Jumbo Frame Settings & Slow or Failing Live Migrations over SMB Direct

The Problem

I recently had to trouble shoot a Windows Server 2012 R2 Hyper-V cluster where SMB Direct is leveraged for live migration. It seemed to work, sometime perfectly but at times it but it was in “slow” motion. The VMs got queued for live migration, it took some time for it started and sometimes it would finish or it would fail. This did not happen between all the nodes. I diligently checked out the SMB Direct network but that was OK on all nodes. Basically the LM network was perfectly fine.

To me this indicated that the hosts potentially had issues communicating with each other to coordinate the live migration. But pings and such looked good, there was connectivity, on the surface all seemed well.  In the event log details we saw indications that this was indeed the case. Unfortunately I did not get the opportunity to take screenshots or copies of the events in this particular situation.

The nodes had a separate 2*1Gbps native team LAN access and backups. But diving deeper I noticed that they had set Jumbo Frames on some of those member NICs and not on others. So these setting differed from node to node and that was leading to the symptoms we described above.


You can use Jumbo Frames on your live migration network. Testing has shown this to be beneficial. When you’re doing SMB direct it won’t make such a big difference but it doen not hurt. When SMB Direct fails you’ll fall back to SMB with Multichannel and there it helps more! See Live Migration Can Benefit From Jumbo Frames. While SMB Direct (infiniband, RoCE & iWarp) know Jumbo frames the limited testing I have ever done there indicates only a small increase (2%) in throughput so I’m not sure it’s even worthwhile when doing RDMA.

When you can use Jumbo Frames on you host LAN NIC or team of NICs (handy is you use it to do backups as well)  you need to be consistent end to end. Meaning ALL hosts, ALL NICS & all switches/ switch ports. Being inconsistent in this on the cluster nodes  was what cause the slow to failing live migrations. You need to have good communications between the hosts themselves and AD. Just unplug the LAN from a Hyper-V cluster host to demo this => live migration from to that node and the rest of the cluster won’t work. Mismatching Jumbo Frames or potentially other network settings make this less obvious.  Another “fun” example to trouble shoot is a NIC team where the member NICs are in different VLANs.

Live Migration Can Benefit From Jumbo Frames

Does live migration benefit from Jumbo frames? This question always comes back so I’d just blog it hear again even if I have mentioned it as part of other blog posts. Yes it does! How do I know. Because I’ve tested and used it with Windows Server 2008 R2, 2012 & 2012 R2. Why? because I have a couple of mantra’s:

  • Assumption are the mother of all fuckups
  • Assume makes an ASS out of U and ME
  • Trust but verify

What can I say. I have been doing 10Gbps since for Live Migration with Hyper-V. And let me tell you my experiences with an otherwise completely optimized server (mainly BIOS performance settings): It will help you with up to 20% more bandwidth use.

And thanks to Windows Server 2012 R2 supporting SMB for live migration we can very nicely visualize this with 2*10Gbps NICS, not teamed, used by live migration leveraging SMB Multichannel. On one of the 10Gbps we enable Jumbo Frames on the other one we do not. We than live migrate a large memory VM back and forth. Now you tell me which one is which.


Now enable Jumbo frames on both 10Gbps NICs and again we live migrate the large memory VM back and forth. More bandwidth used, faster live migration.


I can’t make it any more clear. No jumbo frames will not kill your performance unless you have it messed up end to end. Don’t worry if you have a cheaper switch where you can only enable it switch wide instead op port per port. The switch is a pass through. So unless you set messed up sizes on sender/receiving host that the switch in between can’t handle, it will work even without jumbo frames and without heaven falling down on your head Smile. Configure it correctly, test it, and you’ll see.

Enabling Jumbo Frames Inside Virtual Machines Enhances Throughput & Reduces CPU Load

Let’s play a bit with a Windows Server 2012 R2 Hyper-V cluster with 2*10Gbps Intel X520 teamed switch independent and in dynamic mode, which is optimal for DVMQ. On these NICs we enabled Jumbo frames (and on the switches of course). That team is used to create a virtual switch for consumption by the virtual machines. The switches used are 2*DELL PowerConnect 8132F. So we have full fault tolerance. But the important thing to note is that this is commodity, quality hardware that we can leverage for great results.

Now we’ll compare 2 scenarios. In both test a sending VM will try to saturate a receiving VMs network bandwidth. Due to how this NIC teaming setup works, that’s about 10Gbps in a two member 2*10Gbps team. We work around this by leaving both VMs on the same host, so the traffic doesn’t need to pass across the wire. The VMs have VMQ & vRSS enabled with the host team members having VMQ enabled.

Jumbo frames disabled inside the VM (Both VMs on same host)

Without Jumbo Frames enabled in the Guest VM and all other things being equal the very best we can achieve non-sustained is 21Gbps (average +/- 17Gbps)  receiving traffic in a VM. Not bad, not bad at all.


In the picture you can see the host with DVMQ doing it’s job at the left while vRSS is at work in the VM. Pretty clear.

Jumbo frames enabled inside the VM

Now, let’s enable Jumbo frames in the VM.  Fire up PowerShell or use the GUI.


Don’t forget to do this on both the sending and the receiving VM Smile. Here we get +/- 30Gbps receiving traffic inside of the VM. A nice improvement isn’t it? Not just that but we consume less CPU resources as well! Sweet Smile


Useful Power at your finger tips or just showing off?

vRSS & DVMQ is one of may scalability & performance improvements in Windows 2012 R2. And yes, Jumbo Frames inside the VM do make a difference but in a 10Gbps environment it’s not  “in your face” that visible. The 10Gbps limit of a single NIC team member makes this your bottle neck. But it DOES help to reduce CPU cycles in that case. Just look at the two screen shots below.

No Jumbo frames in VM (sending & receiving VMs on different hosts)


We’re consuming 7% CPU resources on the host and 15% in the VM.

Jumbo Frames in VM (sending & receiving VMs on different hosts)


We’ve dropped down to 3% on the host and to 9% inside the VM. All bits help I say!

How far can we push this?

Again if you take the NIC Team bottleneck out of the way you can see some serious differences. Take a look at the screenshot below, that’s 36,2Gbps inside of a VM courtesy of vRSS during some other experiments. Tallyho!


So let’s face it, I guess we’ll need some faster memory (DDR4) and multiple 40Gbps/100Gbps Cards to see what the limits of Windows Server 2012 R2 are or to find out if we reached it. Right now the operating system is giving hardware people a run for their money.

Also note that if you have a number of VMs doing a lot of network IO you’ll be using quite a number of CPU cycles. While vRSS & DVMQ make this scale you might want to consider leveraging SMB Direct for the various tastes of live migration as this will definitely help you out on that front as the NIC will do the heavy lifting.

Reality Check

But perhaps we also need little reality check. While 100Gbps and DDR4 is very nice you might not need it for your current needs. When the environment is built right you’ll find that your apps are usually your limiting factor before the hardware, let alone Windows Server 2012 R2.  So why is knowing this important? Well I verify Microsoft claims so I can talk from experience and not just from what I read in a Microsoft presentation. Secondly you can trust that your investment in Windows Server 2012 R2 is going to carry you long and far. It’s future proofed and that’s god for you. Both when your when needs grow exponential and for the longevity of your environment. Third, we can leverage this to virtualize high through put environments and get the best possible results and ROI.

Also, please, please test & find out, verify what settings suit your environment best and just just blindly enable stuff. Good luck!

Direct Connect iSCSI Storage To Hyper-V Guest Benefits From VMQ & Jumbo Frames

As I was preparing a presentation on Hyper-V cluster high available & high performance networking by, you guessed it, presenting it. During that presentation I mentioned Jumbo Frames & VMQ (VMDq in Intel speak)  for the virtual machine, Live Migration and CSV network. Jumbo frames are rather well know nowadays but VMQ is still something people have read about, at best have tinkered with, but no many are using it in production.

One of the reason for this that it isn’t explained and documented very well. You can find some decent explanation on what it is and does for you but that’s about it. The implementation information is woefully inadequate and, as with many advanced network features, there are many hiccups and intricacies. But that’s a subject for another blog post. I need some more input from Intel and or MSFT before I can finish that one.

Someone stated/asked that they knew that Jumbo frames are good for throughput on iSCSI networks and as such would also be beneficial to iSCSI networks provided to the virtual machines. But how about VMQ? Does that do anything at all for IP based storage. Yes it does. As a matter of fact It’s highly recommend by MSFT IT in one of their TechEd 2010 USA presentations on Hyper-V and storage.

So yes enable VMQ on both NIC ports used for iSCSI to the guest. Ideally these are two dedicated NICs connected to two separate switches to avoid a single point of failure. You do not need to team these on the host or have Multiple Path I/O (MPIO) running for this mat the parent level. The MPIO part is done in the virtual machines guests themselves as that’s where the iSCSI initiator lives with direct connect. And to address the question that followed, you can also use Multiple Connections per Session (MCS) in the guest if your storage device supports this but I must admit I have not seen this used in the wild. And then, finally coming to the point, both MPIO and MCS work transparently with Jumbo Frames and VMQ. So you’re good to go Smile