Wednesday, May 23, 2012

NFS traffic rate limiting on Juniper switches

"And now for something completely different"... 

The configuration for the ESXi infrastructure I worked with is a bit tricky. VMs are hosted on NFS filer, but the same VMs are mounting NFS exports from the same filer. Much like in the picture below:

The physical bandwidth between the access switch where ESXi are connected and the core switches where the Filer is connected is limited. So, I had a lot of ESXi servers, with a lot more VMs competing over the same physical links. Lucky me, I had control over the access switches - Juniper. The next step was elementary: ensure enough bandwidth for ESXi and give some to the VMs - rate limiting using firewall filtering from JunOS.

And this is how it was done. First, a policer was created for NFS traffic coming from VMs (guest OS) which limits the allocated bandwidth - in my case 200 Mbits with a burst size of 10 MB. When the limit is reached packets are discarded. This way two things are achieved: a decent 200 Mbit bandwidth  is ensured for the VMs and small files (up to 10 MB) are transferred very fast to the Filer (no limits). When the VMs demand a lot of resources, the policer steps in and ensures that the critical ESXi vmk traffic gets its share.

[edit firewall]
set policer policer-NFS-1 if-exceeding bandwidth-limit 200m burst-size-limit 10m
set policer policer-NFS-1 then discard

Then, the firewall filter is created. The filter matches all traffic that goes to the IP address of the Filer and applies the policer to it:

[edit firewall family inet]
set filter limit-vlan100-NFS term term-1 from destination-address
set filter limit-vlan100-NFS term term-1 then policer policer-NFS-1
set filter limit-vlan100-NFS term term-default then accept

Last, add firewall filter is applied to the interface - in this case it is VLAN 100:

set interfaces vlan unit 100 family inet filter input limit-vlan100-NFS

The downside is that firewall filtering adds a bit of a load on the CPUs of the switches. Care should be taken when implementing such solutions (as always).

No comments: