Fibre Channel of Things (FCoT)

The “Internet of Things” is well underway. There are of course the hilarious bad examples of the technology (follow @internetofshit for some choice picks), but there are many valid ways that IoT infrastructure can be extremely useful.  With the networked compute we can crank out for literally pennies and the data they can relay to process, IoT is here to stay.

Hacking a dishwasher is the new hacking a gibson

But there’s one thing that these dishwashers, cars, refrigerators, Alexa’s, etc., all lack: Access to decent storage.

The storage on many IoT devices is either terrible or nonexistent. Unreliable flash storage or no storage at all. That’s why the Fibre Channel T19 working group created a standard for FCoT (Fibre Channel of Things). This gives small devices access to real storage, powered by arrays not cheap and unreliable local flash storage.

The FCoT suite is a combination of VXSAN and FCIP. VXSAN provides the multi-tenancy and scale to fibre channel networks, and FCIP gives access to the VXSANs from a variety of FCaaS providers over the inferior IP networks (why IoT devices chose IP instead of FC for their primary connectivity, I’ll never know). Any IoT connected device can do a FLOGI to a FCaaS service and get access to a proper block storage. Currently both Amazon Web Services and Microsoft Azure offer FCoT/FCaaS services, with Google expected to announce support by the end of June 2017.

Why FCoT?

Your refrigerator probably doesn’t need access to block storage, but your car probably does. Why? Devices that are sending back telemetry (autonomous cars are said to produce 4 TB per day) need to put that data somewhere, and if that data is to be useful, that storage needs to be reliable. FCaaS provides this by exposing Fibre Channel primitives.

Tiered storage, battery backed-up RAM cache, MLC SSDs, 15K RPM drives, these are all things that FCoT can provide that you can’t get in a mass-produced chip with inexpensive consumer flash storage.

As the IoT plays out, it’s clear that FCoT will be increasingly necessary.

 

Video: Newbie Guide to Python and Network Automation

Why We Wear Seat Belts On Airplanes

This post is inspired by Matt Simmons‘ fantastic post on why we still have ashtrays on airplanes, despite smoking being banned over a decade ago. This time, I’m going to cover seat belts on airplanes. I’ve often heard people balking at the practice for being somewhat arbitrary and useless, much like balking at turning off electronic devices before takeoff. But while some rules in commercial aviation are a bit arbitrary, there is a very good reason for seat belts.

airplane

In addition to being a very, very frequent flier (I just hit 1 million miles on United), I’m also a licensed fixed wing pilot and skydiving instructor. Part of the training of any new skydiver is what we call the “pilot briefing”. And as part of that briefing we talk about the FAA rules for seat belts: They should be on for taxi, take-off, and landing. That’s true for commercial flights as well.

Some people balk at the idea of seat belts on commercial airliners. After all, if you fly into the side of a mountain, a seat belt isn’t going to help much. But they’re still important.

84271048

Your Seat Belt Is For Me, My Seat Belt Is For You

In a car, the primary purpose of a seat belt is to protect you from being ejected, and to keep you in one place so the car around you (and airbags) can absorb the impact of an impact. Another purpose, one that is often overlooked, is to keep you from smashing the ever loving shit out of someone who did wear their seat belt.

In skydiving, we have a term that encompasses the kinetic and potential energy contained within the leathery sacks of water and bones known as humans: Meat missiles. Unsecured cargo, including meat missiles, can bounce around the inside of airplanes if there’s a rough landing or turbulence. With all the energy and mass, we can do a lot of damage. That’s why flight attendants and pilots punctuate their “fasten you seat belt” speech with “for your safety and the safety of those around you“.

A lot of people don’t realize that if you don’t wear a seat belt, you’re endangering those around you as much as, or more so, than yourself. Your seat belt doesn’t do much good if a meat missile smashes into you. Check out the GIF below:

In the GIF, there’s some sort of impact and as a result the unsecured woman on the left smashes into the secured woman on the right. It’s hard to tell how bad they were hurt, though it could have been a lot worse having two heads smash into each other. The side airbag doesn’t do much good if one solid head hits another solid head. Had the woman on the left had her seat belt on it’s likely their injuries would be far less severe.

While incidents in commercial aviation are far more rare than cars, there can be rough landings and turbulence, both expected and unexpected, and even planes colliding while taxing. Those events can cause enough movement to send meat missiles flying, hence the importance of seat belts.

Commercial aviation is probably the safest method of travel, certainly safer than driving. But there is a good reason why we wear seat belts on airplanes.So buckle up, chumps.

Did VMware vSphere 6.0 Remove the Layer 2 Adjacency Requirement For vMotion? No.

images

I’ve seen this misconception a few times on message boards, reddit, and even comments on this blog: That Layer 2 adjacency is no longer required with vSphere 6.0, as VMware now supports Layer 3 vMotion. The (mis)perception is that you no longer need to stretch a Layer 2 domain between ESXi hosts.

That is incorrect. VMware did remove a Layer 2 adjacency requirement for the vMotion Network, but not for the VMs. Lemme explain.

It used to be (before vSphere 6.0) that you were required to have the VMkernel interfaces that performed vMotion on the same subnet. You weren’t supposed to go through a default gateway (though I think you could, it just wasn’t supported). So not only did your VM networks need to be stretched between hosts, but so did your VMkernel interfaces that performed the vMotion sending/receiving.

What vSphere added was a separate TCP/IP stack for vMotion networks, so you could have a specific default gateway for vMotion, allowing your vMotion VMkernel interfaces to be on different subnets.

This does not remove the requirement that the same Layer 2 network exist on the sending and receiving ESXi host. The IP of the VM needs to be the same, so the VM network you vMotion to needs to have the same default gateway (for outbound packets) and inbound routing (for inbound packets).

Inside of a data center this adjacency is typically done by simply making the same VLAN available (natively or now through VXLAN) on all the ESXi hosts in the cluster.

If it’s between datacenter, things tend to get a more complicated. As in dumpster fire. Here’s a presentation I recently did on the topic, and Ivan Pepelnjak has far more high-brow explanations of why it’s a bad idea.

You’ll need solutions like LISP (for inbound), FHRP filtering (for outbound), OTV (for stretching the VLAN), and a whole host of other solutions to handle all the other problems long distance vMotion can introduce.

Screen Shot 2016-05-24 at 12.20.33 PM.png

Where is your God now?!?!?

So when you hear that vSphere 6 no longer requires Layer 2 adjacency between ESXi hosts, that’s only for the vmkernel interfaces, not the VM networks. So yes, Virginia, you still need Layer 2 adjacency for vMotion. Even in vSphere 6.0.

 

Long Distance vMotion Is A Dumpster Fire

In this screencast, I go on a rant about why long-distance vMotion is a dumpster fire. Seriously, don’t do it.

Fibre Channel in the Cloud: FCaaS

Public cloud providers such as Amazon Web Services, Microsoft Azure, and Rackspace, as well as private cloud systems such as OpenStack, have dominated the computing landscape for the past several years. And once a joke of a marketing term (remember Larry Ellison’s super villain-monologue on the topic?), the cloud is now A Thing, with a definition and everything.

One technology that seemed like it was getting left behind in all these cloud games, however, was Fibre Channel. Ephemeral compute nodes, object storage, extreme scale, elastic provisioning — all of these were characteristics that were initially thought to be bad fits for Fibre Channel.

giphy

Sad Fibre Channel is Sad

As it turns out, Fibre Channel is right at home in the cloud.

mrp6ibd

Amazon Web Services has recently rolled out Fibre Channel as a Service (FCaaS), as have Rackspace, Digital Ocean, and Microsoft Azure.

All of those public cloud providers have some sort of block storage offerings, but they’re typically based on something like iSCSI or another back-end block protocol. Customers have been demanding the kind of block storage in the public cloud, where they can control zoning and zonesets, just like they do in their traditional data centers worlds.

The problem with that historically is that AWS and the others haven’t been able to provide this to customers because of the limitations of Fibre Channel at scale. I’ll explain.

Fibre Channel uses FC_IDs, which are like IP addresses, to send Fibre Channel frames around a given SAN. Here’s an FC_ID: 0x510121.

It’s a 24-bit number, typically written in hexadecimal notation. The first octet (two digits) is known as the domain ID. This is given to the switch, so that means there’s a limit of about 240 or so switches in a given fabric (some domain IDs are reserved). Plus, the two vendors of Fibre Channel switches (Brocade and Cisco) limit domain IDs to a maximum of 50 or so, so no more than 50 or so switches for a given fabric.

For a private data center with a single tenant, this isn’t a problem as a 50 switch Fibre Channel fabric is huge. But for Amazon, 50 switches is miniscule.

So enter VXSAN. The SNIA introduced VXSAN recently under the T18 working group, which provides an extension of typical Fibre Channel frame formats. Like VXLAN, VXSAN adds a higher degree of segmentation.

Cisco has VSANs of course, and Brocade has Virtual Fabrics. Neither are compatible with each other, and neither provide the additional scale required to handle massive cloud scale. VXSAN fixes both of those. VXSAN will work on a traditional Fibre Channel SAN from either Brocade or Cisco, without modification through use of the Open Virtual Fibre Channel Switch.

Wait, what?

That’s right, part of any VXSAN implementation is the Open Virtual Fibre Channel Switch (kind of a mouthful, even with the acronym OVFCS).

Similar to how VXLAN operates an overlay network on a traditional IP network as an underlay, VXSAN operates as an overlay SAN on top of a traditional Fibre Channel SAN.

Instead of VTEPs, OVFC switches terminate the VXSAN segments into virtualization hosts and VXSAN aware storage arrays (both EMC and NetApp have them in their latest software revs) to terminate the VXSAN-applied LUN to the a given virtual machine.

vsan

A given virtualization host has two virtual Fibre Channel switches (A/B), each connected to their own Fibre Channel interface (A/B).

vfc

The virtual Fibre Channel switches rely on upstream NPIV to get their connectivity, so they can run alongside the hypervisor’s traditional SCSI subsystem. In the example below, both virtual Fibre Channel switches do FLOGIs, as does the hypervisor.

FLOGI

The virtual machines, however, to a vFLOGI into the VXSAN segment, not into the traditional switching infrastructure. The upstream physical switches have no idea a FLOGI happened from the VM.

vflogi

The VXSAN header, like VXLAN, has a 24-bit address space, providing 16 million segments, each with their own VXSAN fabric capable of having a full Fibre Channel fabric with up to 239 virtual Fibre Channel switches each. So while 239 Fibre Channel switches won’t work for Amazon, 3.8 billion will (16 million x 239).

You will have to enable Fibre Channel jumbo frames on your traditional Fibre Channel fabric, as the VXSAN header adds 62 bytes to the frame format.

VXSAN is designed to run on VXSAN-unaware switches, as it takes for new header formats to make it into silicon, but both Cisco and Brocade have said they plane to release VXSAN-aware switches by the end of the year.

VXSAN is built to be mulit-tenant, so customers from Amazon and others can do their own zoning. I got to play with a Beta of the FCaaS from AWS and I did just a quick configuration with a single VM and a virtual LUN.

First, you log into the A or B virtual Fibre Channel switches. There’s no password, you use the keys you’ve uploaded into Amazon.

Linux Foundation Open Virtual Fibre Channel Switch (Read the Apache 2.0 License for licensing details)
switch#
switch# config
switch#(config) zone Host1
switch(config-zone)# member pwwn 20:00:00:12:34:45:67:aa
switch(config-zone)# member pwwn 50:00:00:00:00:ab:cd:ef

I was able to push a zoneset and connected my instance to storage pretty quickly. All in all, it only took about 10 minutes to get it up and running.

OpenStack is prepating to include FCaaS and the Open Virtual Fibre Channel Switch in with the next release (Mikata) due out this month.

So check out FCaaS on Amazon, Azure, and the others. FCaaS should bring Fibre Channel into the cloud world.

Edit: Also, this is an April Fool’s joke. 5 years running.

LACP is not Link Aggregation

So there’s a mistake I’ve been making, for years. I’ve referred to what is link aggregation as “LACP”.  As in “I’m setting up an LACP between two switches”. While you can certainly set up LACP between to switches, the more correct term for the technology is link aggregation (as defined by the IEEE), and an instance of that is generically called a LAG (Link Aggregation Group). LACP is an optional part of this technology.

Here I am explaining this and more in an 18 minute Youtube video.