dvFilter

ESXi Learnswitch – Enhancement to the ESXi MAC Learn DvFilter

04.24.2017 by William Lam // 23 Comments

The ESXi MAC Learn dvFilter Fling was released a little over two years ago and it has become a must have when it comes to running our ESXi Hypervisor within a VM, also referred to as Nested ESXi. The reason this Fling has become such a popular hit amongst our customers and partners is that it greatly improves the performance when “Promiscuous Mode” is enabled on a Virtual or Distributed Virtual Portgroup, which is a requirement for using Nested ESXi. Although this Fling works great, there are a couple of limitations with this solution today. The first of which is called out in the original Fling release notes, that once a MAC Address has been learned, it never ages out which is not ideal for long running Nested ESXi environments that generates a large amount of new MAC Addresses. The second, is the lack of vMotion support where the learned MAC Address table is not transfered to the destination ESXi host and must be re-learned.

To help address both of these limitations, the folks over in the Network and Security Business Unit (NSBU) have been working hard to improve upon the existing solution and have developed a new native MAC Learning VMkernel module called the Learnswitch. This new Learnswitch not only helps improves Nested ESXi workloads but it can also potentially benefit other workloads such as Nested Containers or other 3rd Party network inspection software. One immediate difference from the previous MAC Learn dvFilter solution is that rather than operating on the Network IO Chain, the filtering is now performed within the outer virtual switch layer itself which will provide some additional performance gains. The other added benefit from an internal VMware standpoint is that the Learnswitch is now vmkapi compatible, which means we will have a better backwards compatible story for supporting old releases of ESXi. One downside to this new solution compared to the previous one is that because the dvFilter operated below the virtual switch layer, it could support both a Virtual Standard Switch as well as the Distributed Virtual Switch. With the new Learnswitch, a Distributed Virtual Switch will be required. If you currently do not meet the requirements of the new Learnswitch, you can continue using the dvFilter, but it is recommended that you do not mix both on a single system but you can definitely make use of both solutions across different ESXi hosts depending on the constraints of your environment.

Here are some of the new capabilities provided by the new Learnswitch module:

Overlay Network based that learning and filtering are done in Etherswitch forwarding check
MAC Address learning is based on VLAN ID or VXLAN ID on uplink and leaf port
Packet is filtered on uplink and leaf port if the MAC is learned on a different port
MAC Address table size is 32k per system
MAC Address aging support with default aging time of 5 minutes and configurable
Unknown unicast packet is flooded by default and configurable to drop
vMotion support that the MAC table learned on the port is transferred to destination host and RARP packet is sent
Standalone VMkernel module available as a VIB
net-learnswitch CLI to display MAC Address table, configuration and stats

[Read more...]

Does the ESXi Mac Learn dvFilter work with Nested ESXi on NSX VXLAN's?

09.19.2014 by William Lam // 3 Comments

After publishing my article on the new ESXi Mac Learn dvFilter which helps improve CPU/Network performance when using promiscuous mode with Nested ESXi, I received a couple of questions asking whether the dvFilter would work with NSX VXLAN's? At the time, I had only tested the Mac Learn dvFilter using standard VSS/VDS and not with any VXLAN based networks. I had reached out to a couple of folks asking whether this would work and to my surprise, I actually got back a mix set of answers to it will not work to it could work. One of the reasons that was given to me on why this may not work is that NSX-v (NSX for vSphere) leverages a different "virtual switch" than VSS/VDS and hence the Mac Learn dvFilter would not properly function. This actually would make sense, but because I received other responses negating that fact, I figured I probably should just test it for myself and see.

NSX 6.1 was recently released and I figured this would be a great opportunity for me to learn a bit more about NSX, as I have never played with it before and also test whether Mac Learn dvFilter would in fact work with NSX VXLAN's. In my lab environment I have deployed NSX and I have 3 physical ESXi hosts running VSAN (go SDS!). I deployed both an NSX ESR (Edge Service Router) hosting 2 Logical Networks (aka VXLAN segments) and an NSX DLR (Distributed Logical Router) hosting another 2 Logical Networks.

Here is a screenshot of the 4 Logical Networks, the first two on NSX ESR and the last two on NSX DLR:

Here is a screenshot of both the NSX ESR and DLR:

Note: If you would like to learn more about NSX ESR and DLR, check out this great article by Brad Hedlund who goes into more detail.

For my test, I first enabled Promiscuous Mode and Forged Transmit on the respective Logical Switches which is just a dvPortgroup on the VDS for my NSX ESR setup. I then had 2 Nested ESXi VMs running (without the Mac Learn dvFilter), a Windows "Jump Box" VM and vMA all connected to the same VLXAN network.

I then transfer an ISO from the Windows VM to vMA while running ESXTOP on the physical ESXi host which is hosting these four VMs. As I expected, both the Nested ESXi VMs and vMA were receiving network packets. Next, I installed the Mac Learn dvFilter VIB on the physical ESXi host and added the required VM Advanced Settings to both the Nested ESXi VMs and then re-ran the test. To my surprise, both the Nested ESXi VMs were no longer receiving the erroneous packets! So it seems that using VLXAN with NSX ESR, the Mac Learn dvFilter is working as expected.

To be thorough, I also ran through same test but now for the VXLAN segments backed by NSX DLR. This time, I was really surprised by the results. The test was prior to installing the Mac Learn dvFilter and my expectation was that the two Nested ESXi VMs would be seeing the duplicated network packets from the VDS, but to my surprise, they did not! Both the Nested ESXi VMs were pretty much idling at 0 packets as nothing was being sent to them. I am not exactly sure why I was seeing this behavior, perhaps there is some type of optimization in the DLR? This is something I hope to get an answer from someone in Engineering on why I might be seeing this positive behavior.

To summarize, this myth has been busted and the Mac Learn dvFilter does in fact work with VXLAN networks. If you are using NSX ESR for your VXLAN setup, then you will need to install the dvFilter and if you are using NSX DLR, it seems like you do not need to make any additional changes. After briefly speaking with Christian Dickmann, the creator of the dvFilter as I wanted to share the results with him, I also learned about some interesting tidbits. Christian was not surprised by the results actually, the reason for this is that the VMkernel networking stack was architected and designed to be modular. This meant that, one could switch out the "virtual switch" with other implementations and the underlying dvFilter framework would still continue to work regardless of the "virtual switch" being used.

Additional Note:

I did not get a chance to test with vCNS and VXLAN, but I believe it should work given NSX-v is functional. If you are able to test this, feel free to leave a comment on whether the expected behavior is seen with the Mac Learn dvFilter.
I did not get a chance to test this with vCloud Director with VXLAN based networks, but as I mentioned, this should work. Please leave a comment if you can confirm
I also noticed when creating the Logical Switches, there is a Mac Learning capability, but from my testing, I found it did not benefited Nested ESXi and the Mac Learn dvFilter was still required.

New VMware Fling to improve Network/CPU performance when using Promiscuous Mode for Nested ESXi

08.28.2014 by William Lam // 44 Comments

I wrote an article awhile back Why is Promiscuous Mode & Forged Transmits required for Nested ESXi? and the primary motivation behind the article was in regards to an observation a customer made while using Nested ESXi. The customer was performing some networking benchmarks on their physical ESXi hosts which happened to be hosting a couple of Nested ESXi VMs as well as regular VMs. The customer concluded in his blog that running Nested ESXi VMs on their physical ESXi hosts actually reduced overall network throughput.

UPDATE (04/24/17) - Please have a look at the new ESXi Learnswitch which is an enhancement to the existing ESXi dvFilter MAC Learn module.

UPDATE (11/30/16) - A new version of the ESXi MAC Learning dvFilter has just been released to support ESXi 6.5, please download v2 for that ESXi release. If you have ESXi 5.x or 6.0, you will need to use the v1 version of the Fling as it is not backwards compat. You can all the details on the Fling page here.

This initially did not click until I started to think about this a bit more and the implications when enabling Promiscuous Mode which I think is something that not many of us are not aware of. At a very high level, Promiscuous Mode allows for proper networking connectivity for our Nested VMs running on top of a Nested ESXi VMs (For the full details, please refer to the blog article above). So why is this a problem and how does this lead to reduced network performance as well as increased CPU load?

The diagram below will hopefully help explain why. Here, I have a single physical ESXi host that is connected to either a VSS (Virtual Standard Switch) or VDS (vSphere Distributed Switch) and I have a portgroup which has Promiscuous Mode enabled and it contains both Nested ESXi VMs as well as regular VMs. Lets say we have 1000 Network Packets destined for our regular VM (highlighted in blue), one would expect that the red boxes (representing the packets) will be forwarded to our regular VM right?

What actually happens is shown in the next diagram below where every Nested ESXi VM as well as other regular VMs within the portgroup that has Promiscuous Mode enabled will receive a copy of those 1000 Network Packets on each of their vNICs even though they were not originally intended for them. This process of performing the shadow copies of the network packets and forwarding them down to the VMs is a very expensive operation. This is why the customer was seeing reduced network performance as well as increased CPU utilization to process all these additional packets that would eventually be discarded by the Nested ESXi VMs.

This really solidified in my head when I logged into my own home lab system which I run anywhere from 15-20 Nested ESXi VMs at any given time in addition to several dozen regular VMs just like any home/development/test lab would. I launched esxtop and set the refresh cycle to 2seconds and switched to the networking view. At the time I was transferring a couple of ESXi ISO’s for my kicskstart server and realized that ALL my Nested ESXi VMs got a copy of those packets.

As you can see from the screenshot above, every single one of my Nested ESXi VMs was receiving ALL traffic from the virtual switch, this definitely adds up to a lot of resources being wasted on my physical ESXi host which could be used for running other workloads.

I decided at this point to reach out to engineering to see if there was anything we could do to help reduce this impact. I initially thought about using NIOC but then realized it was primarily designed for managing outbound traffic where as the Promiscuous Mode traffic is all inbound and it would not actually get rid of the traffic. After speaking to a couple of Engineers, it turns out this issue had been seen in our R&D Cloud (Nimbus) which provides IaaS capabilities to the R&D Organization for quickly spinning up both Virtual/Physical instances for development and testing.

Christian Dickmann was my go to guy for Nimbus and it turns out this particular issue has been seen before. Not only has he seen this behavior, he also had a nice solution to fix the problem in the form of an ESXi dvFilter that implemented MAC Learning! As many of you know our VSS/VDS does not implement MAC Learning as we already know which MAC Addresses are assigned to a particular VM.

I got in touch with Christian and was able to validate his solution in my home lab using the latest ESXi 5.5 release. At this point, I knew I had to get this out to the larger VMware Community and started to work with Christian and our VMware Flings team to see how we can get this released as a Fling.

Today, I am excited to announce the ESXi Mac Learning dvFilter Fling which is distributed as an installable VIB for your physical ESXi host and it provides support for ESXi 5.x & ESXi 6.x

Note: You will need to enable Promiscuous Mode either on the VSS/VDS or specific portgroup/distributed portgroup for this solution to work.

You can download the MAC Learning dvFilter VIB here or you can install directly from the URL shown below:

To install the VIB, run the following ESXCLI command if you have VIB uploaded to your ESXi datastore:

esxcli software vib install -v /vmfs/volumes/<DATASTORE>/vmware-esx-dvfilter-maclearn-0.1-ESX-5.0.vib -f

To install the VIB from the URL directly, run the following ESXCLI command:

esxcli software vib install -v http://download3.vmware.com/software/vmw-tools/esxi-mac-learning-dvfilter/vmware-esx-dvfilter-maclearn-1.0.vib -f

A system reboot is not necessary and you can confirm the dvFilter was successfully installed by running the following command:

/sbin/summarize-dvfilter

You should be able see the new MAC Learning dvFilter listed at the very top of the output.

For the new dvFilter to work, you will need to add two Advanced Virtual Machine Settings to each of your Nested ESXi VMs and this is on a per vNIC basis, which means you will need to add N-entries if you have N-vNICs on your Nested ESXi VM.

ethernet#.filter4.name = dvfilter-maclearn
ethernet#.filter4.onFailure = failOpen

This can be done online without rebooting the Nested ESXi VMs if you leverage the vSphere API. Another way to add this is to shutdown your Nested ESXi VM and use either the “legacy” vSphere C# Client or vSphere Web Client or for those that know how to append and reload the .VMX file as that’s where the configuration file is persisted
on disk.

I normally provision my Nested ESXi VMs with 4 vNICs, so I have four corresponding entries. To confirm the settings are loaded, we can re-run the summarize-dvfilter command and we should now see our Virtual Machine listed in the output along with each vNIC instance.

Once I started to apply this changed across all my Nested ESXi VMs using a script I had written for setting Advanced VM Settings, I immediately saw the decrease of network traffic on ALL my Nested ESXi VMs. For those of you who wish to automate this configuration change, you can take a look at this blog article which includes both a PowerCLI & vSphere SDK for Perl script that can help.

I highly recommend anyone that uses Nested ESXi to ensure you have this VIB installed on all your ESXi hosts! As a best practice you should also ensure that you isolate your other workloads from your Nested ESXi VMs and this will allow you to limit which portgroups must be enabled with Promiscuous Mode.