LightTrends Newsletter
Google uses optical circuit switching for its data centers
Summary and highlights of the Sigcomm 2022 networking conference held from August 22-26 in Amsterdam, The Netherlands.
- Google has revealed that it has been using optical circuit switches in its data centers for the past five years.
- The 136x136 port optical circuit switch is Google's design.
- Other Sigcomm-reported research includes:
- Telefónica discussing its experiences with artificial intelligence (AI)
- Understanding mobile access performance
- Avoiding WAN congestion using prediction
- Juniper Network's latest generation Trio routing chipset and how it benefits machine-learning workloads.
- In his keynote, Columbia University's Professor Henning Schulzrinne compared networking science to civil engineering, the oldest engineering discipline. Lessons drawn also apply to telecom.
Venue for the Sigcomm 2022 conference: The Beurs van Berlage is a former stock exchange built by The Government of Amsterdam at the end of the 19th century.
Google is using optical switching in its data centers. The hyperscaler made the revelation at Sigcomm.
"We have eliminated the spine layer in favor of directly connecting the aggregation blocks in a mesh," says Rui Wang, a co-author of Google's Sigcomm paper.
Google rarely details its data center architecture. Indeed, it discussed previously its Jupiter data center network in a 2015 Sigcomm paper.
Now, Google has revealed that it started using optical circuit switches in its data centers five years ago.
Google has since shifted all its Jupiter data center deployments from the classic three-tier Clos interconnect architecture to a mesh-based one using optical switching.
The spine - the top layer of the classic three-tier Clos architecture - is the core traffic aggregation and interconnect layer linking tens of thousands of servers in a data center. The spine's core networking role requires that it is architected before the data center comes online.
“The spine makes Jupiter's evolution expensive,” says Wang.
Data centers are continually being updated with servers, top-of-rack switches, and aggregation blocks. Unfortunately, the upgrading causes the spine to become a performance bottleneck as the links of newly added aggregation blocks fall back in speed to match those of the spine. Equally, upgrading the data center's spine is impractical given the lengthy disruption and lost revenues.
The full text of the research note is available to LightCounting subscribers by logging into their accounts.