SDN and Open Ethernet Switches Empower Modern Data Center Networks Jeff Doyle Big Switch Networks (Jeff.Doyle@bigswitch.com) April 2015 1
What Haunts Data Center Admins? Poor visibility into the network Long lead times for bug fixes Proprietary switch internals Poor scalability April 2015 2
The Value of Open Communities More brainpower Better software security Better hardware reliability Better interoperability Faster vulnerability identification You get a say [Facebook] has tracked over $2 billion in savings via optimizations to its data center, software, and network. The bottom line for us is actually pretty large. --Jay Parikh, VP of Engineering April 2015 3
Traditional Switches: Vertical Integration Feature 1 Feature 2 Provisioning and Management Static, manual configuration Low feature velocity Proprietary Network OS Operating Systems Few API s, only CLI (closed OS) Not externally programmable Proprietary System Hardware Systems Lock-in to a particular vendor Proprietary Silicon System Silicon Slow innovation cycles Expensive, no economies of scale April 2015 4
Open Systems Are Not a New Concept Open Architecture Multi-Level Vendor Choice Innovation Velocity Low TCO Apache MySQL Nagios Linux Custom Web App Hadoop etc. Windows VMware KVM Xen Applications 3 rd party or Custom Operating System Open or closed source Virtualized or bare metal Many support models Dell HP Super Micro Hardware Systems Fierce competition Branded or white box Intel AMD System Silicon Competition and rapid innovation April 2015 5
Open Architecture Multi-Level Vendor Choice Innovation Velocity Low TCO Open Switches: This Looks Familiar SDN OpenFlow BSN Switch Light Pica8 Traditional OSPF STP STP BGP IGMP BGP Cumulus Application Re-create existing distributed protocols New SDN-based applications Network Operating System Open or closed source New Ecosystem Quanta Celestica Accton Dell Alpha ODM and Brand Name Companies Broadcom Mellanox Marvell Centec Intel Merchant Silicon Growing number of startups! April 2015 6
BACKPLANE Traditional Redundant Chassis Switches SUPERVISOR 1 Centralized Control, Distributed Switching: This Looks Familiar Too BACKPLANE SUPERVISOR 2 Supervisor Backplane SDN Controllers AB Spine Switches Singe point of control & management for all chassis linecards User does not log into linecards User does not upgrade linecards No L2/L3 protocols inside the chassis No user access to backplane April 2015 Line Cards Leaf (ToR) Switches Redundant Logical Chassis One Big Switch Extend all the benefits of chassis to SDN Clos fabric 7
Lessons from Altoona Facebook s 4-post aggregated cluster architecture Source: http://nathanfarrington.com/papers/facebook-oic13.pdf Drawbacks stem from very large CSWs and FCs: 1 box failure reduces intra- or inter- cluster capacity by 25% Reduced vendor choice Disproportionately high per-port CAPEX and OPEX Proprietary internals: Complicate management Prevent customization Extend waits for bug fixes Large switches tend to be internally oversubscribed Not all ports can be used simultaneously CSW port densities: Limit scale and bandwidth Slow transition to higher port speeds Difficult to manage machine-to-machine traffic April 2015 8
Lessons from Altoona: Open Switches Wedge currently running at scale as ToR switches - 6-Pack still in test Linux-based operating system choices FBOSS / OpenBMC Big Switch (SDN) Cumulus (local) Soon to be commercially available via Accton Challenge: Smaller switches require a new architecture April 2015 9
Lessons from Altoona: Pod Architecture Standard unit of network 48 ToR switches 40G uplinks 4 Fabric Switches Source: https://code.facebook.com/posts/360346274145943/introducing-data-center-fabric-the-next-generation-facebook-data-center-network/ April 2015 10
Lessons from Altoona: Pod and Spine April 2015 11 Source: https://code.facebook.com/posts/360346274145943/introducing-data-center-fabric-the-next-generation-facebook-data-center-network/
Lessons from Altoona: Pod and Spine April 2015 12 Source: https://code.facebook.com/posts/360346274145943/introducing-data-center-fabric-the-next-generation-facebook-data-center-network/
Lessons from Altoona: Key Takeaways Very large, very scalable architectures can be built with small, open switches Improved: Management Visibility Flexibility / Agility MTTR Resiliency Innovation Hyper-Scalability The same hyperscale principles can be applied to much smaller DC networks April 2015 13
The Bottom Line: Reduced CAPEX 32x 40G $7,495 $234 per 40G port 48x 10G + 4x 40G $5,999 Recommended Comparison Shopping Reading: Arista 7050SX-72-F, 48x 10G + 6x 40G $29,995 Bare Metal Switches Is There a Cost Benefit? Juniper QFX5100-24Q, 24x 40G $28,304 $1179 per 40G port April 2015 http://cumulusnetworks.com/media/cumulus/pdf/misc/business-brief-capex.pdf 14
The Bottom Line: Reduced OPEX Single pane of glass control Zero-touch provisioning Programmability / Automation Integration with orchestration Human error sharply reduced Higher utilization of existing resources (okay, this is CAPEX) April 2015 15
In Summary Open switches are key to CAPEX savings + Customization + Scaling + Flexibility / Agility / Innovation + Fast transitions SDN is key to OPEX saving + Faster provisioning + Faster troubleshooting + Improved visibility + Improved security + Improved planning April 2015 16
Thank You! Questions? Want to discuss Open Switching further? Come see us at the Beer n Pizza Party Open Switching table! Questions: info@bigswitch.com Webinars: www.bigswitch.com/webinars On-line Labs: labs.bigswitch.com April 2015 17