The data center is home to the computational power, storage, and applications necessary to support an
enterprise business. The data center infrastructure is central to the IT architecture, from which all content
is sourced or passes through. Proper planning of the data center infrastructure design is critical, and
performance, resiliency, and scalability need to be carefully considered.
Another important aspect of the data center design is flexibility in quickly deploying and supporting new
services. Designing a flexible architecture that has the ability to support new applications in a short time
frame can result in a significant competitive advantage. Such a design requires solid initial planning and
thoughtful consideration in the areas of port density, access layer uplink bandwidth, true server capacity,
and oversubscription, to name just a few.
The data center network design is based on a proven layered approach, which has been tested and
improved over the past several years in some of the largest data center implementations in the world. The
layered approach is the basic foundation of the data center design that seeks to improve scalability,
performance, flexibility, resiliency, and maintenance.
Basic Layered Design
The layers of the data center design are the core, aggregation, and access layers. These layers are
referred to extensively throughout this guide and are briefly described as follows:
• Core layer—Provides the high-speed packet switching backplane for all flows going in and out of
the data center. The core layer provides connectivity to multiple aggregation modules and provides
a resilient Layer 3 routed fabric with no single point of failure. The core layer runs an interior
routing protocol, such as OSPF or EIGRP, and load balances traffic between the campus core and
aggregation layers using Cisco Express Forwarding-based hashing algorithms.
• Aggregation layer modules—Provide important functions, such as service module integration,
Layer 2 domain definitions, spanning tree processing, and default gateway redundancy.
Server-to-server multi-tier traffic flows through the aggregation layer and can use services, such as
firewall and server load balancing, to optimize and secure applications. The smaller icons within the
aggregation layer switch in represent the integrated service modules. These modules
provide services, such as content switching, firewall, SSL offload, intrusion detection, network
analysis, and more.
• Access layer—Where the servers physically attach to the network. The server components consist
of 1RU servers, blade servers with integral switches, blade servers with pass-through cabling,
clustered servers, and mainframes with OSA adapters. The access layer network infrastructure consists
of modular switches, fixed configuration 1 or 2RU switches, and integral blade server switches. Switches
provide both Layer 2 and Layer 3 topologies, fulfilling the various server broadcast domain or administrative requirements.
This chapter defines the framework on which the recommended data center architecture is based and introduces the primary data center design models: the multi-tier and server cluster models.
Data Center Design Models
The multi-tier model is the most common design in the enterprise. It is based on the web, application,
and database layered design supporting commerce and enterprise business ERP and CRM solutions. This
type of design supports many web service architectures, such as those based on Microsoft .NET or Java
2 Enterprise Edition. These web service application environments are used by ERP and CRM solutions
from Siebel and Oracle, to name a few. The multi-tier model relies on security and application
optimization services to be provided in the network.
The server cluster model has grown out of the university and scientific community to emerge across
enterprise business verticals including financial, manufacturing, and entertainment. The server cluster
model is most commonly associated with high-performance computing (HPC), parallel computing, and
high-throughput computing (HTC) environments, but can also be associated with grid/utility computing.
These designs are typically based on customized, and sometimes proprietary, application architectures
that are built to serve particular business objectives.
“Data Center Multi-Tier Model Design,” provides an overview of the multi-tier model, and “Server Cluster Designs with Ethernet,” provides an overview of the server cluster model.
Later chapters of this guide address the design aspects of these models in greater detail.
Multi-Tier Model
The multi-tier data center model is dominated by HTTP-based applications in a multi-tier approach. The
multi-tier approach includes web, application, and database tiers of servers. Today, most web-based
applications are built as multi-tier applications. The multi-tier model uses software that runs as separate
processes on the same machine using interprocess communication (IPC), or on different machines with
communications over the network. Typically, the following three tiers are used:
• Web- server
• Application
• Database
Multi-tier server farms built with processes running on separate machines can provide improved
resiliency and security. Resiliency is improved because a server can be taken out of service while the
same function is still provided by another server belonging to the same application tier. Security is
improved because an attacker can compromise a web server without gaining access to the application or
database servers. Web and application servers can coexist on a common physical server; the database
typically remains separate.
Resiliency is achieved by load balancing the network traffic between the tiers, and security is achieved
by placing firewalls between the tiers. You can achieve segregation between the tiers by deploying a
separate infrastructure composed of aggregation and access switches, or by using VLANs.
Physical Segregation in a Server Farm with Appliances (A) and Service Modules (B)
VLANs to segregate the server farms. The left side of the illustration (A) shows the physical topology, and the right side (B) shows the VLAN allocation across the service modules, firewall, load balancer, and switch. The firewall and load balancer, which are VLAN-aware, enforce the VLAN segregation between the server farms. Note that not all of the VLANs require load balancing. For example, the database in the example sends traffic directly to the firewall.
Logical Segregation in a Server Farm with VLANs
Physical segregation improves performance because each tier of servers is connected to dedicated
hardware. The advantage of using logical segregation with VLANs is the reduced complexity of the
server farm. The choice of physical segregation or logical segregation depends on your specific network
performance requirements and traffic patterns.
Business security and performance requirements can influence the security design and mechanisms
used. For example, the use of wire-speed ACLs might be preferred over the use of physical firewalls.
Non-intrusive security devices that provide detection and correlation, such as the Cisco Monitoring,
Analysis, and Response System (MARS) combined with Route Triggered Black Holes (RTBH) and
Cisco Intrusion Protection System (IPS) might meet security requirements. Cisco Guard can also be
deployed as a primary defense against distributed denial of service (DDoS) attacks.
Server Cluster Model
In the modern data center environment, clusters of servers are used for many purposes, including high
availability, load balancing, and increased computational power. This guide focuses on the high
performance form of clusters, which includes many forms. All clusters have the common goal of combining
multiple CPUs to appear as a unified high performance system using special software and high-speed
network interconnects. Server clusters have historically been associated with university research,
scientific laboratories, and military research for unique applications, such as the following:
• Meteorology (weather simulation)
• Seismology (seismic analysis)
• Military research (weapons, warfare
Server clusters are now in the enterprise because the benefits of clustering technology are now being
applied to a broader range of applications. The following applications in the enterprise are driving this
requirement:
• Financial trending analysis—Real-time bond price analysis and historical trending
• Film animation—Rendering of artist multi-gigabyte files
• Manufacturing—Automotive design modeling and aerodynamics
• Search engines—Quick parallel lookup plus content insertion
In the enterprise, developers are increasingly requesting higher bandwidth and lower latency for a
growing number of applications. The time-to-market implications related to these applications can result
in a tremendous competitive advantage. For example, the cluster performance can directly affect getting
a film to market for the holiday season or providing financial management customers with historical
trending information during a market shift.
HPC Cluster Types and Interconnects.
In the high performance computing landscape, various HPC cluster types exist and various interconnect
technologies are used. The top 500 supercomputer list at www.top500.org provides a fairly
comprehensive view of this landscape. The majority of interconnect technologies used today are based
on Fast Ethernet and Gigabit Ethernet, but a growing number of specialty interconnects exist, for
example including Infiniband and Myrinet. Specialty interconnects such as Infiniband have very low
latency and high bandwidth switching characteristics when compared to traditional Ethernet, and
leverage built-in support for Remote Direct Memory Access (RDMA). 10GE NICs have also recently
emerged that introduce TCP/IP offload engines that provide similar performance to Infiniband.
The Cisco SFS line of Infiniband switches and Host Channel Adapters (HCAs) provide high
performance computing solutions that meet the highest demands.
Although high performance clusters (HPCs) come in various types and sizes, the following categorizes
three main types that exist in the enterprise environment:
• HPC type 1—Parallel message passing (also known as tightly coupled)
– Applications run on all compute nodes simultaneously in parallel.
– A master node determines input processing for each compute node.
– Can be a large or small cluster, broken down into hives (for example, 1000 servers over 20 hives)
with IPC communication between compute nodes/hives.
• HPC type 2—Distributed I/O processing (for example, search engines)
– The client request is balanced across master nodes, then sprayed to compute nodes for parallel
processing (typically unicast at present, with a move towards multicast).
– This type obtains the quickest response, applies content insertion (advertising), and sends to the
client.
• HPC Type 3—Parallel file processing (also known as loosely coupled)
– The source data file is divided up and distributed across the compute pool for manipulation in
parallel. Processed components are rejoined after completion and written to storage.
– Middleware controls the job management process (for example, platform linear file system
[LFS]).
The traditional high performance computing cluster that emerged out of the university and military
environments was based on the type 1 cluster. The new enterprise HPC applications are more aligned
with HPC types 2 and 3, supporting the entertainment, financial, and a growing number of other vertical
industries.
Logical View of a Server Cluster
The components of the server cluster are as follows:
• Front end—These interfaces are used for external access to the cluster, which can be accessed by
application servers or users that are submitting jobs or retrieving job results from the cluster. An
example is an artist who is submitting a file for rendering or retrieving an already rendered result.
This is typically an Ethernet IP interface connected into the access layer of the existing server farm
infrastructure.
• Master nodes (also known as head node)—The master nodes are responsible for managing the
compute nodes in the cluster and optimizing the overall compute capacity. Usually, the master node
is the only node that communicates with the outside world. Clustering middleware running on the
master nodes provides the tools for resource management, job scheduling, and node state monitoring
of the computer nodes in the cluster. Master nodes are typically deployed in a redundant fashion and
are usually a higher performing server than the compute nodes.
• Back-end high-speed fabric—This high-speed fabric is the primary medium for master node to
compute node and inter-compute node communications. Typical requirements include low latency
and high bandwidth and can also include jumbo frame and 10 GigE support. Gigabit Ethernet is the
most popular fabric technology in use today for server cluster implementations, but other
technologies show promise, particularly Infiniband.
• Compute nodes—The compute node runs an optimized or full OS kernel and is primarily
responsible for CPU-intense operations such as number crunching, rendering, compiling, or other
file manipulation.
• Storage path—The storage path can use Ethernet or Fibre Channel interfaces. Fibre Channel
interfaces consist of 1/2/4G interfaces and usually connect into a SAN switch such as a Cisco MDS
platform. The back-end high-speed fabric and storage path can also be a common transport medium
when IP over Ethernet is used to access storage. Typically, this is for NFS or iSCSI protocols to a
NAS or SAN gateway, such as the IPS module on a Cisco MDS platform.
• Common file system—The server cluster uses a common parallel file system that allows high
performance access to all compute nodes. The file system types vary by operating system (for
example, PVFS or Lustre).
Physical Overview
Server cluster designs can vary significantly from one to another, but certain items are common, such as
the following:
• Commodity off the Shelf (CotS) server hardware—The majority of server cluster implementations are
based on 1RU Intel- or AMD-based servers with single/dual processors. The spiraling cost of these
high performing 32/64-bit low density servers has contributed to the recent enterprise adoption of
cluster technology.
• GigE or 10 GigE NIC cards—The applications in a server cluster can be bandwidth intensive and
have the capability to burst at a high rate when necessary. The PCI-X or PCI-Express NIC cards
provide a high-speed transfer bus speed and use large amounts of memory. TCP/IP offload and
RDMA technologies are also used to increase performance while reducing CPU utilization.
• Low latency hardware—Usually a primary concern of developers is related to the message-passing
interface delay affecting the overall cluster/application performance. This is not always the case
because some clusters are more focused on high throughput, and latency does not significantly
impact the applications. The Cisco Catalyst 6500 with distributed forwarding and the Catalyst
4948-10G provide consistent latency values necessary for server cluster environments.
• Non-blocking or low-over-subscribed switch fabric—Many HPC applications are
bandwidth-intensive with large quantities of data transfer and interprocess communications between
compute nodes. GE attached server oversubscription ratios of 2.5:1 (500 Mbps) up to 8:1(125 Mbps) are
common in large server cluster designs.
• Mesh/partial mesh connectivity—Server cluster designs usually require a mesh or partial mesh
fabric to permit communication between all nodes in the cluster. This mesh fabric is used to share
state, data, and other information between master-to-compute and compute-to-compute servers in
the cluster.
• Jumbo frame support—Many HPC applications use large frame sizes that exceed the 1500 byte
Ethernet standard. The ability to send large frames (called jumbos) that are up to 9K in size, provides
advantages in the areas of server CPU overhead, transmission overhead, and file transfer time.
Physical View of a Server Cluster Model Using ECMP
The recommended server cluster design leverages the following technical aspects or features:
• Equal cost multi-path—ECMP support for IP permits a highly effective load distribution of traffic
across multiple uplinks between servers across the access layer. Although Figure 1-6 demonstrates
a four-way ECMP design, this can scale to eight-way by adding additional paths.
• Distributed forwarding—By using distributed forwarding cards on interface modules, the design
takes advantage of improved switching performance and lower latency.
• L3 plus L4 hashing algorithms—Distributed Cisco Express Forwarding-based load balancing
permits ECMP hashing algorithms based on Layer 3 IP source-destination plus Layer 4
source-destination port, allowing a highly granular level of load distribution.
• Scalable server density—The ability to add access layer switches in a modular fashion permits a
cluster to start out small and easily increase as required.
• Scalable fabric bandwidth—ECMP permits additional links to be added between the core and access
layer as required, providing a flexible method of adjusting oversubscription and bandwidth per
server.
In the preceding design, master nodes are distributed across multiple access layer switches to provide
redundancy as well as to distribute load.
