Conventional Data Center networks (Limitations)
Conventional Data Center networks have 3-4 tires like access layer, aggregation layer, core layer, services layer, etc. Each layer duplicates many of the IP/Ethernet packet analysis and forwarding functions. This adds end-to-end latency.
To reduce costs and make these networks scalable, over-subscription (higher downlink capacity and lower uplink capacity) is typically used for all these tiers. Over-subscription might result in packet losses during network congestion and is not suitable for storage traffic.
STP (Spanning Tree Protocol) is used to block multiple paths (loops) in the network. STP prevents the complete usage of available network bandwidth and the network may have to be rewired to handle changes in application workloads.
Link Aggregation is used to create larger pipes, but it cannot be dynamically modified by the network. Manual intervention is required to create/re-configure link-aggregation to suit the changing application requirements.
Firewalls and load-balancers for each application are dedicated to a group of application servers.
Network management is centered in the switch operating system.
It is difficult to handle Virtual Machine (VM) migration and multi-tenancy (for cloud networks) using conventional data center networks.
Multiple networks maybe used for connectivity (IP/Ethernet), Storage (FC – Fiber Channel), HPC – High Performance Computing (Infiniband), etc.
Modern Data Center networks (Wish-list)
The modern data center network should be able to reduce the 3-4 tiers of the conventional DC network, into a simple 2-tier network (access layer and core layer) in order to easily facilitate the increasing traffic between the servers.
Larger layer-2 domains that are easy to scale, facilitate automatic VM Migration and have lower latencies, could be created.
The modern data center architecture should enable people to partition computing, storage and network resources rapidly and dynamically, in order to accommodate for flexible multi-tenant cloud based services, like a private cloud each for different departments within an organization.
The network intelligence could either be placed in the network core, or at the network edge, or within every server (where the application resides). The third option more readily supports the creation of a dynamic logical network that provides for easier application mobility.
Self aggregating logical ISL (Inter Switch Links) could replace manually configured Link Aggregation Groups (LAG’s).
A common management architecture for servers, storage and networking could replace the current requirement for individual management platforms for each of them.
VM’s (Virtual Machines) could be moved from highly utilized servers to under-utilized servers dynamically, along with their network state/storage attributes that move along with the VM’s. Virtualization could also be automated using a common set of standard tools.
Modern data center networks should be able to be scale dynamically (even with live traffic) using a set of modular hardware and modular software components, while still maintaining low-cost per port.
A network is generally over-subscribed in order to operate/expand in a cost-effective way. But there are certain types of traffic in the modern data centers (like storage – FCOE traffic) that cannot tolerate packet losses due to over-subscribed networks. Hence, it is important to introduce some kind of throttling mechanism in modern data center networks to pause traffic from the source, in case of network congestion.
Network cables can be connected once and dynamically reconfigured using SDN – Software Defined Networking.
This post represents the highlights of the document, ‘ODIN Volume 1: Transforming the Data Center Network’, by IBM’s ODIN (Open Data Center Inter-operable Network) initiative. You can read the complete discussion on this topic by downloading the document from the above link. This is a part of the ODIN series of posts.
You could stay up to date on the various computer networking/IT technologies by subscribing to this blog with your email address in the sidebar box that says,’Get email updates when new articles are published’