Data Center Topologies
This section discusses Data Center topologies and, in particular, the server farm topology. Initially, the discussion focuses on the traffic flow through the network infrastructure (on a generic topology) from a logical viewpoint and then from a physical viewpoint.
Generic Layer 3/Layer 2 Designs
The generic Layer 3/Layer 2 designs are based on the most common ways of deploying server farms. Figure 4-6 depicts a generic server farm topology that supports a number of servers.
NOTE
Notice that the distribution layer now is referred to as the aggregation layer resulting from becoming the aggregation point for most, if not all, services beyond the traditional Layer 2 and Layer 3.
)
Figure
4-6 Generic Server Farm Design
The highlights of the topology are the aggregation-layer switches that perform key Layer 3 and Layer 2 functions, the access-layer switches that provide connectivity to the servers in the server farm, and the connectivity between the aggregation and access layer switches.
The key Layer 3 functions performed by the aggregation switches are as follows:
Forwarding packets based on Layer 3 information between the server farm and the rest of the network
Maintaining a "view" of the routed network that is expected to change dynamically as network changes take place
Supporting default gateways for the server farms
The key Layer 2 functions performed by the aggregation switches are as follows:
Spanning Tree Protocol (STP) 802.1d between aggregation and access switches to build a loop-free forwarding topology.
STP enhancements beyond 802.1d that improve the default spanning-tree behavior, such as 802.1s, 802.1w, Uplinkfast, Backbonefast, and Loopguard. For more information, refer to Chapter 12, "Layer 2 Protocol Essentials."
VLANs for logical separation of server farms.
Other services, such as multicast and ACLs for services such as QoS, security, rate limiting, broadcast suppression, and so on.
The access-layer switches provide direct connectivity to the server farm. The types of servers in the server farm include generic servers such as DNS, DHCP, FTP, and Telnet; mainframes using SNA over IP or IP; and database servers. Notice that some servers have both internal disks (storage) and tape units, and others have the storage externally connected (typically SCSI).
The connectivity between the two aggregation switches and between aggregation and access switches is as follows:
EtherChannel between aggregation switches. The channel is in trunk mode, which allows the physical links to support as many VLANs as needed (limited to 4096 VLANs resulting from the 12-bit VLAN ID).
Single or multiple links (EtherChannel, depending on how much oversubscription is expected in the links) from each access switch to each aggregation switch (uplinks). These links are also trunks, thus allowing multiple VLANs through a single physical path.
Servers dual-homed to different access switches for redundancy. The NIC used by the server is presumed to have two ports in an active-standby configuration. When the primary port fails, the standby takes over, utilizing the same MAC and IP addresses that the active port was using. For more information about dual-homed servers, refer to Chapter 2, "Server Architecture Overview."
The typical configuration for the server farm environment just described is presented in Figure 4-7.
Figure 4-7 shows the location for the critical services required by the server farm. These services are explicitly configured as follows:
agg1 is explicitly configured as the STP root.
agg2 is explicitly configured as the secondary root.
agg1 is explicitly configured as the primary default gateway.
agg2 is explicitly configured as the standby or secondary default gateway.
)
Figure
4-7 Common Server Farm Environment
NOTE
The explicit definition of these critical functions sets the primary and alternate paths to and from the server farm. Notice that there is no single point of failure in the architecture, and the paths are now deterministic.
Other STP services or protocols, such as UplinkFast, are also explicitly defined between the aggregation and access layers. These services/protocols are used to lower convergence time during failover conditions from the 802.d standard of roughly 50 seconds to 1 to 3 seconds.
In this topology, the servers are configured to use the agg1 switch as the primary default gateway, which means that outbound traffic from the servers follows the direct path to the agg1 switch. Inbound traffic can arrive at either aggregation switch, yet the traffic can reach the server farm only through agg1 because the links from agg2 to the access switches are not forwarding (blocking). The inbound paths are represented by the dotted arrows, and the outbound path is represented by the solid arrow.
The next step is to have predictable failover and fallback behavior, which is much simpler when you have deterministic primary and alternate paths. This is achieved by failing every component in the primary path and recording and tuning the failover time to the backup component until the requirements are satisfied. The same process must be done for falling back to the original primary device. This is because the failover and fallback processes are not the same. In certain instances, the fallback can be done manually instead of automatically, to prevent certain undesirable conditions.
NOTE
When using 802.1d. if the primary STP root fails and the secondary takes over, when it comes back up, it automatically takes over because it has a lower priority. In an active server farm environment, you might not want to have the STP topology change automatically, particularly when the convergence time is in the range of 50 seconds. However, this behavior is not applicable when using 802.1w, in which the fallback process takes only a few seconds.
Whether using 802.1d or 802.1w, the process is automatic, unlike when using HSRP, in which the user can control the behavior of the primary HSRP peer when it becomes operational again through the use of preemption. If preemption is not used, the user has manual control over when to return mastership to the initial master HSRP peer.
The use of STP is the result of a Layer 2 topology, which might have loops that require an automatic mechanism to be detected and avoided. An important question is whether there is a need for Layer 2 in a server farm environment. This topic is discussed in the following section.
For more information about the details of the Layer 2 design, see Chapter 20, "Designing the Data Center Infrastructure."
The Need for Layer 2 at the Access Layer
Access switches traditionally have been Layer 2 switches. This holds true also for the campus network wiring closet. This discussion is focused strictly on the Data Center because it has distinct and specific requirements, some similar to and some different than those for the wiring closets.
The reason access switches in the Data Center traditionally have been Layer 2 is the result of the following requirements:
When they share specific properties, servers typically are grouped on the same VLAN. These properties could be as simple as ownership by the same department or performance of the same function (file and print services, FTP, and so on). Some servers that perform the same function might need to communicate with one another, whether as a result of a clustering protocol or simply as part of the application function. This communication exchange should be on the same subnet and sometimes is possible only on the same subnet if the clustering protocol heartbeats or the server-to-server application packets are not routable.
Servers are typically dual-homed so that each leg connects to a different access switch for redundancy. If the adapter in use has a standby interface that uses the same MAC and IP addresses after a failure, the active and standby interfaces must be on the same VLAN (same default gateway).
Server farm growth occurs horizontally, which means that new servers are added to the same VLANs or IP subnets where other servers that perform the same functions are located. If the Layer 2 switches hosting the servers run out of ports, the same VLANs or subnets must be supported on a new set of Layer 2 switches. This allows flexibility in growth and prevents having to connect two access switches.
When using stateful devices that provide services to the server farms, such as load balancers and firewalls, these stateful devices expect to see both the inbound and outbound traffic use the same path. They also need to constantly exchange connection and session state information, which requires Layer 2 adjacency. More details on these requirements are discussed in the section, "Access Layer," which is under the section, "Multiple Tier Designs."
Using just Layer 3 at the access layer would prevent dual-homing, Layer 2 adjacency between servers on different access switches, and Layer 2 adjacency between service devices. Yet if these requirements are not common on your server farm, you could consider a Layer 3 environment in the access layer. Before you decide what is best, it is important that you read the section titled "Fully Redundant Layer 2 and Layer 3 Designs with Services," later in the chapter. New service trends impose a new set of requirements in the architecture that must be considered before deciding which strategy works best for your Data Center.
The reasons for migrating away from a Layer 2 access switch design are motivated by the need to drift away from spanning tree because of the slow convergence time and the operation challenges of running a controlled loopless topology and troubleshooting loops when they occur. Although this is true when using 802.1d, environments that take advantage of 802.1w combined with Loopguard have the following characteristics: They do not suffer from the same problems, they are as stable as Layer 3 environments, and they support low convergence times.
NOTE
The STP standard 802.1d has limitations in addressing certain conditions in addition to its convergence time, yet a fair amount of spanning tree–related problems are the result of misconfiguration or rogue STP devices that appear on the network and "bridge" between Layer 2 domains. More information on this topic is presented in Chapter 12.
The next section discusses an alternate solution for a topology with spanning tree that does not present the STP problems or limitations.
Alternate Layer 3/Layer 2 Designs
Figure 4-8 presents an alternate Layer 3/Layer 2 design resulting from the need to address STP limitations.
)
Figure
4-8 Loopless Topology
Figure 4-8 presents a topology in which the network purposely is designed not to have loops. Although STP is running, its limitations do not present a problem. This loopless topology is accomplished by removing or not allowing the VLAN(s), used at the access-layer switches, through the trunk between the two aggregation switches. This basically prevents a loop in the topology while it supports the requirements behind the need for Layer 2.
In this topology, the servers are configured to use the agg1 switch as the primary default gateway. This means that outbound traffic from the servers connected to acc2 traverses the link between the two access switches. Inbound traffic can use either aggregation switch because both have active (nonblocking) paths to the access switches. The inbound paths are represented by the dotted arrows, and the outbound path is represented by the solid arrows.
This topology is not without its own challenges. These challenges are discussed later in the chapter after other information related to the deployment of services becomes available.
Multiple-Tier Designs
Most applications conform to either the client/server model or the n-tier model, which implies most networks, and server farms support these application environments. The tiers supported by the Data Center infrastructure are driven by the specific applications and could be any combination in the spectrum of applications from the client/server to the client/web server/application server/database server. When you identify the communication requirements between tiers, you can determine the needed specific network services. The communication requirements between tiers are typically higher scalability, performance, and security. These could translate to load balancing between tiers for scalability and performance, or SSL between tiers for encrypted transactions, or simply firewalling and intrusion detection between the web and application tier for more security.
Figure 4-9 introduces a topology that helps illustrate the previous discussion.
Notice that Figure 4-9 is a logical diagram that depicts layer-to-layer connectivity through the network infrastructure. This implies that the actual physical topology might be different. The separation between layers simply shows that the different server functions could be physically separated. The physical separation could be a design preference or the result of specific requirements that address communication between tiers.
For example, when dealing with web servers, the most common problem is scaling the web tier to serve many concurrent users. This translates into deploying more web servers that have similar characteristics and the same content so that user requests can be equally fulfilled by any of them. This, in turn, requires the use of a load balancer in front of the server farm that hides the number of servers and virtualizes their services. To the users, the specific service is still supported on a single server, yet the load balancer dynamically picks a server to fulfill the request.
)
Figure
4-9 Multiple-Tier Application Environments
Suppose that you have multiple types of web servers supporting different applications, and some of these applications follow the n-tier model. The server farm could be partitioned along the lines of applications or functions. All web servers, regardless of the application(s) they support, could be part of the same server farm on the same subnet, and the application servers could be part of a separate server farm on a different subnet and different VLAN. Following the same logic used to scale the web tier, a load balancer logically could be placed between the web tier and the application tier to scale the application tier from the web tier perspective. A single web server now has multiple application servers to access.
The same set of arguments holds true for the need for security at the web tier and a separate set of security considerations at the application tier. This implies that firewall and intrusion-detection capabilities are distinct at each layer and, therefore, are customized for the requirements of the application and the database tiers. SSL offloading is another example of a function that the server farm infrastructure might support and can be deployed at the web tier, the application tier, and the database tier. However, its use depends upon the application environment using SSL to encrypt client-to-server and server-to-server traffic.
Expanded Multitier Design
The previous discussion leads to the concept of deploying multiple network-based services in the architecture. These services are introduced in Figure 4-10 through the use of icons that depict the function or service performed by the network device.
NOTE
Figure 4-10 introduces the icons used through this chapter to depict the services provided by network devices in the Data Center.
The different icons are placed in front of the servers for which they perform the functions. At the aggregation layer, you find the load balancer, firewall, SSL offloader, intrusion-detection system, and cache. These services are available through service modules (line cards that could be inserted into the aggregation switch) or appliances. An important point to consider when dealing with service devices is that they provide scalability and high availability beyond the capacity of the server farm, and that to maintain the basic premise of "no single point of failure," at least two must be deployed. If you have more than one (and considering you are dealing with redundancy of application environments), the failover and fallback processes require special mechanisms to recover the connection context, in addition to the Layer 2 and Layer 3 paths. This simple concept of redundancy at the application layer has profound implications in the network design.
)
Figure
4-10 Network Service Icons
A number of these network service devices are replicated in front of the application layer to provide services to the application servers. Notice in Figure 4-10 that there is physical separation between the tiers of servers. This separation is one alternative to the server farm design. Physical separation is used to achieve greater control over the deployment and scalability of services. The expanded design is more costly because it uses more devices, yet it allows for more control and better scalability because the devices in the path handle only a portion of the traffic. For example, placing a firewall between tiers is regarded as a more secure approach because of the physical separation between the Layer 2 switches. This argument is correct, yet it is likely to be much more related to an existing security policy than a real threat. Having logical instead of physical separation simply requires a consistent application of security policies to ensure that the expanded security zone is as secure logically as it is physically.
This brings the discussion to another alternative of designing the multitier server farm, an alternative in which there is no physical separation, but rather a logical separation between tiers, as presented in the next section.
Collapsed Multitier Design
A collapsed multitier design is one in which all the server farms are directly connected at the access layer to the aggregation switches, and there is no physical separation between the Layer 2 switches that support the different tiers. Figure 4-11 presents the collapsed design.
)
Figure
4-11 Collapsed Multiple-Tier Design
Notice that in this design, the services again are concentrated at the aggregation layer, and the service devices now are used by the front-end tier and between tiers. Using a collapsed model, there is no need to have a set of load balancers or SSL offloaders dedicated to a particular tier. This reduces cost, yet the management of devices is more challenging and the performance demands are higher. The service devices, such as the firewalls, protect all server tiers from outside the Data Center, but also from each other. The load balancer also can be used concurrently to load-balance traffic from client to web servers, and traffic from web servers to application servers.
Notice that the design in Figure 4-11 shows each type of server farm on a different set of switches. Other collapsed designs might combine the same physical Layer 2 switches to house web applications and database servers concurrently. This implies merely that the servers logically are located on different IP subnets and VLANs, yet the service devices still are used concurrently for the front end and between tiers. Notice that the service devices are always in pairs. Pairing avoids the single point of failure throughout the architecture. However, both service devices in the pair communicate with each other, which falls into the discussion of whether you need Layer 2 or Layer 3 at the access layer.
The Need for Layer 2 at the Access Layer
Each pair of service devices must maintain state information about the connections the pair is handling. This requires a mechanism to determine the active device (master) and another mechanism to exchange connection state information on a regular basis. The goal of the dual–service device configuration is to ensure that, upon failure, the redundant device not only can continue service without interruption, but also seamlessly can failover without disrupting the current established connections.
In addition to the requirements brought up earlier about the need for Layer 2, this section discusses in depth the set of requirements related to the service devices:
Service devices and the server farms that they serve are typically Layer 2–adjacent. This means that the service device has a leg sitting on the same subnet and VLAN used by the servers, which is used to communicate directly with them. Often, in fact, the service devices themselves provide default gateway support for the server farm.
Service devices must exchange heartbeats as part of their redundancy protocol. The heartbeat packets might or might not be routable; if they are routable, you might not want the exchange to go through unnecessary Layer 3 hops.
Service devices operating in stateful failover need to exchange connection and session state information. For the most part, this exchange is done over a VLAN common to the two devices. Much like the heartbeat packets, they might or might not be routable.
If the service devices provide default gateway support for the server farm, they must be adjacent to the servers.
After considering all the requirements for Layer 2 at the access layer, it is important to note that although it is possible to have topologies such as the one presented in Figure 4-8, which supports Layer 2 in the access layer, the topology depicted in Figure 4-7 is preferred. Topologies with loops are also supportable if they take advantages of protocols such as 802.1w and features such as Loopguard.
NOTE
To date, most common implementations use Layer 2 at the access layer and rely on the Spanning Tree Protocols and Cisco enhancements to lower convergence times and achieve stability, as depicted in Figure 4-7. Few use the loopless topology. The main reasons relate to whether it is possible to have a loopless topology, given the restrictions imposed by the requirements, and, if possible, whether the setup is simple enough for support, maintenance, and management reasons. Dual-homing requires Layer 2 adjacency between access switches to carry the same VLANs, and redundant stateful service devices need Layer 2 adjacency to work properly. Therefore, it is important to carefully consider the requirements when designing the server farm network infrastructure.
The following section discusses topics related to the topology of the server farms.