Server virtualisation is a growing reality in data centres. The economics are firmly behind the trend. Server virtualisation reduces the total cost of ownership by reducing the number of physical servers, requiring less cooling and less power while increasing flexibility. This is all good for the business and the server group, but what effect does it have on the management of the network? The truth is that it complicates network management.
There are two big network problems associated with server virtualisation. The first is configuring virtual LANs. Network managers need to make sure the VLAN used by the virtual machine (VM) is assigned to the same switch port as the physical server running the VM.
One solution is for the server virtualisation group to tell network management team every possible server the VM can be started on and preconfigure the switch ports. This is not a perfect solution because it can cause the VLAN to be defined on a very large percentage of the switch ports. It can get even more complicated because the server group may not be aware of all the servers that images can be started on, especially during a recovery situation when they are taking emergency measures.
The second problem is assigning QoS and enforcing network policies, such as access control lists (ACLs). Traditionally this is done in the network switch connected to the server running the application. With server virtualisation there's a software switch running under the hypervisor in the physical server - not the traditional physical network switch that connects to the physical server.
It is still important that policy be enforced in the the software switch. For example, if two VMs running on the server are not allowed to communicate with each other, someone who gained control of VM1 could open connections to VM2 and steal its data. If ACLs are applied by the soft switch in the server then this would be blocked.
Before virtualisation, this was prevented because the applications in VM1 and VM2 would run in different servers and the ACLs defined in the network switch would prevent the communication. Having policies applied in the software switch maintains the security. The issue is how to get the software to apply the policies.
Overcoming these two challenges is critical to making server virtualisation work smoothly. It would have been nice if the vendor community had created a uniform standard that works with all the different virtualisation vendors. As is normally the case with rapidly growing new technology, this did not happen. The industry has implemented four ways to address these problems.
Virtualisation vendors' solution
The market-leading virtualisation vendor is EMC VMware but many other virtualisation products exist, including Citrix's Zen, Microsoft's Hyper-V, KVM and offerings from many other smaller vendors. The most widely available products are for VMware, which is why I use it here as the example.
VMware's vCenter controls the virtualisation process and directs where the VMs are started. The hypervisor controls the server and the VMs running on the physical server. VSwitch is a software Layer 2 switch provided by VMware. Each VM has a virtual NIC. The vNIC uses a MAC address from either the virtualisation vendor's pool of MAC addresses or one created and assigned by the enterprise.
Step 1 is for the server group to define all network characteristics and policies for the VM machines. The operator tells vCenter to start VM2 in Step 2. This process includes multiple messages between vCenter and the hypervisor on the server, one of which pushes the network policy information to the hypervisor. In Step 3, the hypervisor configures the vSwitch with the correct VLAN, QoS and policy information. When the application on VM2 starts to send packets, the policy is applied in the vSwitch.
This solves the problem of applying policies at the first switch but it does not solve the VLAN configuration problem in the network switch. The virtualisation groups need to tell network management to configure the VLAN on the switch port before the VM starts sending traffic, which requires quick coordination or the switch has to be preconfigured. The coordination can get more complicated when the virtualisation group moves the VM on the fly. Then the virtualisation group needs to coordinate with the networking group as it moves the server, and the network group needs to clean up the configuration on the old switch after a successful move.
One of the biggest concerns with this approach is the amount of coordination required between the virtualisation and network groups. The virtualisation group must configure parameters in vCenter that are controlled by the networking group, such as VLAN numbers, QoS, and ACLs. This means that good ongoing coordination is needed between the server virtualisation group and the networking group. Any change in VLANs or policies must be immediately reflected in the virtual server configuration, which introduces another possible failure point.
Another concern is the lack of visibility to what is going on within a networking component, the vSwitch, by the networking group. The vSwitch is under vCenter's control, not traditional network management software. Additionally, the network management team has little visibility into the VM. This visibility problem has been addressed by several networking vendors by having vCenter notify the networking team of changes or polling for changes and then displaying this information along with the traditional network data, which greatly helps with problem determination.
Some ways to address the problem
Blade Networks has an application that runs on its switch and Force10's next release of its OS addresses the VLAN problem. The switches poll vCenter looking for any changes, or alternatively listens for vCenter to send out a message announcing a change. If the switch finds any changes, it will automatically perform the configuration. The virtualisation operator doesn't have to coordinate the change with network operation, allowing start up of the VM to go smoothly. The polling interval does need to be smaller than the time it takes to start a VM to make sure the switch sees the change fast enough. In Force10's first release, the only parameter it monitors for is VLAN. Blade Network goes further by also applying the full range of policies at the network switch based on the vNIC or VM's UUID. This solution still requires that policies be implemented in the vSwitch.
The second way to solve the configuration problem is with orchestration software from vendors such as Hewlett-Packard and Juniper, which work on their own switches. There are also solutions from management vendors such as Scalent and CA. In this case, the orchestration software talks to both the network switch and vCenter and coordinates changes in configuration between the two environments. This approach has the potential benefit of being able to work with a wide range of switch and virtualisation vendors.
Cisco provides its own soft switch solution to replace vSwitch called the 1000V. There are two components to the 1000V. The VSM is the Virtual Switch Module and replaces the vSwitch software running inside the hypervisor. The Virtual Element Manager (VEM) is where the network policies for the VSM are configured and stored.
In Step 1, the VM's VLAN and policies configured based on the VM's UUID or vMAC addresses in the VEM. VCenter starts a new VM or moves a VM in Step 2. The Hypervisor informs the VSM in Step 3. The VSM then retrieves the policy information from the VEM in Step 4. If the network switch is from the Nexus product line, it also retrieves the necessary VLAN and policy information from the VEM. At this point both the switch in the hypervisor and the Nexus switch have all the correct information on how to handle VM2. When VM2 starts sending traffic, all the correct policies are applied in the 1000V switch in the hypervisor.
The benefits of Cisco's approach are the same as with the first approach. It blocks any communication between the two VMs if it is not allowed and applies the appropriate policies the first time traffic hits a switch. If the 1000V is used with the Nexus switch that is virtualisation-ready, it will solve the VLAN problem in the network switch. It has the additional benefit of moving the switch in the hypervisor under the control of the network management software, returning the clear accountability to the network group. There are downsides. Currently Cisco has a solution for only VMware, not for Xen and HyperV.
The fourth approach takes a network equipment centric view; see Figure 3. In Step 1, the VM is defined in the network management software by its virtual NIC. In Step 2, vCenter directs the hypervisor to start up VM. The hypervisor sends out an advertisement packet announcing it is starting VM2, in Step 3. The advertisement has VM2's vNIC and its UUID. In Step 4, the switch sees the advertisement and sends a request for its VLAN and other policy information. The switch then applies the policies to any traffic entering the network.
The key point is the switch only applies policy in the network switch, and not at the vSwitch. The switch also monitors for messages from the hypervisor that indicate the VM has moved and then removes the VLAN and policy information associated with the vNIC. Vendors using this solution include Arista Networks, Blade, and Enterasys, along with HP and Juniper through their orchestration approach. Other vendors such as Brocade plan to offer this solution. Extreme Networks uses this technique for QoS and policy but not for VLANs.
This approach tries to remove the need to involve the virtualisation group in enforcing network policies. However, there are still two problems. The first is that the server virtualisation and networking groups must still coordinate VLAN numbers. Currently, Enterasys has the ability to automatically provide the vSwitch with the VLAN number, with Arista planning to add it shortly.
The biggest problem with this approach is that it does not apply policies at the vSwitch, thus allowing traffic between VMs on the server to bypass ACLs and other security policies. Enterasys and Arista plan to overcome this problem by adding the ability to push the policies down to the vSwitch.
A future way to overcome this problem is by allowing "hairpin turns" in the switch. With hairpin turns, the vSwitch is configured to send all traffic, even VM1 to VM2 traffic, directly to the network switch. The network switch then applies the policies and assigns QoS. The VM1-to-VM2 traffic would be sent back to the vSwitch which then delivers it to VM2.
This would make vSwitch a "dumb" switch with only a forwarding role. The problem is that 802.1d, the bridging standard that all Layer 2 switches are based on, does not allow traffic that came from a port to be sent back down the same port the traffic came from. Thus under the current rules the network switch could not return the packet from VM1 addressed to VM2 because doing so would break this rule. This was added to prevent loops from being formed. The IEEE is working a revision to 802.1d that allows the switch to perform hairpin turns and additional work is underway to standardise the dump switch in the hypervisor. When this becomes widespread it addresses the problem and has the benefit of removing most of the coordination between the network and server group.
Enterasys has a workaround that directs the vSwitch to put each VM in a separate VLAN. They select VLAN numbers that are not being used to prevent any potential problems. Because the VMs are in different VLANs, they cannot communicate with each other. When the packet arrives at the network switch, the switch replaces the VLAN numbers with the real one assigned to the VM making it appear to the network and their destination that they have always been in the correct VLAN.
The VLAN connection issue
There is another VLAN configuration problem that the techniques described thus far do not address. When a new VLAN number is assigned to a port, it needs to be connected to all the other ports with that VLAN number. This requires that all the aggregation switches in the path need to have the VLAN defined on them.
For example, let's say VLAN 5 supports an accounts payable application. All the VMs that support the application are located on one top-of-rack switch that has been configured for VLAN 5. For workload reasons, one of the VMs running the accounts payable application is moved to another rack with its own switches in another part of the data center that is reached by way of several aggregation switches.
Preserving the VLAN means that all the intermediate switches need to be configured with VLAN5; if they are not then the VLAN is broken. Currently none of the products mentioned here deal with how to automatically assign the VLAN in the aggregation switch. This means that if a VM can move across a data center, the VLAN number must be preconfigured in all the aggregation and core switches. Additionally, it is not likely that this problem will be solved in the immediate future.
The industry has worked out a set of adequate solutions for the port VLAN and policy problems. There is no magic, most require the networking and virtualisation groups to coordinate their activities in the short term to make it work smoothly. The hairpin turn is the best long-term solution and the industry is moving toward it. It is important that network managers understand how the various solutions work with the different virtualisation vendors they support, as the full range of solutions are not available for all the virtualisation solutions currently in the market.
Layland is head of Layland Consulting. He can be reached at [email protected]