OpenStack - Two Compute Nodes

Getting two Compute  nodes to work was fairly straightforward.

You basically just install openstack-nova-compute, and your Neutron  network plugin (linuxbridge-agent in my case).

The only question I had was whether two Compute Nodes can belong to the same OpenStack Region.

Thank goodness I found a ppt where a guy made it clear that one could run a slew of nodes in a single region (he had multiples in Region 1, and Region 2).

At one point, I decided I would install the OpenVSwitch on this second Compute Node. I'll probably write a second post on that. It did not appear to me that you could mix and match OpenVSwitch and LinuxBridge on different Compute Nodes (at least not easily?). This is because the Neutron L3 Agent config file has a driver field and only seems to accept one mode or the other. I could be wrong about this; more testing necessary. But I backed OpenVSwitch out and enabled LinuxBridge-Agent. Things seem to be working very well with the Linux Bridge Agent.

The Linux Bridge Agent creates Layer 2 Tap interfaces and puts these interfaces on a bridge. If you are using VXLAN protocol it will also manage those interfaces as well.

OpenVSwitch

Today I added a 2nd Compute Node (KVM).

I thought I would use OpenVSwitch on it.

This took me down a deep rabbit hole, as OpenVSwitch is a complex little bugger.

I installed the OpenVSwitch package, then the driver agent (on Compute Node). I wanted it to run in a Layer 2 mode because I had LinuxBridge Agent running on the first Compute Node and the Controller.

After setting OpenVSwitch up on the 2nd Compute node, I realized my external NIC was a bridge, so I tried to use veth pairs to make it work. Nope. As it turns out, the Controller (and L3 agent) seems to use drivers for OpenVSwitch OR LinuxBridge (not both). It appears that it is all or nothing and you cannot mix and match between LinuxBridgeAgent and OpenVSwitchAgent.

I backed it out and used / installed LinuxBridgeAgent.

OpenStack Functional Demo

Originally, with one CentOS 7 server (32 Gb RAM) that was set up to run Ansible and LibvirtD at my disposal, I installed Openstack on a single box.

I put the Controller in a VM and used the host as the Nova Compute Node.

I had all sorts of issues initially. The Keystone and Glance were fairly straightforward. I did not have DNS, so I used IP addresses for most urls, which is a double-edged sword. The complexity in OpenStack is with Nova (virtualization management) and Neutron (networking).

I did not create a "Network Node". I used only a Controller Node and a Compute Node. What one would normally put on a Network Node, runs on the Controller Node (L3 agent, DHCP Agent, Metadta Agent).

One issue was libguestfs was not working. I finally removed it from the box only to realizs that there was a yum dependency with the openstack-nova-conpute package. So I installed nova compute using an rpm with the --nodeps flag.

Getting linuxbridge agent to work took some fiddling. One issue is that it was not clear if I needed fo run LinuxBridgeAgent on the Controller. The instructions make it seem that it is only for the Conpute Node. Well, not so. Neutron creates a tap for every dhcp agent, and every port. ON THE CONTROLLER if that is where you run those services. So you install it both places.

The Neutron configuration file...is about 10,000 lines long, leaving many opportunities for misconfiguration (by omission, incorrect assumption/interpretation, or just plain typos). It took a while to sleuth out how OpenStack uses Nova, Neutron and the l3 agent and linuxbridge agent to create bridges, vnets and taps (ports). But - confusing again - is whether you need to configure all parms exactly same on both boxes, of if some are ignored on one node or the other. I was not impressed with these old style ini and config files. Nightmares of complexity.

Another major challenge I had was the external network. I failed to realize (until I did network debugging) that packets that leave the confines of OpenStack need to find their way back into OpenStack. This means having specific routes to internal OpenStacks networks via the OpenStack external gateway port on the OpenStack router from VMs sitting outside OpenStack.

Another confusing thing is that OpenStack runs namespaces (separate and distinct network stacks) to avoid IP Overlays (by default - the way Neutron  is configured). Knowing how to navigate namespaces is / was a new topic for me and makes it harder to debug connectivity issues.

Finally, when I worked all of this out, I realized that the deployment of VMs was taking up almost 100% CPU. This led me down a rabbit hole to discover that I needed to use the kvm virt_type, and a CPU mode of host-passthrough to calm the box down.

Once I got this done, I could deploy efficiently.

Another thing (maybe this should be its own post) is the notion of setting ports that you can use on deployment (instead of saying "deploy to this network", you can say "use this port on this network" - which has its own IP and port assignment). Because you can attach multiple submets to a single network, I figured I could create ports for nodes that I wanted to reside on that submet. And I COULD! But - the ETSI MANO standards have not caught up with this kind of cardinality / association (per my testing anyway) so it only works if you use OpenStack GUI to deploy. Therefore, having a "one subnet to one network" rule is simpler and will work better for most situations I think.

In the end, I was able to do everything smoothly with, OpenStack. Save Images, create Flavors, Networks, and Deploy. But it all has to be configured "just so".

Service Orchestration and Automation with Open Baton

I haven't blogged in a while, but I've been busy working on Cloud Automation technology of late.

Originally, some guys in the company did an evaluation between Puppet, Chef and Ansible for automating the deployment of virtual machines into the Cloud (they hosted their own cloud and did not rely on the commercial cloud providers we see today).

It took me a while, but I finally had the time to examine their stuff, and before long I was hacking the scripts for my own purposes, so that I could build different versions of our SD-WAN solution, and different topologies of this solution (e.g. we had an L2 solution, an L3 solution, an L3 solution with Routing, et al).  I fell in love with Ansible. I could spin a virtual network up in a matter of minutes, and I could start with raw virtual machines (Linux - CentOS) that would download the packages and install them (with yum installer), install and configure the software, etc. I could probably write a book on the topic of Ansible alone. But - I took someone else's hard work, and ran with the ball and it is always easier to do that then start from scratch yourself.

Then - I was asked to get a prototype of ETSI Mano working.

Years back, when I was at Nokia Networks, we examined Service Orchestration, but back then there were no standards and it was a HUGE integration clusterf$k to get that kind of technology working. We tried it with BEA and JNetX, and message queues. It was a mess.

This time, I read through the standards, and indeed, it looked to me like we HAVE standards drafted up. But - do we have any working solutions? I looked at a solution called OpenBaton, which is open source, out of Berlin, Germany. I put it on a box, went through the tutorials, and it seemed to "kinda sorta" work. So I was able to get this working with a stub "dummy" module that doesn't do anything.

Originally, I put OpenBaton on one virtual machine. It is designed to run on Ubuntu 14.04 (at least that is what the developers tested it on). So, not being heavily familiar with Ubuntu I installed 14.04 on a virtual machine on a KVM Host (32 Gig RAM, 8 core CPU and 1Tb Disk) and immediately the Ubuntu upgraded it to 16.04. This created some problems right away.  One HUGE issue is that all of the software is written in Java, and they stated that they wanted JDK 1.7. But - guess what? Oracle had just deprecated 1.7 that very week, and took the 1.7 JDK link down, which broke all of Open Baton's scripts. Don't ask me how I got around this...it was very difficult. I installed OpenJDK 1.7, and then "faked things" so that Open Baton's scripts would believe Oracle JDK was on the box.  I wound up having to download and compile many packages from GitHub and compiling them myself. I also wound up having to hack and manipulate the SystemD unit files so that the services would start up properly.

Initially, I only installed the Orchestrator (NFVO), and the Generic VNFM (Virtual Networking Function Manager) modules. But, to really vet the technology out, OpenBaton needs a "real" system to talk to. So, in a 2nd Virtual Machine on the KVM host, I installed an OpenStack Controller Virtual Machine on Centos 7, and it ran along side the OpenBaton Virtual Machine. On the KVM Host, I installed the Neutron Compute Module, which is responsible for interacting with the KVM host and launching the virtual machines.

I got it to launch machines, but that gets boring quick. I wanted to examine the ability to run scripts and configure the VMs dynamically, and have the VMs inter-communicate. I then learned that OpenBaton - though called an Orchestrator - cannot actually pull any of this off without using an EMS (Element Management System). And Open Baton uses Zabbix for this. So, I had to install a Zabbix Server and a Zabbix Plugin - and I installed these on the Open Baton Virtual Machine, thinking that I would alleviate issues if I put them all together on the same box (more on this later).

In the end, I am able to get Open Baton to launch VMs consistently, but I get a TON of timeout errors. As I debug things, I realize that threads and message queues are timing out because the process of deploying and configuring the VMs is so CPU and Disk intensive that the VMs just get overwhelmed and Open Baton gets impatient waiting for things to happen.

I run top (as well as htop and other tools) and realize that I need to take a step back if I am going to take a step forward. I need to get another box - a second box - and distribute some load, and move some things out of these virtual machines.

Okay, that's it for now. I will update more on the next post.

NetFlow with Ntop

I had heard that Ntop supports Netflow on Linux.

I found a link / blog where someone else has played with this package for same or similar purposes. Let me share that here:
https://devops.profitbricks.com/tutorials/install-ntopng-network-traffic-monitoring-tool-on-centos-7/

I downloaded the Ntop package, and immediately it barked about the fact that I did not have kernel headers on the system.

This is bad, in my mind.

What box, running out in the field, would have kernel headers installed on it? That would be a bad security practice because it would mean that the box has a lot of stuff on it that it probably shouldn't have...specifically this would mean compilers, et al?

I also noticed that the package runs with a license code. There is a limited license it can run as, which is default configured.  But I'm not sure I like having software, at least for this purpose, that is dependent on licensing. I did not study whether it is a key license that is time expired, or if it calls out to a remote server to authenticate the license, et al.

I kind of stopped there. I did not play with it any further. I may come back to it, and if I do I will update this accordingly.

NetFlow with nfcapd and fprobe

I spent some time researching and using NetFlow this week (about a day).

Basically, you download the nfdump package, which has the collector (nfcapd), and a GUI (nfsen) and a command line tool called nfdump.

You run the collector, which listens on a standard or specified port, and "something" (i.e. a router) that knows how to capture flows, will write netflow formatted files. Then you can use nfdump or nfsen to view these flows.

There are multiple versions of NetFlow - from version 5 all the way up to 9 (see the NetFlow Wiki). The different versions provide additional data (or extensions as they refer to them).

The tricky part in testing this is to mimic or simulate a router. To do this:

fprobe is a tool you can install to generate flows. But it does not appear to install with the yum package manager, so you need to download the source and compile it, or there is an rpm that can be downloaded and installed.

frpobe-ulog is another tool, but it runs over iptables and requires iptables rules to work. I was surprised to see that yum COULD find and install this program, but not fprobe.

There are a few other tools as well, but these were the two I tried out.

Both of these worked, although there is not a lot of documentation or forum discussion on the fprobe-ulog approach. I wound up using fprobe.

There is the question of what defines and constitutes a network flow. The Wikipedia defines this. I think that if you have a bunch of udp traffic, it is harder for Netflow to stitch the traffic together into a flow for hindsight analysis. But TCP of course is straightforward.

System Tap

I spent some time reading the Beginner's Guide to System Tap.

https://sourceware.org/systemtap/SystemTap_Beginners_Guide/

I learned the basics of writing, reading and compiling / running the system tap scripts.

I also enjoyed running the sample System Tap scripts that are mentioned specifically in Chapter 5 of this guide.

I wound up downloading a bunch of these; especially the networking ones.

Writing these efficiently would take some practice. But there is a good Reference Guide that can make the process of writing the scripts easier. The question is, could there be a use case for writing one of these scripts that someone hasn't already thought up, and written?