Horizontal virtualization cluster of data center Switches

Horizontal virtualization cluster of data center Switches

Virtualization Technology is a buzzword in the data center, including horizontal virtualization, vertical virtualization, one virtual machine and multiple virtualization, and NVO3 virtualization. Today, we will focus on horizontal virtualization. Taking Huawei CloudEngine 12800 series as an example, let our friends know the origins and development history of this technology, this article briefly introduces the features of various horizontal Virtualization Technologies and selection strategies in various scenarios.

Origins of horizontal virtualization Clusters

In the early stages of data center network development, there was no dedicated data center switch. What should I do? First, take the campus switch and use the most traditional VRRP + STP. Let's use it together. This is the classic campus network below.

This network model is classic and reliable. After a long time, the problem arises:

◆ The traffic increases and STP blocking leads to low Link Utilization;

◆ Non-Shortest Path forwarding. The root node has a bandwidth bottleneck and the forwarding latency is high;

◆ VRRP single-active gateway, with the standby node equipment idle;

◆ Limited STP network scale, poor convergence performance;

◆ Multiple management nodes, complicated logic topology, and difficult maintenance.

These problems have brought about the demand for horizontal virtualization, and the box switch cluster took the lead.

Stack

A typical box Switch stack includes CISCO's VSS (Virtual Switch System), Huawei's CSS (Cluster Switch System), and H3C's IRF (Intelligent Resilient Framework ). VSS, CSS, and IRF are essentially stacked, but they only wear different vests. Of course, manufacturers also develop some differences.

Stack technology is essentially a merge of management plane, control plane, and forwarding plane. Stack the Main Control Board of the system and manage all the line cards and network boards of the two physical devices to form a logical large switch.

However, it should be noted that the purpose of stacking is not only to increase, but also to look at the logical topology from the network perspective and become "rich and handsome "!

The main performances of "Gao fushuai:

◆ Super nodes with almost twice the exchange capability;

◆ Layer-2 and layer-3 forwarding traffic are fully load-balanced, making full use of all links;

◆ Single-node logic, comprehensive business support, and simple network solution design;

◆ Supports Fault Protection for physical nodes by deploying cross-frame link-failure;

◆ The integration of network elements facilitates network management and maintenance.

There are also a number of benefits:

◆ Shortest Path forwarding with low latency;

◆ Compared with traditional STP, a larger L2 network can be established;

◆ Convergence performance of link-failure and network fault convergence block.

In a stack system, the bandwidth of the stack link is always insufficient relative to the Service port. This requires that the forwarded business traffic should be avoided through stacked links as much as possible. This is the so-called local traffic priority forwarding.

As shown in, the Huawei data center switch stack system provides local priority for layer-3 ECMP and link bundling. Local preferential forwarding saves the bandwidth of the stack link and reduces the forwarding latency.

In addition to the above general stack technology, the Huawei CloudEngine 12800 series data center high-end switches have also made significant physical improvements to the reliability of stack.

Stack Optimization

Reliability Optimization (stack of transfer control separation)

The stacking of transfer control separation, also known as out-of-band stacking, is mainly aimed at high reliability.

Most of the industry's box switches are stacked. The control channels and forwarding channels between stack members use one channel. The CloudEngine 12800 Series Data Center Switches of Huawei have independently developed a stack system for forwarding control separation. Here, "transfer" refers to the business data forwarding channel; "control" refers to the Control Message (also known as "signaling") channel.

In traditional frame-based stack systems, business data channels and Control Message channels use the same physical channel, that is, stacked links. As shown in:

In this stack system, control messages and data are run together. If the data communication volume of the stack channel is large, the control messages may be impacted and lost, thus affecting the reliability of the control plane. Strictly speaking, this design does not meet the design requirements of "data, control, and management Plane Separation. In addition, the establishment of the stack system depends on the startup of the Line Card, which leads to an increase in software complexity and affects the startup speed of the stack.

The stack system of the transfer control separation adopts the following Architecture:

The hardware stack architecture brings about a series of reliability improvements:

◆ Control the physical isolation between message channels and business data channels to ensure that business data does not affect control messages;

◆ Threefold dual-master fault protection, including stack management link (4-way), stack forwarding Link (at least 2-way), Business port/Management port DAD;

◆ Stack System Establishment, no longer dependent on the start of the line card, no Software Timing dependency, simplified software implementation, and simple means reliable;

◆ Stack System Establishment, no longer waiting for the start of line cards/network boards, shortening the stack system establishment time;

◆ Short Message channel control path, less fault points, low latency.

Limitations of stack Improvement

The stack system brings the advantages of the above series, but the unpleasant problems are gradually exposed, which is determined by the nature of the stack principle.

As shown in, the two switches form a logical switch by tightly coupled management plane, control plane, and data plane. This leads to the following three risks or problems.

◆ Overall system-level reliability risks

For common faults, the stack system can implement fault protection through link switching, Master/Slave board switching, and frame switching. However, because the two physical Switches of the entire system are tightly coupled in the software (management plane and control plane), this increases the possibility of software faults spreading from one switch to another. Once such a fault occurs, the failure of the entire stack system will affect all services connected to the stack system.

◆ Service Interruption Duration of Version Upgrade

Because the stack itself provides business protection, when the stack system is upgraded, traffic protection cannot be performed by another node when the VRRP member node is upgraded, and the interruption time is relatively long.

In this regard, various manufacturers have developed two-frame RoundRobin and ISSU upgrade methods, which shorten the service interruption time during the upgrade, but do not solve the upgrade risks described below, the upgrade risks are even magnified due to the increase in technical complexity and software engineering complexity.

◆ Overall system upgrade risks

The Software Version Upgrade of a device is a risky network operation even if the most traditional and simple upgrade method is used. If the device fails to be upgraded, the Service carried by the device will become invalid. In this case, you must use all means including rollback to restore the service as soon as possible.

Due to the tight coupling between member switches, the stack system can only be upgraded by two devices. An upgrade failure will interrupt all business networks in the stack system. Stack systems often assume the role of Dual-protection access on servers at the access layer, or the role of High-reliability gateway at aggregation. This means that the upgrade failure may lead to paralysis of the entire business.

Link-virtualization Virtualization (M-LAG)

Horizontal virtualization is required to meet the requirements of Layer 2 cross-device redundancy at the access layer and aggregation layer, and cross-device redundancy at the L3 gateway at the aggregation layer. Are there other technologies that support horizontal virtualization without stacking?

The answer is of course, The M-LAG (Multichassis Link Aggregation Group) of Huawei CloudEngine Series Data Center switches supports such virtualization technology. This technology implements layer-2 virtualization only on the link-plane of two devices. The management and control planes of the two member devices are independent.

Note: Wikipedia calls this technology a MC-LAG (Multi-Chassis Link Aggregation Group), CISCO calls it a Virtual Port-Channel ). In this article, Wikipedia is used, which is abbreviated as MC-LAG.

MC-LAG, supporting cross-device link bundling, supporting Dual-Active L3GW. On the access side, from the peer device perspective, the server perspective, the MC-LAG is similar to the stack.

However, from the perspective of L3 network, the two member nodes of the MC-LAG have their own independent IP addresses, the two nodes have their own independent management and control plane. From an architecture perspective, the two member devices of the MC-LAG only have the coupling of the data plane and the lightweight coupling of the protocol plane:

The architecture of MC-LAG determines that this technical solution does not have three problems that stack is difficult to solve:

So, said the MC-LAG so many benefits, is there no disadvantages? Of course not. You have an inch of strength and a short size. The last section compares the advantages and disadvantages of stack and MC-LAG, as well as the scenario selection recommendations.

Stack and MC-LAG comparison and selection recommendations

Stack and MC-LAG have their own advantages and disadvantages, according to the above comparison table. In general, for Network Design/maintenance personnel, stack wins in the management and maintenance of simple, MC-LAG wins in the risk of reliability and low upgrade.

When designing a data center network solution, consider the following considerations:

◆ Strategy 1: convergence layer priority reliability, upgrade convenience, choose M-LAG; access layer because of the large amount of equipment, priority to business deployment and maintenance convenience, choose stack.

◆ Policy 2: priority is given to reliability and low upgrade risks. M-LAG is used for aggregation and access.

◆ Policy 3: give priority to ease of service deployment and maintenance, and stack is used for aggregation and access.