• /
    • Archives /
  • 30 November 1999

Archives For 30 November 1999

How I replaced the CloudStack Virtual Router with my own physical Linux router

30 August 2012 — 13 Comments

Back in March I wrote a blog on how to create a network without a Virtual Router.  I received a lot of questions about it. It’s also a question that pops up now and then on the CloudStack forums. In the meanwhile I’ve worked hard to implement this setup at work. In this blog I’ll describe the concept of working with a CloudStack setup that has no Virtual Router.

First some background. In Advanced Networking, VLAN’s are used for isolation. This way, multiple separated networks can exist over the same wire. More about VLAN technology in general on this wikipedia page. For VLAN’s to work, you need to configure your switch so it knows about the VLAN you use. VLAN’s have an unique id between 1 and 4096. CloudStack configures this all automatically, except for the switch. Communication between Virtual Machines in the same CloudStack network (aka VLAN) is done using the corresponding VLAN-id. This all works out-of-the-box.

It took me some time to realize how powerful this actually is. One can now combine both VM’s and physical servers in the same network, by using the same VLAN for both. Think about it for a moment. You’re now able to replace the Virtual Router with a Linux router simply by having it join the same VLAN(s) and using the Linux routing tools.

Time for an example. Say we have a CloudStack network using VLAN-id 1234, and this network is created without a Virtual Router (see instructions here). Make sure you have at least 2 VM’s deployed and make sure they’re able to talk to each other over this network. Don’t forget to configure your switch. If both VM’s are on the same compute node, networking between the VM’s works, but you won’t be able to reach the Linux router later on if the switch doesn’t know the VLAN-id.

Have a separate physical server available running Linux and connect it to the same physical network as your compute nodes are connected to. Make sure the ip’s used here are private addresses. In this example I use:

compute1: 10.0.0.1
compute2: 10.0.0.2
router1: 10.0.0.10
vm1: 10.1.1.1
vm2: 10.1.1.2

The Linux router needs two network interfaces: one to the public internet (eth0 for example) and one to the internal network, where it connects to the compute nodes (say eth1). The eth1 interface on the router has ip-address 10.0.0.10 and it should be able to ping the compute node(s). When this works, add a VLAN interface on the router called eth1.1234 (where 1234 is the VLAN-id CloudStack uses). Like this:

ifconfig eth1.1234 10.1.1.10/24 up

Make sure you use the correct ip-address range and netmask. They should match the ones CloudStack uses for the network. Also, note the ‘.’ between the eth1 and the VLAN-id. Don’t confuse this with ‘:’ which just adds an alias ip.

To check if the VLAN was added, run:

cat /proc/net/vlan/eth1.1234

It should return something like this:

eth1.1234 VID: 1234 REORDER_HDR: 1 dev->priv_flags: 1
 total frames received 14517733268
 total bytes received 8891809451162
 Broadcast/Multicast Rcvd 264737
 total frames transmitted 6922695522
 total bytes transmitted 1927515823138
 total headroom inc 0
 total encap on xmit 0
Device: eth1
INGRESS priority mappings: 0:0 1:0 2:0 3:0 4:0 5:0 6:0 7:0
 EGRESS priority mappings:

Tip: if this command does not work, make sure the VLAN software is installed. In Debian you’d simply run:

apt-get install vlan

Another check:

ifconfig eth1.1234

It should return something like this:

eth1.1234 Link encap:Ethernet HWaddr 00:15:16:66:36:ee 
 inet addr:10.1.1.10 Bcast:0.0.0.0 Mask:255.255.255.0
 inet6 addr: fe80::215:17ff:fe69:b63e/64 Scope:Link
 UP BROADCAST RUNNING MASTER MULTICAST MTU:1500 Metric:1
 RX packets:14518848183 errors:0 dropped:0 overruns:0 frame:0
 TX packets:6925460628 errors:0 dropped:15 overruns:0 carrier:0
 collisions:0 txqueuelen:0 
 RX bytes:8892566186128 (8.0 TiB) TX bytes:1927937684747 (1.7 TiB)

Now, the most interesting tests: ping vm1 and vm2 from the linux router, and vice versa. It should work, because they are all using the same VLAN-id. Isn’t this cool? You just connected a physical server to a virtual one! 🙂

You now have two options to go from here:

1. Use a LoadBalancer (like Keepalived) and keep the ip’s on the VLAN private using Keepalived’s NAT routing. The configuration is exactly the same as if you had all physical servers or all virtual servers.

2. Directly route public ip’s to the VM’s. This is the most interesting one to explain a bit further. In the example above we’ve used private ip’s for the VLAN. Imagine you’d use public ip addresses instead. For example:

vm1: 8.254.123.1
vm2: 8.254.123.2
router1: 8.254.123.10 (eth1.1234; eth1 itself remains private)

This also works: vm1, vm2 and router1 are now able to ping each other. A few more things need to be done on the Linux router to allow it to route the traffic:

echo 1 > /proc/sys/net/ipv4/ip_forward
echo 1 > /proc/sys/net/ipv4/conf/eth1/proxy_arp

Finally, on vm1 and vm2, set the default gateway to router1; 8.254.123.10 in this example.

How does this work? The Linux router also answers arp requests for the ip’s in the VLAN. Whenever traffic comes by for vm1, router1 answers the arp request and routes the traffic over the VLAN to vm1. When you’d run a traceroute, you’ll see the Linux router appear as well. Of course you need to have a subnet of routable public ip’s assigned by your provider for this to work properly.

To me this setup has two major advantages:

1. No wasted resources for Virtual Routers (one for each network)
2. Public ip’s can be assigned directly to VM’s; you can even assign multiple if you like.

The drawbacks? Well, this is not officially supported nor documented. And since you are not using the Virtual Router, you’ll have to implement a lot of services on your own that were normally provided by the Virtual Router. Also, deploying VM’s in a network like this only works using the API. To me these are all challenges that make my job more interesting 😉

I’ve implemented this in production at work and we successfully run over 25 networks like this with about 100-125 VM’s. It was a lot of work to configure it all properly and to come up with a working solution. Now that it is live, I’m really happy with it!

I realize this is not a complete step-by-step howto. But I do hope this blog will serve as inspiration for others to come up with great solutions build on top of the awesome CloudStack software. Please let me know in the comments what you’ve come up with! Also, feel free to ask questions: I’ll do my best to give you some directions.

Enjoy!

In Cloud Orchestration , , , , , , ,
permalink

High cpu load on CloudStack management servers after leap second 30/6/2012@23:59:60 UTC

1 July 2012 — 5 Comments

After the leap second insert last night, my CloudStack 3.0 servers (or its Java processes actually) started to use a lot of CPU. Here’s how to fix it (sets the date):

date ; date -s "`date -u`" ; date

Just run this on your management- or compute node. On my CloudStack system it only occurred on the management servers. Confirmed: it occurred both on management- and compute nodes on our CloudStack system. No restart required afterwards . The load will drop immediately.

Note: restarting cloud-management alone does not fix the issue. Rebooting the machine does, however, but I’d prefer not to reboot them 🙂

MySQL seems to be affected as well, though I didn’t experience problems with it.  Thanks to the guys @Mozilla for blogging about this problem and suggesting a fix.

I’ve also posted this to the CloudStack forums,so there might be some discussion as well.

In Cloud Orchestration, Linux sysadmin , , , ,
permalink

Meet our shiny new CloudStack Cloud

28 June 2012 — Leave a comment

This is what our Cloud looks like in the CloudStack Dashboard. Pretty powerful 🙂


These are the boxes the hardware was in.


This image is taken in our lab while testing CloudStack.


Very handy: a tray for spares in your rack!

 

This is the front: 6 compute nodes, 2 management servers and 2 Linux routers that manage all traffic. We also have 2 big storage servers that you cannot see on this image.

This is how the back of the rack in the data center looks like after we’ve built everything in.


Close up of power management and storage network.


And the final image shows the serial, public, and manage networks.

We’ve labelled and documented every cable and used a separate color for each connection type (i.e. mgt network, storage network, uplinks, cross links, serial connections, etc).

In Cloud Orchestration , , ,
permalink

High Availability with automatic failover for CloudStack management servers

2 May 2012 — 12 Comments

I’m using CloudStack for some months now and our cloud is close to going live. It’s an awesome piece of software that is just rock solid :-). One thing I couldn’t really find is how to create high available management servers with automatic failover. I’ve come up with a solution that I’ll share in this blog post.

From the CloudStack manual:

The CloudStack Management Server should be deployed in a multi-node configuration such that it is not susceptible to individual server failures. (…) MySQL may be configured to use replication to provide for a manual failover in the event of database loss.

Of course, when building a cloud one cannot just have one management server, as that would create a big single-point-of-failure . Even though there is no impact on already running VM’s, you and your customers, for example, won’t be able to stop and start VM’s. The manual suggests looking into “MySQL replication” and when problems occur, “manually failover” to another server.

How does that work? The management server itself is stateless, which means you can have two management servers and if you’d issue a command to either of them, the result would be the same. You can distribute the load, it just doesn’t matter which management server you’ll talk to. So there’s no master nor slave: they’re just all the same. The challenge is where CloudStack stores its data: in a MySQL server. We should have one MySQL master server that handles the requests from the management servers. MySQL supports replication, which means you can add MySQL slave servers that would just stay in sync with the master using the binary logs. You cannot query them directly, that’s what the master is for. When the master dies, you can promote a slave to be the new master. But this is a manual step to take.

Personally, I’d like to automate this. Further more, my experience with MySQL master/slave in the past, is that it isn’t rock solid. Sometimes slaves would get out of sync due to some error. You at least need some monitoring to warn you when this happens. It is almost always possible to fix this, but again this is manual work and I was looking for an automatic solution. So I came up with an alternative..

Since 2005 I’m building Linux clusters at work for our webhosting and e-mail business. Using Open Source techniques, that is. One of the things I’ve been using for years is DRBD. You can think of DRBD as a network based RAID-1 mirror. Using a dedicated high speed network (Gigabit or better), DRBD keeps two disks in sync. Using another Open Source tool, Heartbeat, one can automatically fail-over from one server to another and keep the service online. Heartbeat and DRBD have a sub-second failover and in case of a complete (power) failure of the hardware, the automatic failover takes just 10 seconds. Now that’s cool!

How can this help solve the management server issue? Imagine two management servers, that use DRBD to keep a given disk in sync. This disc is mounted on /var/lib/mysql on the management server that is primary. It is not mounted on the secondary management server. Heartbeat makes sure MySQL is only run on the primary server. To make it all easy to manage, Heartbeat also takes care of an extra ip-addres that is always mounted on the primary server. We call this the “virtual ip-address”. It looks like this:

Wat do we have then? Two management servers, both run the CloudStack management software and can be used to visit the webserver, call the API etc. Both management servers use the MySQL server which is run on the primary server. Tell CloudStack to use the “virtual ip-address” as MySQL host address. DRBD will make sure the stand-in server has an up-to-date version of the MySQL disk.

If the secondary server dies, nothing happens apart from losing redundancy. What if the primary server fails?

When either server goes offline, the MySQL disk and MySQL service is run on the server that is still alive. Of course also the CloudStack management is still available then. This way, you have an automatic failover for the CloudStack management server.

To extend this setup, one could easily setup a loadbalancer that distributes the traffic between the management servers. Both keepalived and haproxy can do that for you.

I hope this brings some inspiration to others working with CloudStack. If you’ve suggestions, questions or improvements let me know!

In Cloud Orchestration , , , , , ,
permalink

The power of coding your infrastructure – CloudStack and Puppet in action!

25 March 2012 — Leave a comment

Lately I’ve been coding to automate the deployment of Virtual Machines that will run in our cloud. It’s inspired by blogs and code from Jason Hancock but implemented in a different way to meet our specific needs. Since CloudStack has a very flexible API, you can make it do anything you want. It works pretty cool so far.

From a command line script we can deploy a new VM with a certain “role”. This role has a name, like “web server” or “test-vm-db” and contains the OS template name, the datacenter name, the networks to connect to, the hardware offering (CPU/Memory/Disk) and the account and domain to use. These are all CloudStack specific settings. When calling the deployVirtualMachine() API command, CloudStack will create a VM with exactly these specs. We use a simple php cli script to add a new server, like this:

Result:

Since it is in a specific domain (defined by the role), other users in that domain can “see” and work with the new VM immediately. When you deploy for a client, they will “see” the new VM like this:

When the VM boots, it is receives its role name, “web server” for example, and sends this info to Puppet. This way, Puppet knows how to configure the new VM. It will do so and when it’s finished you have an exact same “web server” as the others. Run the script again and you’ll have another one. Puppet does not only do initial configuration, it will make sure the configuration stays consistent across the all servers it manages. When you want to change a setting, Puppet distributes it and makes sure it’s the same on all servers. All automatically!

Imagine the power behind this concept. All of a sudden you can create a script that can deploy a new web server in just minutes. Now even the loadbalancer itself can spin up a new web server to handle the higher load it receives. Or destroy a web server when there are too many idle web servers. That’s what you call elastic!

What you need to build this yourself is:
1. PHP CloudStack Client (to talk to CloudStack from PHP)
2. CloudStack API docs
3. Script to grab meta-data when VM has booted

Please see Jason Hancock‘s blog for some working examples of how he implemented it. As I said, CloudStack’s API is very flexible. But you can even extend the API, if you want to go further. Thanks for the inspiration Jason!

In Cloud Orchestration , ,
permalink