Latency Across Cloud Backbones Varies Significantly

Azure Network

One of the most exciting developments to happen in enterprise networking has been the rise of private backbones. Alibaba, Azure, AWS, and the Google Cloud Platform (GCP) — all can connect your SD-WAN appliances, offering lower, more predictable latency than the open Internet. 

We’ve already spoken about the problems of the open Internet and the benefits of private backbones. You can read those results here. Between Internet regions, such from the US or Europe to the Asia Pacific, latency is often too high and unpredictable for enterprise-grade connectivity. However, even within well-developed Internet regions, Internet performance can be too erratic for some applications depending on the route.

A recent report evaluated the performance of the major independent backbones gathering latency and loss metrics from several locations across each backbone.  It provides a picture that’s relevant to any company looking to reduce the cost of their global MPLS network without exposing their SD-WAN to the unpredictability of the Internet. Latencies between backbones varied as much as 22 percent but that was only part of the story. Even backbones with less latency may not provide optimum end-to-end performance.

Backbone Performance Starts Before the Backbone

The report, Cloud Performance Benchmark from ThousandEyes, took a close look at the differences between cloud backbones. Normally, backbone analysis starts by comparing latency between the PoPs of the backbone. But, as the report demonstrates, one of the biggest differences between cloud backbones happens before traffic enters those networks. 

The backbones differ significantly in where they ingress incoming traffic, a difference that significantly impacts your end-to-end performance.  Traffic destined for datacenters on AWS and Alibaba Cloud enter their respective backbone closest to the target region. This means that the traffic will be subject to the unpredictability of the global Internet longer. By contrast, traffic carried across GCP and Azure enter the backbones closest to the source, minimizing exposure to the unpredictability and latency of the global Internet. 

Backbone Latency Differs Significantly

We were particularly interested in understanding how latency compared across backbones. ThousandsEyes measured bidirectional latency between Availability Zones(AZ) and regions within cloud providers. We took a closer look at three routes of interest to SASE Experts clients: 

  • Across the continental United States (Redmond, Virginia to Santa Clara, California);
  • Transpacific (Santa Clara, California to Mumbai, India)
  • Europe to AsiaPac (London to Tokyo)

Apples-to-apples comparison of the numbers has to be taken carefully as provider PoPs did not all sit in the same cities. Within Europe, the Azure PoP monitored by ThousandEyes is based in Ireland, not England. Within the US, the GCP West PoP monitored by ThousandEyes is in Dallas, OregonR, while others are based in Santa Clara, California. 

Overall, latency across the US was very comparable for AWS, Azure, and GCP. Between Mumbai and the US, GCP exhibited 10 to 15 percent less latency than the other backbones. London to Tokyo, latency was more comparable with AWS, the best, exhibiting only 4 percent less latency than the worst, Alibaba. 

Not shown here is the performance between Europe and Mumbai. The ThousandEyes report calls out that route in particular. GCP continues to show weaker performance on that route due to a lack of direct connectivity on the GCP Backbone, says the report. The report also found that GCP exhibits 2.5-3.0x the network latency in comparison to AWS, Azure, and Alibaba Cloud from Europe to the respective cloud provider regions in India.

Our Take

The development of private independent backbones has given global infrastructure managers a key weapon in their arsenal of IT tactics for reducing global WAN costs. All of the backbones exhibit significantly lower latency than what we’ve seen across the Internet. 

However, looking at backbone performance alone is misleading. You should also consider the location of your own offices, backbone provider policies for ingressing traffic, the other services being provided with the backbone, and the degree of management. 

You might also find these articles of interest:

How to calculate the Latency of a Wide Area Network or IEPL

Internet Variability and SD-WAN

Share this post