Dell 32-bit HPCC Solutions

Dell 32-bit High Performance Computing Cluster Solutions

Scalable Systems Group

Dell Inc.

August 2003

Scalable Systems Group

Table of Contents

1Introduction

1Design Architecture

5Management Solution

6In-band Monitoring and Management

6Dell IT Assistant

7Ganglia

8Out-of-Band Monitoring and Management

9Embedded Remote Access, or ERA

10Digital KVM for the PE1655MC Cluster Configuration

10HPCC Software Package

11Cluster Configuration

11Operating System Configuration

11Cluster Management

12Parallel Virtual File System

12Performance Characteristics

12Processor

13Memory Subsystem & Scalability

15I/O Subsystem

15Interconnect

16Best Fit

17Conclusion

18References

Introduction

Over the years, clustering technology using commodity hardware and software components to form a supercomputing environment has become legitimate and well-accepted for High Performance Computing solutions. High Performance Computing Cluster, or HPCC, technology provides a cost effective method for delivering a parallel computing system platform, targeted towards compute- and data-intensive applications. Through Dells HPCC solutions, due to low price points and excellent scalability, users can aggregate standard-based server and storage resources into powerful supercomputers to solve complex computational problems. Although well understood, HPCC is still considered a complex solution due to the variety of choices of its components. Depending on the applications characteristics as well as the budget of procurement, the design for a cluster is quite unique for each customer to have a best-fit solution. In this article, we present three currently offered HPC cluster configurations based on three Dell PowerEdge 32-bit server platforms, the PowerEdge 2650, PowerEdge 1655MC, and PowerEdge 1750 servers. Then we discuss the common design considerations and the differences between the three solutions according to the cluster management features and performance characteristics.

Design Architecture

For HPC Cluster solutions, Dell leverages high-density and rack-optimized PowerEdge servers as the building blocks. The PowerEdge 1750 and PowerEdge 2650 are dual-Xeon processor systems, with 1U and 2U form factors respectively. The standard offerings include 8, 16, 32, 64 and 128-node configurations with Fast Ethernet, Gigabit2Ethernet or Myrinet used as the interconnect. The PowerEdge 1655MC, on the other hand, is a Pentium III bladed server with up to 6 blades or 12 P-III processors consolidated in a 3U chassis. Using PowerEdge 1655MC as the compute node, Dell offers 6, 18, 36, 66 and 132-node standard configurations interconnected with Gigabit Ethernet.

The three options mentioned above use the PowerEdge 2650 platform as the master node and management node. The following figures show samples of the design diagrams for the three HPCC solutions.

Figure 1 illustrates a 132-node, PowerEdge 1655MC with Gigabit Ethernet configuration. Its components include the PowerEdge 2650 for both the master and management nodes with a SCSI enclosure, and a PowerVault 220S attached to the master node for external storage. Three different fabrics are incorporated for managing this cluster: the cluster fabric, the Embedded Remote Access (ERA) fabric and a KVM over IP fabric. Details of these fabrics are described below in the Monitoring and Management section. In addition, the optional Dell PowerVault 725N NAS (Network Attached Storage) device, which provides for expansion of external storage, is also shown in the figure.

Figure 1: 132-node PowerEdge 1655MC Configuration

A 128-node PowerEdge 2650 Gigabit Ethernet configuration demonstrating 2U/2P Xeon architecture is shown below in Figure2. This design has the PowerEdge 2650 for all compute nodes as well as the master and management nodes. The PowerEdge 2650 bundles offer Fast Ethernet, Gigabit Ethernet, as well as Myrinet for a high bandwidth, low latency interconnect.

Figure 2: 128-node PowerEdge 2650 Configuration

Figure 3 shows the 1U/2P Xeon platforms, the PowerEdge 1750 128-node Myrinet configuration. In this diagram, the PowerEdge 2650 is used as the master node to enable the use of external storage and extra PCI slots for additional devices. The figure below shows the administration and ERA fabrics both networked with Dells PowerConnect switches while the cluster fabric is interconnected with Myrinet.

Figure 3: 128-node PowerEdge 1750 Myrinet Configuration

The following table summarizes the feature comparison of the three cluster configurations.

PowerEdge 1655MC

PowerEdge 1750

PowerEdge 2650

Master Node

PowerEdge 2650

PowerEdge 2650

PowerEdge 2650

ITA Node

1 CPU PowerEdge 2650

1 CPU PowerEdge 2650

1 CPU PowerEdge 2650

Compute Node

PowerEdge 1655MC

PowerEdge 1750

PowerEdge 2650

Cluster Configurations

6,18,36,66,132

8,16,32,64,128

8,16,32,64,128

Interconnect

GbE

GbE/FE/Myrinet

GbE/FE/Myrinet

CPU

PIII-1.26 or 1.4 GHz

Xeon Dual Processors 2.4 or 3.06 GHz

Xeon Dual Processors 2, 2.4,2.8, or 3.06 GHz

Storage

SCSI-PowerVault 220S and NAS PowerVault 725N

SCSI-PowerVault 220S

SCSI-PowerVault 220S

Out-of-band Management

ERA/MC

ERA-O

ERA

Power Requirement

850W

420W

415W

Thermal Requirement

2899 BTU/hr

1434 BTU/hr

1415 BTU/hr

Form Factor

3U (6 Blades)

1U

2U

Management Solution

Dells HPCC solutions provide 4 methods of managing and monitoring the cluster: ITA, Ganglia, ERA and digital KVM. ITA and Ganglia are two in-band management tools that use the cluster fabric for monitoring and management traffic. IT Assistant, or ITA, is Dells server management solution that provides a centralized management console used to discover nodes on the network and examine hardware sensor data to prevent failures at the system level. Ganglia is an OS-level cluster monitor that can be used to look at resource usage, detect node failures, and troubleshoot performance problems. Both ITA and Ganglia require OS support and use the cluster fabric for communication. ERA and the digital KVM are the tools used for out-of-band management, providing a redundant fabric that allows access to the cluster if the in-band fabric is unavailable. ERA provides remote systems management for Dell PowerEdge servers.

For the PowerEdge 1750 and PowerEdge 2650 cluster configurations, the master and ITA nodes are connected to an analog KVM switch to access the cluster. From the KVM switch, the administrator can remotely connect to the compute nodes via the master node and can monitor and manage the hardware via the ITA node. Additionally, for the PowerEdge 1655MC cluster configurations, the KVM Fabric goes through a digital KVM switch, Dell 2161DS KVM over IP Console Switch.

In-band Monitoring and Management

Dell IT Assistant

Dell ITA is a remote server and desktop management application that employs remote management technology. IT Assistant provides asset, configuration, event, security, and storage management for Dell systems and other systems equipped with industry-standard instrumentation. ITA enables the system administrator to easily specify the parts of the network that are required to be monitored in a HPCC environment. The process of locating and viewing systems on the network, called discovery, is extremely versatile, with the ability to include network nodes on a subnet, range of addresses, or individual-system basis.

Dell OpenManage ITA includes two main components: the ITA Web-based user interface and ITA services. ITA services consist of:

ITA connection service: establishes the communication with the ITA Web-based user interface

ITA network monitoring service: discovers systems across the HPCC network

ITA data repository: stores configuration and system discovery information

Command-line utilities: enables setting up CIM discovery and security, manage the ITA data repository, and write scripts to perform ITA tasks without using the Web-based user interface.

ITA offers versatile installation capabilities that allow you to install the ITA services on one system and the Web-based user interface on the same or another system. The interface seeks out individual installations of the services, making it possible to set up several services systems to "cover" different parts of the network, and manage them through a single interface. Conversely, one installation of the services can support multiple interfaces dispersed strategically throughout the network.If a Web server is installed on the system hosting the ITA Web-based user interface, remote Web-server interfaces can be run on any system that has Internet Explorer. Figure 4 shows an example of an ITA web interface.

Figure 4: IT Assistant Web Interface

Ganglia

Ganglia, an open source monitoring tool developed at the University of California at Berkley, is another in-band management tool offered in the Dell HPC Cluster configuration. From the monitoring station, Ganglia monitors and automatically graphs over 20 metrics such as the nodes load average, number of running processes, number of incoming and outgoing network packets, total and free memory on every node of the cluster, etc.

Ganglia provides graphical views of cluster information. The Cluster Report view (Figure 5) shows the overall status of the cluster and summarizes total node count, number of nodes that are up, overall load average, CPU and memory utilization for the cluster. Color-coding is used to represent CPU utilization to enable quick identification of overloaded systems. A crossbones icon indicates a node is down. Selecting a different metric in this view redisplays the screen with the value of this metric for each node, and uses the metric as a sort index when displaying the nodes.

Figure 5: Cluster Report View of Ganglia

Using Ganglia allows administrators to define and add parameters in the cluster that they want to monitor. Ganglias GUI will automatically graph those values in addition to the pre-set metrics for every node. Ganglia also simplifies cluster management by providing a remote execution environment. This feature is used for remote management, and to execute commands in parallel on multiple nodes.

Additionally, Ganglia provides the ability to monitor multiple clusters. This is especially useful in large compute centers where computational resources are grouped in smaller clusters for specialized use. The centralized console enables an administrator to monitor multiple clusters at once, while maintaining a high level of security by defining trust relationships.

Out-of-Band Monitoring and Management

During heavy communication between application components or compute nodes, in-band management and monitoring can inaccurately report network or server problems, since they share fabric with the applications. In addition, monitoring/management traffic will consume resources that are used by parallel applications. Also, if a machines OS is not responding or if the network is down, neither ITA nor Ganglia can access the nodes and provide the ability to fix the problem since both methods rely on OS and network support.

In these situations, system administrators can use the out-of-band network management methods to communicate with the cluster hardware, and diagnose or fix problems. Dells HPCC solution provides 2 out-of-band management routes: ERA and KVM.

Embedded Remote Access, or ERA

In the PowerEdge 1750 platform, the ERA controller resides on the ERA/O daughter card. The PowerEdge 1655MC has a controller known as the ERA/MC, which is on its management blade. The ERA controller on the PowerEdge 2650 is located on the riser card. All three servers include a designated port on the back of the machine, which is referred to as the ERA port or ERA. The ports are operated through 10/100 base-T IP-addressable Ethernet networking. The ERA network can be configured as an out-of-band management route, which is independent to the regular networks and utilizes its own dedicated processor, memory, buses and network connection on the ERA chip without consuming the cluster computing or network resources. If the cluster nodes become unresponsive, ERA allows the administrator to view and access the nodes remotely to troubleshoot the system. ERA provides the following functionality:

Initial configuration

Local and remote management

Scripting for automation

Remote firmware updates

Remote monitoring of fans and sensors

Remote power cycle, power down and power up

Configuration of servers, network switches, and the digital KVM through console redirection (on the PowerEdge 1655MC)

The use of ERA within a HPC cluster simplifies cluster management and allows a system administrator to monitor the hardware components remotely either through a CLI (through the serial port) or a web-based GUI console. One main utility used in ERA is racadm (remote access control and administrator), which provides the interface for monitoring and configuring the system. The racadm utility can be used through a serial port using a communications program such as minicom or HyperTerminal, through a remote interface or through a web-based console across the network. Through the serial interface, the administrator can view or modify the configuration settings on the compute nodes. For instance, the administrator can change the IP configuration of the ERA port to be able to access the GUI available on the web console (Figure 6).

An administrator can use the automated scripting feature to run configuration commands on multiple nodes. This proves to be a useful tool for making identical changes within large cluster configurations. The remote interface is currently only supported through the MS-DOS environment using Microsoft Windows, allowing the use of the racadm command for managing the nodes. The web interface can be accessed through any supported web browser using the ERA IP address or through IT Assistant. It allows the user to utilize the features of the ERA in a graphical interface.

Figure 6: Web Based ERA Console

One of the main features of out-of-band management is the ability to control and monitor the hardware from a remote location. The racadm commands on the PowerEdge servers allow the administrator to view the health status of the servers within the cluster. By allocating appropriate IP addresses to the ERA ports of all the nodes within a cluster, the administrator can assign names to each system, allowing access to individual nodes in order to utilize specific resources. Using racadm, there are multiple commands to use to troubleshoot the cause of a failure. Administrators can power-cycle the nodes individually, reset configurations and cause LEDS to blink or glow to easily identify systems within a cluster.

Digital KVM for the PE1655MC Cluster Configuration

The PowerEdge 1655MC contains an embedded digital KVM switch, which allows video, keyboard and mouse access to each blade. All access to the blades is from the management card on the chassis, which can either be through the standard analog PS2 keyboard, mouse and video ports or through the Analog Rack Interface port with a CAT5 cable. The Analog Rack Interface port can be connected directly to a port on the Dell 2161DS Digital KVM Switch with a CAT5 cable, which cascades the switches and allows them to be accessed from one central station. In large cluster configurations with several PowerEdge 1655MC chassis, cascading the embedded KVM switch with the Dell 2161DS KVM over IP Consol Switch can greatly minimize the cable organization and management.

HPCC Software Package

The 32-bit HPC Cluster configurations from Dell come with the Felix 3.1 Deployment solution stack. Felix is a collaborative effort between MPI Software Technologies Inc. (MSTI) and Dell. It presents a highly automated and scalable infrastructure for easy cluster set up and configuration for Red Hat 8.0 and 9.

The benefits of using Felix span from end-users to system administrators. System administrators have a convenient way to install/uninstall/upgrade packages, add/remove nodes, make configuration changes, distribute modifications, re-image nodes, and monitor/manage the cluster using a clean and easy to use Graphical User Interface. This ultimately allows the end-users to have a clean computing environment with a single system image and a predictable set of resources. Felix is installed on top of the operating system on the master node. Although it installs a specific kernel release, it does not require any kernel modifications and it maintains the master and compute nodes as fully functional nodes.

Felix allows users to specify various options during the steps of cluster installation. The following is a sequence of steps that Felix uses to complete a cluster deployment phase.

Cluster Configuration

Users can configure the cluster using various parameters outlined below. The configuration scheme is platform independent.

1) Starting IP address to be assigned to compute nodes. IP addresses are incremented by one and assigned to every compute node thereafter.

2) Number of compute nodes in the clusters.

3) Naming scheme for the compute nodes and their NIC (Network Interface Card) interface name (i.e., eth0).

4) An IP for the master node for both the external and cluster networks.

5) Master node name and its NIC interface to the cluster and the external network.

6) An IP scheme to use the Myrinet network (adding Myrinet to the cluster is optional).

Operating System Configuration

Felix allows users to specify custom OS parameters that can be used to install the OS on the compute nodes. A user can assign a specific kernel, ram disk, or kickstart for a compute node installation on all three platforms: PowerEdge 1655MC, PowerEdge 1750, and PowerEdge 2650. Each platform and OS release has associated kickstart files, kernels and ramdisk, although Felix is designed so that users can create and specify other components for installation. Depending upon the platform of compute nodes in the cluster, an appropriate kickstart file can be selected. Felix allows users to specify and copy the different RPMs (Red Hat Package Manager) from the Red Hat installation CDs.

Cluster Management

Felix is able to handle various scenarios where cluster functionality may be hampered by node hardware failure (i.e., hard disk failure, NIC failure, etc). Felix is able to reinstate a compute node after failure, even if it requires a complete system replacement. It is able to deal with network interface card changes. If a NIC device fails on a compute node, Felix can configure the compute node to start using an alternate NIC if one is present. Felix can also efficiently handle a combined failure of the hard drive and NIC. The master node can be similarly recovered after hardware failure. Felix provides the option of backing up the master node configuration to the compute nodes. After a hard disk swap out, the master node can be re-installed and the configuration then recovered from the compute nodes thereby returning the system to its original state.

Felix also provides a parallel command tool, which enables users to execute a command simultaneously across a selected set of compute nodes. This can be effective for restarting services, installing RPMS, copying files, etc. on a set of nodes without the burden of typing commands explicitly for each node.

Parallel Virtual File System

Felix is capable of installing a variety of packages typically used in the cluster environment, such as Lapack, Scalapack, Ganglia and PVFS (Parallel Virtual File System). The user can select a list of packages and a set of compute nodes to install the packages on allowing them to customize the cluster. In particular, to provide the cluster with a parallel file system, Felix has the option to install PVFS.

The Parallel Virtual File System (PVFS) Project by Clemson University was designed to create a parallel file system for PC clusters. The advantages of a parallel file system include a global name space, distribution of data across multiple disks, the use of multiple user interfaces, and can include additional I/O interfaces to support larger files. PVFS is designed to eliminate bottlenecks and to increase the bandwidth available for the I/O resources throughout the cluster. Instead of using message passing libraries, PVFS uses TCP/IP for communication. PVFS is relatively easy to install and start using. It is a user-level implementation, so no modifications need to be made to the kernel.

For more information on PVFS, refer to the Power Solutions article, The Parallel Virtual File System for High-Performance Computing Clusters, Power Solutions, November 2002.

Performance Characteristics

Processor

The PowerEdge 1750 platform offers the most processing power in the densest form factor among the three Dell HPCC offerings. The PowerEdge 1750 supports two Xeon processors up to 3.0GHz with 512K L2 cache and Hyper Threading support. Xeon processors support the SSE2 instruction set, which enables the processor to perform two floating-point operations (FLOPS) in one clock cycle. Therefore, a PowerEdge 1750 equipped with dual 3.0GHz processors has a theoretical peak computation throughput of 12Gflops. Another Xeon based platform that is a part of Dells HPCC offerings is the PowerEdge 2650. The PowerEdge 2650 offers the same type of processors as the PowerEdge 1750 in a 2U form factor. The PowerEdge 2650 includes 3 PCI-X slots instead of 2 for the PowerEdge1750 and five hard drives instead of 3, combined with the performance of Intel Xeon processors. The PowerEdge 2650 is suitable for compute-intensive applications, where internal storage and I/O capacity have priority over chassis size.

The PowerEdge 1655MC offers Pentium III processors up to 1.4GHz in a form factor that can fit 12 CPUs in a 3U space. This makes it an excellent choice for applications in which cost and processor density are of primary importance and applications where integer arithmetic is involved such as certain bioinformatics and life sciences applications.

Memory Subsystem & Scalability

Both the PowerEdge 2650 and the PowerEdge 1750 are based on the ServerWorks GC-LE chipset and offer either a 400MHz or 533MHz front-side bus (FSB) with 8GB and 12GB maximum memory, respectively. The PowerEdge 1655MC offers a 133MHz FSB with 2GB maximum memory. In-the-box scalability of performance for MPI-based applications depends greatly on the memory subsystem performance. This is due to the fact that the processes running on the compute nodes contend for the same memory resource. This becomes more apparent as the CPU clock speed increases since faster processors will perform the computations quicker putting more stress on the memory subsystem. Figure 10 below shows the impact of processor speed in scalability of Linpack benchmark, which is a memory intensive linear equation-solving application. The experiment was performed on a PowerEdge 2650 with 400 MHz front side bus with different processor speeds. As the processor speed increases, the overall performance also increases. On the other hand the scalability in going from a single process to two processes decreases. With the 533MHz FSB, the PowerEdge 1750 and PowerEdge 2650 show impressive scalability.

Figure 10: High Performance Linpack scalability comparison on PowerEdge 2650 (Tested by Dell Scalable Systems Group, July 2003)

To understand the effect of FSB in in-the-box scalability, take a look at Figure 11 below. The chart shows Linpack results on a PowerEdge 1550 with 133MHz FSB and 1.26MHz processor. The PowerEdge 2650 can achieve an equivalent scalability factor despite the fact that the CPU runs at a much higher clock rate. This is due to the fact that the FSB also runs faster to serve the data from memory to the CPU quicker.

Figure 11: High Performance Linpack scalability comparison on PowerEdge 2650 and PowerEdge 1550 (Tested by Dell Scalable Systems Group, June 2003)

I/O Subsystem

The PowerEdge 2650 is equipped with two 64b/100MHz PCI-X and one 64b/133MHz PCI-X slots sharing two PCI-X bus segments and the PowerEdge 1750 has two 64b/133MHz PCI-X slots on independent bus segments. This makes these PowerEdge servers a good choice for high-speed cluster interconnects that support operations at PCI-X speeds, such as Myrinet (D-interface) and InfiniBand. Below are the DMA transfer rates obtained on a PowerEdge 1750 using Myrinet D-interfaces, which would not be attainable on a 64b/66MHz PCI bus.

bus_read

(send)

=

821MBytes/s

bus_write

(recv)

=

1040 MBytes/s

Chart 1: DMA Transfer Rates for PowerEdge 1750 using GM_Debug tool (Tested by Dell Scalable Systems Group, May 2003)

Interconnect

Dells HPCC program offers bundles based on three interconnects, that is Fast Ethernet, Gigabit Ethernet and Myrinet. Fast Ethernet is the most cost effective solution for applications that dont require heavy communication between cluster nodes. Gigabit2 Ethernet is the high-performance, standards based interconnect and Myrinet is the low-latency, high throughput interconnect suitable for communication intensive applications.

Using the PowerEdge 1750 (533MHz version) and on-board Gigabit Ethernet ports, it is possible to achieve an MPI-level latency as low as ~50us. This is due to the fact that the Gigabit Ethernet NICs are embedded within the chipset rather than being attached through a PCI bus to achieve the best performance.

Figure 12: High Performance Linpack scalability comparison on PowerEdge 2650 (Tested by Scalable Systems Group, July 2003)

Just as in-the-box scalability depends heavily on the memory bandwidth, the out-of-box scalability depends on the I/O and interconnect performance. The figure above shows Linpack performance obtained from a cluster of PowerEdge 2650s with Myrinet interconnect. In this experiment, the same benchmark is run over an increasing number of compute nodes and the resulting performance is plotted with respect to the amount of memory involved in the computation. As can be seen from the figure, there is almost perfect scaling in going from 16 nodes to 32 nodes; this is due to the fact that Myrinet interconnect provides a low latency and high-throughput communications path between processes running on different nodes.

Best Fit

Each of the HPCC solutions addresses different segments of the applications domain in the high performance computing field. This section explains the best fit applications domain for each solution.

The PowerEdge 1750 cluster configuration offers the most processing power in the densest form factor among the three DellHPCC offerings with its Xeon processors capable of performing two floating-point operations in one clock cycle due to its built-in SSE2 instruction set. This makes this HPCC solution an excellent choice for applications in which heavy floating point computations are involved, such as computational fluid dynamics, manufacturing, applied physics and EDA applications. In addition to the processing power of Xeon, the 1750s 400 MHz front-side-bus or the 533 MHz FSB provides faster memory access to move data in and out of memory compared to the Pentium III architecture. A majority of the applications types mentioned above are also very memory intensive. They require fast data movement from memory to cache to processing units since the data sets are too large to fit in the cache. Therefore, it makes this offering the best solution for highly data oriented as well as floating-point intensive applications. Each compute node has Gigabit Ethernet for high-bandwidth networking. The PCI-X 64b/133MHz slot on compute nodes may be used for additional expansion for low latency high-bandwidth networking using high-performance networking solutions such as Myrinet. These features make this solution very suitable for applications sensitive for bandwidth and/or latency due to the amount and/or frequency of data being transferred during computation.

Another Xeon-based offering from Dell is the PowerEdge 2650 cluster configuration. The PowerEdge 2650 offers the same type of processors as the PowerEdge 1750, but in a 2U form factor. The applications domain of this solution will be very similar to the PowerEdge 1750, but allows for more internal storage and PCI slot expansion. The 2U form factor allows each compute node to hold up to 5 hard disks, therefore making this solution very suitable for applications where a lot of local disk space is required in each compute node. This cluster solution can also be configured as an I/O cluster for serving a large parallel file system for another HPCC cluster for fast file access through TCP/IP. Having 3 PCI-X capable slots available as well as two onboard Gigabit Ethernet NICs on each node makes this solution highly flexible with a lot of expansion capability in terms of local storage and I/O capabilities. In short, the PowerEdge 2650 cluster configuration is suitable for compute-intensive applications where internal storage and I/O capacity has priority over the processor density.

In the Pentium III PowerEdge 1655MC configuration, each blade can hold up to two processors and two hard drives, so a single enclosure (3U) can have 12 CPUs and 12 drives, which provides the highest density per U compared to the other two configurations. This makes the Pentium III PowerEdge 1655MC configuration an excellent choice for applications in which cost, processor and storage density are of primary concerns and for the applications where integer arithmetic is most involved. Applications such as bioinformatics or genome sequencing in life science, and distributed data mining, are good examples of applications that would benefit from adopting this solution.

For information regarding the Dell HPC Cluster products, please visit the Dell HPCC web site.

Conclusion

Dells HPCC 32-bit configurations offer a complete solution for most high-performance computing environments. The design choices of hardware, software and management solutions need to be carefully considered for various types of applications. The in-band and out-of-band cluster management features offered in the solutions provide system administrators a reliable, highly available, and easily managed operational environment. The Felix software stack, specifically designed for Dell hardware along with commonly used open source cluster software, can be ordered with each cluster. With the PowerEdge 1750, PowerEdge 2650 and PowerEdge 1655MC cluster configurations, Dell delivers high computational power and performance at a cost effective price in an all-inclusive HPC Cluster solution.

Footnotes

U=1.75 inches

2This term indicates compliance with IEEE standard 802.3ab for Gigabit Ethernet, and does not connote actual operating speed of 1 Gb/sec. For high speed transmission, connection to a Gigabit Ethernet server and network infrastructure is required

References

1. Carns, Philip H. and Walter B. Ligon III. "PVFS: A Parallel File System for Linux Clusters." Parallel Architecture Research Laboratory, Clemson University. Proceedings of the 4th Annual Linux Showcase and Conference, Atlanta, GA, October 2000.

2. Ligon III, Walter and Robert B. Ross. "An Overview of Parallel Virtual File Systems" Clemson University. Proceedings of the 1999 Extreme Linux Workshop, June, 1999.

3. Garg, Sharad and Jens Mache. "Performance Evaluation of Parallel File Systems for PC Clusters and ASCI Red." Proceedings of the 2001 International Conference on Cluster Computing (Cluster '01)

4. DellTech/Support : http://delltech.us.dell.com/support/

5. Ganglia: Distributed Monitoring and Execution System :http://ganglia.sourceforge.net/

6. Serial and Remote Execution of CLI Commands for Blade Server Management, Dell PowerSolutions, August 2002: http://www.dell.com/us/en/esg/topics/power_ps3q02-suniti.htm

THIS WHITE PAPER IS FOR INFORMATIONAL PURPOSES ONLY, AND MAY CONTAIN TYPOGRAPHICAL ERRORS AND TECHNICAL INACCURACIES. THE CONTENT IS PROVIDED AS IS, WITHOUT EXPRESS OR IMPLIED WARRANTIES OF ANY KIND.

Dell, PowerEdge, PowerVault and PowerConnect are trademarks of Dell Inc. Intel and Pentium are registered trademarks and Xeon is a trademark of Intel Corporation. Microsoft and Windows are registered trademarks of Microsoft Corporation. Other trademarks and trade names may be used in this document to refer to either the entities claiming the marks and names or their products. Dell disclaims proprietary interest in the marks and names of others.

Copyright 2003 Dell Inc. All rights reserved. Reproduction in any manner whatsoever without the express written permission of Dell Inc. is strictly forbidden. For more information, contact Dell.

Information in this document is subject to change without notice.

2

ii

_1117289116.bin

_1117352803.bin

_1117439505.bin

_1117287983.bin