This is part 1 in series on Windows Network Load Balancing. Part 2 can be found here: https://workinghardinit.wordpress.com/2010/07/23/reflections-on-getting-windows-network-load-balancing-to-work-part-2/
This will not be an extensive NLB installation & configuration manual. You’ll find plenty of material on that searching the internet. I would like to reflect on some issues and options when using Windows Network Load Balancing.
I will not be discussing NLB solutions using just one NIC with multicast. I think they lack so badly in resilience, configuration and troubleshooting capabilities that I never consider using them, not even in the lab. Even in a lab you need to work like in real live, bar some exceptions. Apart from no available slots in a server to add NICs you have no excuse not to and even then, just make sure you do. NIC ports are very cheap nowadays and especially in a virtual environment there is nothing stopping you from adding some extra virtual ports. Do yourself a favor and always use two or more NIC ports. Even in the year 2000 I grinned when I read that one of the drawbacks was the cost of the extra NIC. Really, you have a real business need and are prepared to pay for multiple servers to set up a Windows Network Load Balancing cluster but you can’t spring for an extra NIC? Remember in those days servers really meant hardware and in the Windows 2000 era you needed Windows 2000 Advanced Server or Windows 2000 Datacenter Server.
What I also will not discuss any further beyond the following is hardware load balancing. Yes good hardware load balancers have extra functions and features that can be very valuable and even necessary for certain deployments. They can be rather expensive for some budgets but they are very capable devices. It is up to you as an engineer to look at the needs, the budget, the risks and benefits of technologies for a business case and come up with good, affordable and working solutions. In some cases that solution will be Windows Load Balancing, in other cases it will be hardware load balancing. Needs, circumstances and environments differ, so do the solutions.
Another thing I’ll wipe of the map from the start is the use of a cross over cable to connect the private NIC. Do not use one. It is not supported and will cause issues or fail.
Then there is the confusion around the use of default Gateways, the fact if the private and the NLB NIC must or must not be on the same subnet, routing and forwarding differences between of Windows 2003 & Windows 2008 (R2). These are the issues I’ll address later in Part 2. But first we need to talk about unicast & multicast a bit. This is unavoidable when using Windows Network Load Balancing. To complete the information here I will provide some examples using two NICs on the same and on different subnets with different default gateway and routing solutions, and also an example using multiple independent clusters (3 NICs)
Things to consider when using unicast & multicast
A topic I will not address too much is which is better: unicast or multicast. Well that depends on the needs, the environment and if the products or solutions uses support it. For example when using VMware guests you’ll have to use multicast if you want it to work without breaking things like VMotion. Another example, ISA server 2006 didn’t support multicast until the release of a hotfix that was later included in SP1 and higher). It also depends on the network gear that’s available, etc.
My take on it all is the following. Use what works best given the circumstances. I you have no access to the switch configuration or your networking gear has issues with multicast NLB you can whine all day long that it’s better than unicast but you’ll won’t get anywhere. When practical I use unicast with multiple NICs and when the circumstances or the products used allow for it, I use multicast with multiple NICs. Which is best is a discussion that sometimes smells of “mine is bigger than yours” and I hope you never had that phase and if you did, you’ve left that far behind together with your other growing pains. Thank you.
Why are Unicast & Multicast so Important
Unicast or multicast mode defines how the cluster virtual IP its MAC address is handled. The network traffic sends packets for the cluster virtual IP based on the cluster MAC address advertised by the cluster. The cluster virtual IP MAC address is used because all traffic for the NLB cluster need be delivered to all nodes.
I will not go into detail on how unicast and multicast works. That has been done very well on CISCO’s web site http://www.cisco.com/en/US/products/hw/switches/ps708/products_configuration_example09186a0080a07203.shtml), TechNet (http://technet.microsoft.com/en-us/library/cc782694(WS.10).aspx) and by Thomas Shindler (http://www.isaserver.org/articles/basicnlbpart2.html)
Unicast issues to consider
- You need two NICs ports. This is because of the “bogus MAC address” (see the CISCO link above for an explanation). Oh please … give me a break already! Again don’t even consider using a single NIC NLB solution in production.
- Port Flooding can’t be stopped on the switch level. A valid argument in many cases.
- It does work in most environments and with just about all network gear.
The good news is that you can prevent flooding by using a hub or a switch configured as a hub to in front of the upstream switch. If you have enough nodes in the NLB this might be a good way to go as you will be attaching 8, 16 or more nodes anyway. If you have only two or three nodes that might be a bit overkill that takes up room in the rack and uses power. Another ways is to uses VLAN to separate the traffic. This works well unless you have a need for the NLB subnet to be the same as the rest or can’t get it configured (politics, rules, existing environment …)
Multicast issues to consider
- You can use a one NIC solution. Multicast allows setting up an NLB cluster with only one NIC which, by some, is considered a benefit. I think I was very clear already about this. I never implement single NIC Windows Network Load Balancing solutions.
- Port Flooding. But here we have some good news for switch admins. Multicast also allows you to stop port flooding by using static arp entries on the switches upstream of your server. This is very valuable. When you only have a couple of nodes in the NLB or can’t create or use VLANs to separate the NLB traffic this is a very good reason to use multicast. See also http://www.cisco.com/en/US/products/hw/switches/ps708/products_configuration_example09186a0080a07203.shtml. This one of the reasons multicast is considered better by some people, but as mentioned you can prevent flooding by using a “hub” in front of the upstream switch or by separating the traffic using another VLAN which for lager NLB clusters is not that much overhead. You might still need to do that if for some reason the static arp solution on the switch ports of the NLB NICs can’t be done. You can also use IGMP snooping to examine the contents of multicast packets and associate a port with a multicast address. If this is not possible the static arp entries come mentioned above do the job.
- As mentioned on TechNet (http://technet.microsoft.com/en-us/library/cc782694(WS.10).aspx)upstream routers might not support mapping a unicast IP address (the cluster IP address) with a multicast MAC address. In these situations, you must upgrade or replace the router. If that’s not possible than you can’t use multicast.
- So you’ll need to talk to your network people (or to yourself if you do the networking as well) to get it figured out and see what they prefer, allow, tolerate and support.
Virtualization comes into the picture
In a virtualize environment the discussion on the “best” way of preventing port flooding also changes a bit. You don’t need so many physical ports but they do often become more scares and valuable as the number of NIC ports on the virtualization hosts are limited. Also a lot of virtualization technologies need their specific little tweaks to get stuff working right depending on the version etc.
Closing thoughts on unicast/multicast
So in the end when choosing between unicast and multicast NLB take a long had look at the environment, the possibilities and needs, the politics, available skillsets than pick the one that is best suited for that particular situation. It’s not that on an issue until you meet some CISCO or Juniper networking guru’s who’ll jab on for hours on how the NLB/multicast implementation sucks.
In part 2 we’ll talk a bit about subnets, default gateways, routing, forwarding and the strong host model in Windows 2008 (R2).