What Makes A Good Network Fabric?
As I wrote in my earlier blog post, the data center network is going through a major transformation to support server virtualization and cloud computing, convergence of data storage, application-to-application traffic and new high-performance applications. To address these needs, the data center network fabric, the system network architecture that interconnects server devices and storage devices in a data center environment, has become a critical lynchpin of data center architecture. I am frequently asked the question, “What makes a good network fabric?” The term “fabric” has nearly as many definitions and permutations as “cloud,” so I would like to suggest the top 10 attributes that a data center network fabric should possess.
- Open and Interoperable – Modern data center networking is best accomplished when it is standards-based, when multiple vendors equipment can co-exist and interoperate, and clients can choose between multiple vendors’ wares without paying a pricing penalty or needing to rip-and-replace to meet growth needs and implement next-generation approaches. Standards-based Ethernet is a fabric essential.
- High Speed – To meet the performance needs of big data, cloud computing and workload-optimized systems, data centers are increasingly implementing 10 Gigabit Ethernet on the server and in the access and aggregation layers, which is driving interest in 40 Gigabit Ethernet with 100 Gigabit Ethernet on the horizon.
- Low Latency - Applications such as high-frequency trading require the lowest possible latency. The race to zero latency is enabled by ultra-low-latency switches that deterministic with the same connection speeds across every port combination.
- Loss-less – To meet the needs of machine-to-machine applications and converged data and storage networking, Ethernet networks must be loss-less. Network equipment supports the Data Center Bridging (DCB) standards to ensure loss-less operation.
- Flat – Clos and fat tree network designs are becoming increasingly prevalent for the flow-based, non-blocking, shortest path network fabrics required in highly virtualized and cloud data centers and for converged data and storage traffic. With standards such as Transparent Interconnection of Lots of Links (TRILL) and Shortest Path Bridging (SPB) on the horizon and other alternatives for flat networks requiring proprietary implementations, network architects favor existing approaches to meshed networking such as Virtual Link Aggregation (vLAG) to maximize network efficiency, reduce congestion and address Spanning Tree limitations by creating active/active Layer 2 network paths for load balancing and redundancy.
- Virtualization Aware – To ensure that network security, performance and access policies move dynamically when live virtual machines migrate from server to server anywhere in the virtualized data center or private cloud, the Edge Virtual Bridging (EVB) 802.1Qbg standard specifies the interaction between virtual switching environments in a hypervisor and the first layer of the physical switching infrastructure. The EVB 802.1Qbg standard addresses the conventional vSwitch’s lack of network management, monitoring and security.
- Scalable – Linearly scalable is a requisite of network fabric design, which can be achieved by non-blocking, non-oversubscribed topologies implemented using top-of-rack switches.
- User Control – The emerging OpenFlow specification enables network infrastructure providers to deliver open virtual networking systems that that are easy for users to control, optimize performance dynamically and minimize complexity. OpenFlow-based networks enable the network administrator to easily configure and manage virtual networks that control traffic on a per-flow basis. The administrator can easily program an OpenFlow controller to create or delete multiple independent virtual networks and related policies that span the system network without having to deal with the complexities of the underlying physical network and protocol details.
- Unified Management – Network devices that make up the fabric should be managed, configured and provisioned as if they were a single logical device, including the ability to can track virtual machines by switch or IP address, and pre-provision network characteristics for VMs.
- Low Lifecycle Costs – And finally, per-switch costs should be affordable so networks can scale on an incremental basis, the need for expensive chassis switches should be minimized, and low power requirements and cooling efficiencies should enable low energy costs particularly for massive networks fabrics that interconnect thousands of server and storage systems.
We have designed IBM System Networking’s Unified Fabric Architecture, our fast, smart, interoperable and proven data center Ethernet network fabric, to deliver across all 10 of these attributes. Unified Fabric Architecture along with the ability to replicate pre-provisioned racks and integrated systems, enables clients that have already invested billions of dollars in their data centers to take advantage of the best innovations in the industry and achieve the lowest possible cost of ownership for their IT infrastructure.