10G Ethernet is a Small Form-Factor Pain in the butt

I found out last week that my coworkers were all unaware of what’s been happening to Ethernet while they weren’t looking. See, Ethernet, all of the way up through Gigabit was mostly reasonable. First, they’d make some bad standards, like 10BASE5 or all of the weird not-quite-100BASE-TX standards or the time delay between 1000BASE-CX and 1000BASE-T where Gigabit Ethernet required shielded-twisted pair. Then they’d figure it out and we’d all get a reasonable standard for things that everybody was cool with, where the early-adopters generally were annoyed, but not overly so because they knew what they were getting into.

Not 10 Gigabit.

See, in prior days, they’d have switches where most of the ports had some variant on the 8P8C “RJ-45” modular jack, but then maybe a special-purpose “uplink” port. And it turns out that the uplink is sometimes better as fiber, if for no other reason than grounding and lightning suppression issues in the campus setting.

Eventually, this was standardized as the GBIC port. This way, you could plug in standard Cat5e/6/et al in with one transceiver, or you could use any variety of optical cabling, either single-mode or multi-mode, depending on what you were doing.

Then the manufacturers got greedy. GBIC was too big, so they made SFP, which became SFP+ when 10 gigabit came out. SFP and SFP+ act like standards, but they aren’t. They are hellish nasty things in pseudo-standard multi-source agreement form.

Ergo, If you have a SFP or SFP+ port on your device, it may or may not be checking for a firmware code. A generic device costs $20-30. But the genuine issue costs $300. And the generic device, programmed with the same ID as the genuine issue, works perfectly well. Yep! They’ve found new fun ways to separate you from your money with nay but a few bits of EEPROM.

Oh, but I’m not finished.

I remember Twinax when I got curious about minicomputers and mainframes. The AS/400 used it, for example. Pretty much, it’s coax meets twisted pair. So you have the coax thing where it’s a shielded thing. But it’s also twisted pair, instead of a single conductor like regular coax cable. And so they cooked up DirectAttach SFP+, to go straight from one SFP+ port to the other, without all of that encoding and transceiver stuff.

It turns out that there’s some not-necessarily-marginal advantages in using a super-fancy twinax cable instead of using 10GBASE-T. This, although I’m not the sort who gets to walk a bunch of data centers, it does sound like they are popular. If nothing else, it’s still reasonable to spend $70 on a 6 foot cable than it is to spend $30 * 2 + $10 for two transceivers and a cable. But apparently you can’t be guaranteed that your particular DirectAttach SFP+ cable will actually work with a given configuration for the same reason as everything else SFP+.

Also, if you have networking gear that you don’t totally understand, you can have two identical-looking ports that totally don’t work together. Because most of the time a SFP transceiver works in a SFP+ port, but a DirectAttach SFP+ cable definitely won’t work between a SFP and a SFP+ port.

Which, naturally, means that you can now be forced to not only buy Cisco-branded SFP+ transceivers at a ruinous markup, you can ALSO be forced to buy Cisco-branded SFP+ Direct Attach cables. Pretty much like the Thunderbolt cables, except that it makes the Thunderbolt cable look reasonably priced.

I was able to avoid all of this until last week, of course. And then I had to dig in and the resulting investigation annoyed me. Because System76 charges you $50 what you can get on Amazon for $20. But at least some hardware doesn’t check for a magic EEPROM bit on the SFP+ connector.

I’d be more annoyed if it wasn’t that all of the normal people have Wifi-only networks these days, so it’s mostly a way to screw over large companies. Then again, it fuels the argument that everyone who still needs to deal with physical infrastructure is going to be happier with OpenCompute.