Latency kills

Introduction

I was investigating a very problematic Windows Server 2016 Hyper-V cluster. That cluster was just performing horribly. “Everything” was hanging, stalling, crashing and RHS.exe errors where flying around while WER dumps got created by the dozen. Things were extremely slow up to the points functionality was just failing. The “fun” thing was that the cluster validation wizard while slow gave that cluster a big thumb up and a supported status as all was well.

Prying around

Time to pry around a bit and see if we could find something wrong. We save live migrations stall, fail, last forever in pending or get stuck at a certain percentage, sometimes finally succeeding with ridiculous blackout times. We could not open up virtual machine properties or very slowly. The FCI GUI was highly unresponsive but so was the Hyper-V Manager GUI or even PowerShell. Those were hanging at even loading the virtual machines or enumerating them with Get-VM. Everything was slow to the point it timed out or crashed. Restarting the services (Cluster, Hyper-V) didn’t do anything and restarting VMMS was super slow or just got stuck. It was a depressing sight for which people tended to blame Hyper-V / Microsoft.

As the title gives away it was latency. Not just ordinary high latency. Real bad latency. That kind of latency kills. Extreme latency produces symptoms that are similar to bugs or corrupt components of roles and features. We have a tendency to look at those first in the event logs and then we look at the network and its usual suspects (VMQ, SET, DCB). But nothing pointed to an issue that I could find.

So, storage maybe?Well we did find one Hyper-V host in the cluster with one HBA port producing too many error so we disabled that FC port for testing. No joy the Hyper-V cluster after a clean reboot of all nodes remained problematic. So on to the storage array itself.

Well holy smoke! On the two volumes for CSV in those cluster we saw latencies that were so bad I could not even believe a single VM would boot. It actually made my appreciation for Hyper-V and clustering grow as it managed to do at least a couple of things. With such latencies I would expect the services to just crash & call it a day.

clip_image002

The horrific latency on one of the CSV LUNs.

Looking at the logs we saw that the latencies occurred on the FC HBAs of the controllers. Each one above 50ms, peaking to 150-250ms and one huge peak at almost 500ms. We saw this on all four HBA’s.

clip_image004

The latency on one of the 4 FC HBA’s on one of the controllers. Not a good day. All HBAs had high latencies like this.

The issues were not at the host level (host HBA’s) or not even at the IOPS/bandwidth level of the storage itself. The latency for some reason was spiking. Further investigation lead to the conclusion that the issue was related to synchronous replication going totally wrong. Moving the replication mode to asynchronous fixed that. We’re now investigating why this happened and how to prevent this from happening again. But that’s another story.

clip_image006

Latency on one of the 4 FC HBAs on one of the controllers after we fixed the issue.

Do not assume anything

So, there you go. Everything depends on everything in some direct or indirect way. It’s all connected and that my friends, is why I’m a proponent of “service resilience engineering” where the responsible team owns the entire stack. That’s is how you can act fast.

A First look at Cloud Witness

Introduction

In Windows Server 2012 R2 Failover Clustering we have 2 types of witness:

  1. Disk witness: a shared disk that can be seen by all cluster nodes
  2. File Share Witness (FSW): An SMB 3 file share that is accessible by all cluster nodes

Since Windows Server 2012 R2 the recommendation is to always configure a witness. The reason for this is that thanks to dynamic quorum and dynamic witness. These two capabilities offer the best possible resiliency without administrator intervention and are enabled by default. The cluster dynamically assigns a quorum vote to node when it’s up and removes it when it’s down. Likewise, the witness is given a vote when it’s better to have a witness, if you’re better off without the witness it won’t get a vote. That’s why Microsoft now advises to always set a witness, it will be managed automatically. The result of this is that you’ll get the best possible uptime for a cluster under any given circumstance.

This is still the case in Windows Server 2016 but Failover clustering does introduce a new option witness option: cloud witness.

Why do we need a cloud witness?

For certain scenarios such a cluster without shared storage and especially when a stretched cluster is involved you’ll have to use a FSW. It’s a great solution that works as well as a disk witness in most cases. Why do I say most? Well there is a scenario where a disk witness will provide better resiliency, but let’s not go there now.

Now the caveat here is that you’ll need to place the FSW in a 3rd independent site. That’s a hard order for many to fulfill. You can put in on the desktop of the receptionist at a branch office or on a virtual machine on the cluster itself but it’s “suboptimal”. Ideally the FSW is independent and high available not dependent on what it’s supposed to support in achieving quorum.

One of the other workarounds was to extend AD to Azure, deploy a SOFS Cluster with an non CA file share on a cluster of VMs in Azure and have both other sites have access to it over VPN or express route. That works but in a time of easy, fast, cheap and good solutions it’s still serious effort, just for a file share.

As Microsoft has more and more use cases that require a FSW (site aware stretched clusters, Storage Spaces Direct, Exchange DAG, SQL Availability Groups, workgroup or multi domain clusters) they had to find a solution for the growing number of customers that do not have a 3rd site but do need a FSW. The cloud idea above is great but the implementation isn’t the best as it’s rather complex and expensive. Bar using virtual machines you can’t use Azure file services in the cloud as those are primarily for consumption by applications and the security is done via not via ACLing but access keys. That means the security for the Cluster Name Object (CNO) can’t’ be set. So even when you can expose a cloud file to on premises to Windows 2016 (any OS that supports SMB 3 actually) by mapping it via NET USE the cluster GUI can’t set the required security for the cluster nodes so it will fail. And no you can’t set it manually either. Just to prove this I tried it for you to save you the trouble. Do NOT even go there!

clip_image002

So what is possible? Well come Windows Server 2016 Failover Clustering now has a 3rd type of witness. The cloud witness. Functionally wise it’s like a FSW. The big difference it’s a dedicated, cloud based solution that mitigates the need and costs for a 3rd data center and avoids the cost of the workarounds people came up with.

Implementing the cloud witness

In your Azure subscription you create a storage account, for this purpose I’ve create one named democloudwitness in my resource group RG-Demo. I’m using a separate storage account to keep thing tidy and separated from my other demo storage accounts.

A storage account gets two Access keys and two connection strings. The reason for this is that we you need to regenerate the keys you can have your workloads use the other one this can be done without down time.

clip_image004

In Azure the work is actually already done. The rest will happen on premises on the cluster. We’ll configure the cluster with a witness. In PowerShell this is a one liner.

clip_image006

If you get an error, make sure the information is a correct and you can reach Azure of HTTPS over the internet, VPN or Express Route. You normally do not to use the endpoint parameter, just in the rare case you need to specify a different Azure service endpoint.

The above access key is a fake one by the way, just so you know. Once you’re done Get-ClusterQuorum returns Cloud Witness as QuorumResource.

clip_image008

In the GUI you’ll see

clip_image010

When you open up the Blobs services in your storage account you’ll see that a blob service has been created with a name of msft-cloud-witness. When you select it you’ll see a file with a GUID as the name.

clip_image012

That guid is actually the same as your cluster instance ID that you can find in the registry of your cluster nodes under the HKLM\Cluster key in the string value ClusterInstanceID.

Your storage account can be used for multiple clusters. You’ll just see extra entries each with their own guid.

clip_image014

All this consumes so few resources it’s quite possibly the cheapest ever way of getting a cluster witness. Time will tell.

Things to consider

• Cloud Witness uses the HTTPS REST (NOT SMB 3) interface of the Azure Storage Account service. This means it requires the HTTPS port to be open on all cluster nodes to allow access over the internet. Alternatively an Azure Site-2-Site VPN or Express Route can be used. You’ll need one of those.

• REST means no ACLing for the CNO like on a SMB 3 FSW to be done. Security is handled automatically by the Failover Cluster which doesn’t store the actual access key, but generates a shared access security (SAS) token using the access key and stores it securely.

• The generated SAS Token is valid as long as the access key remains valid. When rotating the primary access key, it is important to first update the cloud witness (on all your clusters that are using that storage account) with the secondary access key before regenerating the primary access key.

• Plan your governance between cluster & Azure admins if these are not the same. I see Azure resources governance being neglected and as a cluster admin it’s nice to have some degree of control or say in the Azure part of the equation.

For completeness I’ll mention that the entire setup of a cloud witness is also very nicely integrated in to the Failover Cluster GUI.

Right click on the desired cluster and select “Configure Cluster Quorum Settings” from menu under “More Actions”

clip_image015

Click through the startup form (unless you’ve never ever done this, then you might want to read it).

clip_image017

Select either “Select the quorum witness” or “Advanced quorum configuration”

clip_image019

We keep the default selection of all nodes.

clip_image021

We select to “Configure a cloud witness”

clip_image023

Type in your Azure storage account name, your primary access key for the “Azure storage account key” and leave the endpoint at its default. You’ll normally won’t need this unless you need to use a different Azure Service Endpoint.

clip_image025

Click “Next”to review what you’re about to do

clip_image027

Click Next again and let the wizard run.

clip_image029

You’ll get a report when it’s done. If you get an error, make sure the information is a correct and you can reach Azure of HTTPS over the internet, VPN or Express Route.

Conclusion

I was pleasantly surprised by how it easy it was to set up a cloud witness. The biggest hurdle for some might be access to Azure in secured environments. The file itself contains no sensitive information at all and while a VPN or Express Route are secured connectivity options this might not be allowed or viable in certain environments. Other than this I have found it to be very reliable, effective cheap and easy. I really encourage you to test it and see what it can do for you.

Client Access & Windows Server 2016 Site Aware Stretched Clusters

Introduction

There’s more to business continuity than having multiple locations. When it comes to high availability, or perhaps more accurately disaster recovery and business continuity people tend to focus on the good news. Some managers don’t want to be bothered by the details of our incompetency (i.e. reality and laws of physics) and vendors only like to focus on what they can sell with the biggest profit margin. Anything raining on that party falls under annoying details. When such a manager and such a sales man find each other it’s a match made in heaven. You’re the one who’s bringing the rain. It comes under the form of a simple question. How are we going to expose the failed over services internally and externally to the users and customers? What you mean that million-dollar investment in multiple SANs, clusters and consultants isn’t sufficient? Nope!

clip_image002

One piece of very good news is that in Windows Server 2016 Failover Clustering we can now leverage a cloud witness as well, next to a file share witness. This has the benefits we do not need a 3rd site for the file share witness. Which was not always feasible, sometimes a bit convoluted to achieve in the cloud via IAAS or depended on a rather less dependable server or PC somewhere in a branch office.

What’s the problem?

The problem is that failing over the workload with the services (VMs, SQL, File Servers, …) in a healthy, consistent state is only part of the challenge. The other part is to make sure that your clients (human or machines) can actually access those failed over services. If required or possible without noticing or with the smallest possible interruption possible. Even when you can achieve failover with only seconds of service interruption, some applications just can handle this gracefully or not at all.

The thing is when you have multiple sites that often means distinct separate subnets / networks. So when that VM with IP address of 10.10.100.124 on default untagged VLAN 100 fails over to the other site how will the clients in the various branch offices or on the internet access it services? DNS point to 10.10.100.124 under normal conditions.

Well when the IP address can be updated for the DNS record thanks to “Multi-Subnet Resource Configuration” (SQL Server, File Share) thing will work again, eventually, given enough time.

clip_image004

Multi-Subnet Resource Configuration works as follows. We have a single network name resource which we make dependent on multiple IP Address resources. In cluster terms that’s a “OR” decency when looking at the validation report. The secret sauce is that only one of the IP address resources of the network name resource is online at any given time. This gets registered in DNS and that’s what the clients use to access the service.

This works but the DNS record need to be upgraded, DNS replication needs to happen, client their DNS cache needs to expire and update etc. You can be looking at half an hour of down time actually.

But what if Multi-Subnet Resource Configuration isn’t an option or we’re in a hurry? What are options and how well and fast do these work? That’s the point at which the storage vendor is already counting the profits, the PM states the job’s done and the boss has already decided the project is a success and the network guys have some questions about YOUR problems. Let’s discuss some of options to deal with accessing services after a site activation.

Note: Hyper-V replica has the ability inject an alternate IP address on failover but we’re talking about a stretched cluster here, where replication happens at the storage level, not at the application level (Hyper-V) for the virtual machines.

Software Defined Networking Aka Network Virtualization.

Using Hyper-V Network Virtualization (HNV) abstract VMs logical subnet boundaries. This gives each virtual network the illusion it is running as a physical network. The typical example for this is multiple tenants that have the same IP space. The fact that it overlays physical network is also very handy when it comes to one and the same tenants in multi-site scenarios. Virtual networks allow VMs to move across different physical networks without re-configuring IP address in guest OS.

clip_image005

This totally abstracts the networks and it works great for virtual machines (Hyper-V). It doesn’t have to be limited to a single DC or site. Do note that there’s things to discuss around CSVs, Live Migration cluster wise and routing, gateways, DNS, geo load balancing access wise but you get the idea. When it comes to different subnets, different sites in regards to clustering things are not as easy as it seems. For this discussion we’re limiting ourselves to client connectivity to resources that move to another site and don’t dive into the details of network virtualization either.

Network Name Properties

There’s two cluster network name resource property setting you can configure to help reducing downtime after a failover.

RegisterAllProvidersIP cluster network name resource property

Remember our first story of “Multi-Subnet Resource Configuration” with the DNS updates and cache that has to expire? Well this can be enhanced as long as the applications can hand handle it. We can configure the DNS registration behavior via the RegisterAllProvidersIP property of a cluster networks name resource.

Get-ClusterResource MySQLServer |

Set-ClusterParameter RegisterAllProvidersIP 1

By setting this to TRUE all the IP address resources, on line and off line, are registered in DNS. If you have a “enlightened” application that can check for and handle multiple IP addresses and determine which one to use it allows for faster client reconnects. This works great with SQL Server.

HOstRecordTTL cluster network name resource property

This is great but has limited scope as the application has to have the logic to handle multiple registered IP addresses for the same resource and figure out when to use which one. SQL Server can do this, so can Exchange. What about a file server? RegisterAllProvidersIP won’t work but we can reduce the time to live of the DNS record for a cluster network resource IP address on the client from 20 minutes to 5 minutes or lower.

Get-ClusterResource MyFileShare |

Set-ClusterParameter HostRecordTTL 300

This is not an option for Hyper-V, there network virtualization works better or we use other options. Read on!

Stretch your VLANs

Here the VLAN(s) stretch across the sites. This means that the IP address of the service (VMs, SQL Servers, File Shares, …) never changes making it very easy to have the clients reconnect very fast.

clip_image007

Easy for the apps and the system administrators. Well sort of, chances are that the networks admins will chip in and put a kill contract out on you with some assassins. Just saying. In a perfect world this would be a good idea. In reality layer 2 and spanning tree are making sure you’ll sort of regret it or at least deal with the drawback and fall out. Choose wisely.

Abstract the network devices

This is a network vendors provided solution and I don’t see it very much in the wild. In this approach the network devices use a 3rd IP address that get registered in DNS for use by the clients. The fact that the workload switches between subnets when failing over between sites is irrelevant to the clients.

clip_image018

Cisco has this in a couple of solutions where NAT or a VIP is used to achieve this. As this is network appliance/ hardware based it works with any workload.

SLA your way out

Some people “mitigate” the prolonged down time by having a separate SLA for local failover versus site failover. Cool, but if I was cynical I could state that this is just lawyer behavior. You create fine print and “cover your ass” for that scenario. It’s not really solving anything but accepting longer down time and having all involved parties recognize and accept that fact. This is a valid approach.

Be creative & drive towards maximum portability

In an ideal world you can provision apps & services so fast you only need to protect and failover the persistent data. A world of micro services, containers where servers and virtual machines are cattle. But many of us will have to deal with servers being holey cows for now.

The above approaches are the most common options. There are more variations to these. One of those could be bases around the use of a dedicated management domain on both sites. It’s a concept I’ve used a couple of time where and when allowed.

It has some drawbacks or at least some complexities to deal with and one such example might be when configuring host based backups that need access to the guest VMs. This requires some extra firewall configuration. Nothing that would prevent you from doing so with good backup products like Veeam and it’s something you’re probably used to doing already for monitoring and backups across domains anyway.

But it also has serious benefits as the actual business domains are completely separate from the management domain and potentially 100% virtualized but that’s not a hard requirement as long as you keep the remaining physical servers in their own site dependent subnet which routes, these don’t move anyway, and they should have workloads that are distributed anyway like AD, Exchange DAGs, etc. The big benefits compared to a stretched cluster is that you can have the same subnet(s) on both sides of the stretched cluster for your virtual machines and you change the routing and endpoints for your public and private access to the services. Instead of making the changes to the cluster resources you do so higher up at the stack. It’s a bit like moving your data center to new location “as is” and directing the clients to the new location. This removes the need for stretched VLAN, or implementing network virtualization, at the cost of a bit more down time & work to “switch”. It’s worth considering.

It helps to leveraging DNS and geo load balancing technologies in this but the core infrastructure (the site ware stretched cluster) can run in a fully routed / Layer 3 fashion.

Sure you’ll still need to make sure the traffic from the offices goes to the correct data center now and it really rocks if you have your internet presence geo-load balanced in some way but let’s face it. But you needed to have that in order for any approach anyway.

Closing thoughts

There is a lot more detail and complexity to all of this than I covered in this short article. This is meant an eye opener, a point from where to start the discussion with the business demanding 24/7, 99.999% a zero cost and effort. Like Amazon or Azure but then better, cheaper and on premises. Ouch! As you might expect, this can’t be dealt with in just a few pages. Getting a solid, working disaster avoidance, recovery and business continuity plan & process is going to take some effort to create and maintain.

Fully failing over without any work or a second of downtime is a very expensive illusion and you might be better off with 15 to 20 minutes of down time for 90% of the workload and 30 to 60 minutes for the remaining 10% that trying to chase the ultimate perfection of 100% zero downtime ever for all services. Chances are you’ll go broke trying and pretending, which means failing. Remember that when your primary data center was just taken down or worse, burnt down, dealing with a couple of hours of down time to get you secondary site up and running 100 % isn’t actually as bad as it seems when discussing 2 or 3 hours of down time in a management meeting. Somehow it always seems a bigger deal when not faced with the alternative of the business being wiped out.

One final note, don’t forget to tell your bosses you’re going to have to practices this a couple of times per year. Doing it for real count’s a practice only if it’s the 3rd time you do it. Good luck!

The Hyper-V Amigos Episode 10

It’s with great pride that the Hyper-V Amigos ride again and for The Hyper-V Amigos Episode 10 they dive into what’s new and improved in Windows Server 2016 Failover Clustering.

image

Well OK we only discuss a few subjects in this web cast as there is only a limited amount of time. I’ll present an overview of during my session at the German Cloud and Datacenter conference on May 12th in Germany. An hour is not enough for a deep dive into everything but we will build on our session we did at the Technical Summit (November 2014) in Germany on Improvements in Failover Cluster 2012 R2 ad get you up to speed so you can select what to investigate further.

Until then, enjoy the webcast and I hope it helps prepare you for what’s coming and entices you to join us at the Cloud and Datacenter Summit in Germany on May 12th! And if clustering alone is not enough to bring you over check out the agenda and you might realize what great gathering of experts is happing at the conference. Just look at the content, the breath and depth of the cloud and datacenter technologies being discussed is vast!