Networks, teaming and heartbeats for clusters

Introduction

In this guide, a fabric is a separate network infrastructure, be it SAN, WAN or LAN. A network may or may not be connected to a dedicated fabric. Some fabrics have more than one network.

The cluster nodes should be connected to each other over at least two independent networks/fabrics. The more independent the better. Ideally, the networks should share no components at all, but as a minimum they should be connected to separate NICs in the server. Ergo, if you want to use NIC teaming you should have at least 4 physical network ports on at least two separate NICs. The more the merrier, but be aware that as with all other forms of redundancy, higher redundancy equals higher complexity.

If you do not have more than one network port or only one network team, do not add an additional virtual network adapter/vlan for “heartbeat purposes”. The most prevalent network faults today are caused by someone unplugging the wrong cable, deactivating the wrong switch port or other user errors. Having separate vlans over the same physical infrastructure rarely offers any protection from this. You are better off just using the one adapter/team.

Previously, each Windows cluster needed a separate heartbeat network used to detect node failures. From Windows 2008 and newer (and maybe also on 2003) the “heartbeat” traffic is sent over all available networks between the cluster nodes unless we manually block it on specific cluster networks. Thus, we no longer need a separate dedicated heartbeat network, but adding a second network ensures that the cluster will survive failures on the primary network. Some cluster roles such as Hyper-V require multiple networks, so check what the requirements are for your specific implementation.

Quick takeaway

If you are designing a cluster and need a quick no-nonsense guideline regarding networks, here it comes:

  • If you use shared storage, you need at least 3 separate fabrics
  • If you use local storage, you need at least 2 separate fabrics

All but a few clusters I have been troubleshooting have had serious shortcomings and design failures in the networking department. The top problems:

  • Way to few fabrics.
  • Mixing storage and network traffic on the same fabric
  • Mixing internal and external traffic on the same fabric
  • Outdated faulty NIC firmware and drivers
  • Bad, poorly designed NICs from Qlogic and Emulex
  • Converged networking

Do not set yourself up for failure.

IPv6

If you haven’t implemented IPv6 yet in your datacenter, you should disable IPv6 on all cluster nodes. If you don’t, you run a high risk of unnecessary failovers due to IPv6 to IPv4 conversion mishaps on the failover cluster virtual adapter. As long as IPv6 is active on the server, the failover cluster virtual adapter will use IPv6, even if none of the cluster networks have a valid IPv6 address. This causes all heartbeat traffic to be converted to/from IPv4 on the fly, which sometimes will fail. If you want to use IPv6, make sure all cluster nodes and domain controllers have a valid IPv6 address that is not link local (fe80:), and make sure you have routers, switches and firewalls that support IPv6 and are configured properly. You will also need IPv6 dns in the active directory domain.

Disabling IPv6

Do NOT disable IPv6 on the network adapters. The protocol binding for IPv6 should be enabled:

clip_image001

Instead, use the DisabledComponents registry setting. See Disable IPv6 for details.

clip_image003

Storage networks

If you use IP-based storage like ISCSI, SMB or FCOE, make sure you do not mix it with other traffic. Dedicated physical adapters should always be used for storage traffic. Moreover, if you are one of the unlucky few using FCOE you should seriously consider converting to FC or SMB3.

Hyper-V networks

In a perfect world, you should have six or more separate networks/fabrics for Hyper-v clusters. Sadly though, the world is seldom perfect. The absolute minimum for production clusters is two networks. Using only one network in production will cause nothing but trouble, so please do not try. Determining whether or not to use teaming also complicates matters further. As a general guide, I would strongly recommend that you always have a dedicated storage fabric with HA, that is teaming or MPIO, unless you use local storage on the cluster nodes. The storage connection is the most important one in any form of cluster. If the storage connection fails, everything else falls apart in seconds. For the other networks, throughput is more important than high availability. If you have to make a choice between HA and separate fabrics, chose separate fabrics for all other networks than the storage network.

7 Physical networks/fabrics

· Internal/Cluster/CSV (if local)/Heartbeat

· Public network for VMs

· VM Host management

· Live Migration

· 2*Storage (ISCSI, FC, SMB3)

· Backup

5 Physical networks/fabrics

· Internal/Cluster/CSV (if local)/Heartbeat/Live Migration

· Public network for vm, VM guest management

· VM Host management

· 2*Storage (ISCSI, FC, SMB3)

4 Physical networks/fabrics

· Internal/Live Migration

· Public & Management

· 2*Storage

Example

clip_image004

Most blade server chasses today have a total of six fabric backplanes, grouped in three groups where each group connects to a separate adapter in the blade. Thus, each network adapter or FC HBA is connected to two separate fabrics/backplanes. The groups could be named A,B and C, with the fabrics named A1, A2, B1 and so on. Each group should have identical backplanes, that is the backplane in A1 should be the same model as the backplane in A2.

If we have Fibrechannel (FC) backplanes in group A, and 10G Ethernet backplanes in group B & C, we have several possible implementations. Group A will always be storage in this example, as FC is a dedicated storage network.

clip_image005

Here, we have teaming implemented on both B and C. Thus, we use the 4 networks configuration from above, splitting our traffic in Internal and Public/Management. This implementation may generate some conflicts during Live Migrations, but in return we get High Availability for all groups.

clip_image006

By splitting group B and C in two single ports, we get 5 fabrics and a more granulated separation of traffic at the cost of High Availability.

Hyper-V trunk adapters/teams on 2012

If you are using Hyper-V virtual switches bound to a physical port or team on you Hyper-V hosts, Hyper-V Extensible Virtual Switch should be the only bound protocol. Note: Do not change these settings manually, Hyper-V manager will change the settings automatically when you configure the virtual switch. If you bind the Hyper-V Extensible virtual switch protocol manually, creation of the virtual switch may fail.

clip_image007

Teaming in Windows 2012

In Windows 2012 we finally got native support for nic teaming. You access the nic teaming dialog from Server Manager. You can find a short description of the features here: http://technet.microsoft.com/en-us/library/hh831648.aspx, and a more detailed one here: Windows Server 2012 NIC Teaming (LBFO) Deployment and Management.

Native teaming support rids us of some of the problems related to unstable vendor teaming drivers, and makes setup of nic teaming a unified experience no matter what nics you are using. Note: never use nic teaming on ISCSI networks. Use MPIO instead.

A note on Active/Active teaming

It is possible to use active active teaming, thus aggregating the bandwidth of two or more adapters to support higher throughput. This is a fantastic technology, especially on 1G ethernet adapters where bandwidth congestion can become a problem. There is, however a snag; a lot of professional datacenters have a complete ban on active/active teaming due to years of teaming problems. I have my self been victim of unstable active/active teams, so I know this to be a real issue. I do think this is less of a problem in Windows 2012 than it was on previous versions, but there may still be configurations that just does not work. The more complex your network infrastructure is, the less likely active/active teaming is to work. Connecting all members in the team to the same switch increases the chance of success. This also makes the team dependent on a single switch of course, but if the alternative is bandwidth congestion or no teaming at all, it does not really matter.

I recommend talking to your local network specialist about teaming before creating a design dependent on active/active teaming.

Using multiple vlans per adapter or team

It has become common practice to use more than one vlan per team, or even more than one vlan per adapter. I do not recommend this for clusters, with the exception of adapters/teams connected to a Hyper-V switch. An especially stupid thing to do is mixing ISCSI traffic with other traffic on the same physical adapter. I have dealt with the aftermath of such a setup, and it does not look pretty unless data corruption is your kind of fun. And if you create a second vlan just to get an internal network for cluster heartbeat traffic on the same physical adapters you are using for client connections, you are not really achieving anything other than making your cluster more complex. The cluster validation report will even warn you about this, as it will detect more than one interface with the same MAC address.

Author: DizzyBadger

SQL DBA Principal Analyst

Leave a Reply