Through Makes Sense

There are a very small number of cases where cut-through switching makes sense, most notably high-frequency trading. In these rare cases where latency absolutely needs to be cut down, cut-through can be achieved. However, there’s lots of compromises to be made.

If you want cut-through, your switches cannot be chassis. They need to be top-of-rack switches with a single ASIC (no interconnects). The interface speed needs to be the same throughout the network to avoid speed changes. You can only do Layer 2, no Layer 3 and of course no VXLAN.

The network needs to be vastly overprovisioned. Anytime you have two packets trying to leave an interface at the same time, one has to be buffered, and that will dramatically increase latency (far beyond store-and-forward latency). The packet sizes will also need to be small as to reduce latency.

Too-Long; Didn’t Read

The bad news is you probably can’t do cut-through switching. But the good news is that you don’t need to.

2 Responses to Cut-Through Switching Isn’t A Thing Anymore

  1. Ned says:

    Thank you for the article. Could u pls post a link to the discussion with Ivan that you refer to in “ As Ivan Pepelnjak noted, I got this part wrong (about Layer 3 and probably VXLAN, the other reasons stand, however).…”?

    Does cut-thru switching affect serialization delay? Thx

    • tonybourke says:

      I had thought that the IP header checksum needed to have the entire packet stored to run, but it is a just-in-time algorithm, so it can be computed in cut-through as well as store-and-forward. That was my mistake.

Leave a Reply

Fill in your details below or click an icon to log in: Logo

You are commenting using your account. Log Out /  Change )

Facebook photo

You are commenting using your Facebook account. Log Out /  Change )

Connecting to %s

This site uses Akismet to reduce spam. Learn how your comment data is processed.

%d bloggers like this: