Compaq ProLiant CL380 ServerNet II SAN Interconnect for Scalable Computing Clu - Page 3

Introduction Traditional Networks Versus System Area Networks - linux

Page 3 highlights

WHITE PAPER (cont.) Doc Number TC000602WP ... INTRODUCTION The enterprise computing systems of many businesses were built around centralized, proprietary computers, which are expensive to manage and upgrade. The networks that evolved around these centralized systems use traditional network technologies such as Ethernet and TCP/IP, which are designed for heterogeneous computing environments. Today, businesses can satisfy their computing goals with a collection of low-cost, industry-standard servers distributed within a cluster. A cluster is a group of two or more interconnected servers that act as a single computing unit. Servers in a cluster can be physically distributed across various locations, yet to a user, the cluster appears as a single, unified computing resource. In a parallel application cluster, for example, each server simultaneously runs a copy of an application and the operating system. Client requests are divided among the servers' CPUs (central processing units), and the servers exchange information about the portion of the client's request they are processing. If one server fails, the jobs it was processing are distributed across the remaining servers. Consequently, clustered servers collectively generate tens of thousands of messages per second to keep track of the jobs they are executing. The high volume of server-to-server messaging in a cluster requires a very efficient software communication interface and a highly reliable, high-speed, low-latency hardware interconnect. Without an efficient server-to-server interconnect, the performance and scalability of a cluster is severely limited. Traditional network technologies produce excessive software overhead, due to their heterogeneous environments, and are too inefficient for intense server-to-server communication. Therefore, in December 1997 Compaq and other industry leaders developed the Virtual Interface (VI) Architecture specification for a distributed messaging interface that allows more efficient server-to-server communication through a system area network (SAN). The VI Architecture provides a common software and hardware interface standard so that customers can choose the optimum SAN interconnect (Compaq ServerNet II, gigabit Ethernet, cLAN, ATM, or others). The Compaq ServerNet II SAN interconnect is the most complete industry-standard implementation of the VI Architecture specification. Measurement of CPU utilization under certain conditions shows that ServerNet II provides three times the performance of gigabit Ethernet and TCP/IP. As a result, ServerNet II is quickly emerging as the interconnect of choice for implementing clusters of industry-standard servers. Operating system vendor and independent software vendor support for ServerNet II includes Windows 2000 Data Center, Windows 2000 Advanced Data Server, Linux, and SCO UnixWare 7.1. TRADITIONAL NETWORKS VERSUS SYSTEM AREA NETWORKS Clustered servers can communicate with each other by using a traditional network or a SAN. Traditional network technologies allow several different types of devices to communicate with each other in complex, heterogeneous networks. Network applications manage communications between the devices using multipoint protocols like TCP/IP and network interface controllers. These protocols generate a large amount of software overhead (error checking and control information) to ensure that messages are sent and received reliably. This software overhead continually interrupts the servers' CPUs. Consequently, the high volume of server-to-server messaging in a cluster can overwhelm traditional network protocols like TCP/IP. In fact, clusters that use traditional networks for messaging can lose as much as 20 to 30 percent of their capability during intense messaging. This loss of system capability results in slower response times for users and reduces the scalability, availability, and flexibility of clusters. When you consider the unpredictable nature of Internet traffic, it becomes apparent that traditional network technologies cannot handle server-to-server messaging in e-commerce environments. 3

  • 1
  • 2
  • 3
  • 4
  • 5
  • 6
  • 7
  • 8
  • 9

W
HITE
P
APER
(cont.)
3
Doc Number
TC000602WP
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
I
NTRODUCTION
The enterprise computing systems of many businesses were built around centralized, proprietary
computers, which are expensive to manage and upgrade.
The networks that evolved around these
centralized systems use traditional network technologies such as Ethernet and TCP/IP, which are
designed for heterogeneous computing environments.
Today, businesses can satisfy their
computing goals with a collection of low-cost, industry-standard servers distributed within a
cluster.
A cluster is a group of two or more interconnected servers that act as a single computing
unit.
Servers in a cluster can be physically distributed across various locations, yet to a user, the
cluster appears as a single, unified computing resource.
In a parallel application cluster, for
example, each server simultaneously runs a copy of an application and the operating system.
Client
requests are divided among the servers’ CPUs (central processing units), and the servers exchange
information about the portion of the client’s request they are processing.
If one server fails, the
jobs it was processing are distributed across the remaining servers.
Consequently, clustered servers
collectively generate tens of thousands of messages per second to keep track of the jobs they are
executing.
The high volume of server-to-server messaging in a cluster requires a very efficient software
communication interface and a highly reliable, high-speed, low-latency hardware interconnect.
Without an efficient server-to-server interconnect, the performance and scalability of a cluster is
severely limited.
Traditional network technologies produce excessive software overhead, due to
their heterogeneous environments, and are too inefficient for intense server-to-server
communication.
Therefore, in December 1997 Compaq and other industry leaders developed the
Virtual Interface (VI) Architecture specification for a distributed messaging interface that allows
more efficient server-to-server communication through a system area network (SAN).
The VI
Architecture provides a common software and hardware interface standard so that customers can
choose the optimum SAN interconnect (Compaq
ServerNet II
, gigabit Ethernet, cLAN, ATM, or
others).
The Compaq
ServerNet II
SAN interconnect is the most complete industry-standard
implementation of the VI Architecture specification.
Measurement of CPU utilization under
certain conditions shows that ServerNet II provides three times the performance of gigabit Ethernet
and TCP/IP.
As a result, ServerNet II is quickly emerging as the interconnect of choice for
implementing clusters of industry-standard servers.
Operating system vendor and independent
software vendor support for ServerNet II includes Windows 2000 Data Center, Windows 2000
Advanced Data Server, Linux, and SCO UnixWare 7.1.
T
RADITIONAL
N
ETWORKS VERSUS
S
YSTEM
A
REA
N
ETWORKS
Clustered servers can communicate with each other by using a traditional network or a SAN.
Traditional network technologies allow several different types of devices to communicate with each
other in complex, heterogeneous networks.
Network applications manage communications
between the devices using multipoint protocols like TCP/IP and network interface controllers.
These protocols generate a large amount of software overhead (error checking and control
information) to ensure that messages are sent and received reliably.
This software overhead
continually interrupts the servers’ CPUs.
Consequently, the high volume of server-to-server
messaging in a cluster can overwhelm traditional network protocols like TCP/IP.
In fact, clusters
that use traditional networks for messaging can lose as much as 20 to 30 percent of their capability
during intense messaging.
This loss of system capability results in slower response times for users
and reduces the scalability, availability, and flexibility of clusters.
When you consider the
unpredictable nature of Internet traffic, it becomes apparent that traditional network technologies
cannot handle server-to-server messaging in e-commerce environments.