Server-based open networking taxes the CPU, so offload to an intelligent server adapter

Server-based open networking taxes the CPU, so offload to an intelligent server adapter

This vendor-written tech primer has been edited by Network World to eliminate product promotion, but readers should note it will likely favor the submitter’s approach.

Networking using commercial off the shelf (COTS) servers has been around for several years, thanks to the proliferation of Linux-based servers and network technologies like Open vSwitch (OVS).  The hope is that the switch world follows the servers’ successful path, hence the birth and popularity of the term “open networking.”

Network devices like top of rack (TOR) switches have traditionally been closed -- the operating systems and protocols that run on the switches were proprietary, could not be disaggregated from the hardware and were not open source.

Switches got disaggregated a bit when the switch vendors adopted switch ASICs from merchant silicon vendors like Broadcom. Next came OpenFlow and OpenFlow-based SDN controllers like Floodlight, which proposed that the switch control plane protocols be removed from the switch and placed in an open source controller. This in some ways disaggregated the OS from the switch box.

Subsequently, switch operating systems like Cumulus Linux came to market. These can be installed and run on switch boxes from multiple vendors, like Quanta and Dell. But such disaggregated switch OSes are not necessarily open source.

More recently, open source switch operating systems like SONiC and Open Network Linux have been in the news. The open source controller ecosystem has further evolved as well, focusing on feature completeness and carrier grade reliability (i.e., OpenDaylight and ONOS).

All in all, significant action and news in the realm of open networking has been related to switches, geared toward helping the industry manage the switch supply chain more effectively and deploy efficiently, similar to the COTS server model.

Figure 1: Switch disaggregation follows server model

Figure 1: Switch disaggregation follows server model

Open networking on servers

What seems to get overlooked in these discussions is how open networking on servers (or server-based open networking) has evolved and enabled open networking on switches.

Over the last several years, TOR switches have become simpler because data center traffic patterns have changed and network infrastructure efficiency requirements have increased. When using leaf (TOR) and spine switches, the imperative has shifted to moving east-west traffic most efficiently, which requires more bandwidth, more ports and lower latency. As a result, the feature requirements in hardware and software in leaf and spine switches have been reduced to a simpler set. This has made open networking in switches easier to implement and deploy.

Figure 2: Networking smarts moving to servers

Figure 2: Networking smarts moving to servers

However, the smarts of networking did not disappear – they just moved to the server, where such smarts are implemented using the virtual switch – preferably an open one such as OVS – and other Linux network features like IP tables. Many new features related to network security and load balancing have been added to OVS.

OpenStack, as an open source and centralized cloud orchestration tool, has rapidly come to prominence, with more than 60% of OpenStack networking deployed today using OVS (with OpenStack Neutron). Server-based open networking has evolved relatively quietly compared to open networking in switches, but it has made major contributions toward bringing deployment efficiencies and flexibility.

Figure 3: Poor server efficiency leads to server sprawl

Figure 3: Poor server efficiency leads to server sprawl

Today, in many high growth cloud, SDN and NFV applications, server-based open networking is running into server sprawl and related TCO challenges. As the networking bandwidths increase and the number of VMs proliferates on servers, OVS processing is taking up an increasingly large number of CPU cycles, which is limiting the number of CPU cycles available for processing applications and VMs. Data center operators cannot economically scale their server-based networking using traditional software-based virtual switches. So implementing server-based networking in x86 architectures and software is a double whammy: it increases costs as too many CPU cores are consumed, and it lowers performance as applications are starved for resources. 

Offloading network processing to networking hardware is an option that has worked well in the past.  However, software-defined and open source networking is evolving at a rapid pace; such innovation stops the moment data center operators look to inflexible network hardware for performance and scale.

Figure 4: ISA offloads networking from servers

Figure 4: ISA offloads networking from servers

The solution to this challenge is to offload OVS processing to an intelligent server adapter (ISA). With an ISA handling OVS processing, performance is boosted by up to 5X, and the data center operator frees as many as 11 CPU cores from network-related processing, enabling greater VM scalability and lower costs. An ISA is programmable and hence can evolve rapidly with new features, preserving the pace of innovation.  Although server-based networking by itself can cause server sprawl, ISAs are making the case for efficient and flexible open networking from the COTS server side. 

See www.netronome.com

IDG Insider

PREVIOUS ARTICLE

«Twitter beta adds Snapchat-style QR codes to make following faster

NEXT ARTICLE

Apple saves iPhone call history to iCloud, but barely mentions it»

Add Your Comment

Most Recent Comments

Resource Center

  • /view_company_report/775/aruba-networks
  • /view_company_report/419/splunk

Poll

Crowdfunding: Viable alternative to VC funding or glorified marketing?