Azure VNET to VNET latency

The client I’m currently working for is starting a fairly large migration of it’s applications to the azure cloud platform. One aspect of this that I’ve been interested in exploring is the performance of the network stack, both throughput and VNet latency. A logical way of separating environments is by creating VNET’s and joining them with VNET to VNET VPN’s which are built in natively.

I’m a little surprised about the VNet’s latency, even in the same datacenter. I have deployed two machines in two separate VNET’s in the same geographic region (North Europe – Dublin) and the standard latency is 2-3ms but it frequently jumps to 6-9ms. While this is just an observation using a standard windows ping, I plan to look into this more closely, a little more scientifically. iperf and throughput stats to come!

Are there any applications that are so latency dependent that this alone would prevent you from moving said applications into cloud infrastructure? Low latency on an on premises network is almost so this might be important to you when considering a move away from on-prem. ‘Noisy Neighbours’ are more easily dealt with when you look after the network end to end of course, troubleshooting this latency variation could be difficult.

 

2 comments

  1. Hi Stew,

    You mentioned that you will do some more test for the VMs in seperate Vnets in same region. Do you mind to share some the details with me?
    Thanks a lot!

    1. Hello. Unfortunately as we have moved to express route (rather than VPN’s) I no longer are in a position to be able to do this testing. I experience similar latency ‘strangeness’ even on express route and frequently see changes of 10-20ms in the space of a few hours. In the morning I have response times of around 9ms to the Amsterdam data center, and then later that day it’s 30ms. We have started to design our applications with this in mind.

Leave a Reply

Your email address will not be published. Required fields are marked *

This site uses Akismet to reduce spam. Learn how your comment data is processed.