Software Sizing Guide
White Paper[ This document is also available in PDF format. ]
Contents
Information contained in this sizing guide is based upon guidelines developed by Sun Microsystems for the Solaris PC NetLink Blueprint currently in development, Readers should monitor the blueprint website (http://www.sun.com/blueprints) for availability of this document, as well as other related information and documentation.
IntroductionOrganizations need to deliver network services and business-critical applications while controlling costs, maximizing uptime, and maintaining flexibility for future growth. LANs and workgroups consisting primarily of personal computers often fail to meet the demands of the heterogeneous, web-centric enterprise. Sun understands this dilemma, and continues to deliver innovative solutions like new interoperability technology that address these needs.Today, corporate computing environments support multiple workgroups, frequently populated with personal computers connected through a local area network. Many of these workgroup installations have been deployed using Microsoft Windows NT protocols for managing users accounts and resources, hosted from PC servers. Unfortunately, the reliance on PC servers is not without problems-their proliferation has resulted in higher management costs and frequent service interruptions. The regrettable truth is that PC servers remain PCs-they are significantly less scalable, and are not built for the high reliability and availability needed in enterprise computing environments. If managers could only find a way to move departmental network services from PCs to more reliable, scalable platforms without changing their infrastructure, the benefits would be immediate and obvious.
Solaris PC NetLink SoftwareSun Microsystems, aware of the need to accelerate the performance and efficiency of heterogeneous enterprise workgroups, has recently announced the Solaris PC NetLink software. Fully compatible with network technology from Microsoft, Solaris PC NetLink provides the key network services and resource sharing facilities required of Microsoft Windows 3.11, Windows NT, Windows 95, and Windows 98 clients.Based on AT&T's Advanced Server for UNIX®, Solaris PC NetLink is Sun's next generation PC LAN integration server. A powerful tool for integrating PCs into the enterprise network, Solaris PC NetLink provides transparent and seamless access to key Microsoft Windows NT 4.0 services, such as file, print, and directory services, in addition to addressing critical enterprise computing needs:
Solaris PC NetLink software running on scalable Sun servers and the Solaris Operating Environment finally gives workgroup administrators the ability to configure departmental servers for unprecedented levels of reliability, availability, serviceability, and economy. Indeed, this combination enables enterprises to reach a long sought after goal-the ability to provide PC clients with the functionality they need using fewer, highly reliable servers running in a more robust, proven environment. In short, Solaris PC NetLink provides numerous important benefits to growing workgroups:
The Importance of Sizing Network Server EnvironmentsScalability is critical to companies needing to make the most cost-effective use of their computing resources, gracefully handle peak workloads, and grow their computing environment as their business grows. There are three aspects to scalability:
Sizing MethodologyBefore an organization can determine how best to configure and utilize servers running the Solaris PC NetLink software, the computing environment must be analyzed to ensure it is sized appropriately. In particular, organizations should:
Define System LoadOrganizations need to determine their user load in order for any sizing analyses to be useful. The user load is defined as a mix of file operations, including multiple writes for autosave functions. Email and printing activity must also be taken into account. It is not unreasonable, for example, for a typical user to consume 20 MB/day of throughput-a value which will be used in this document to reflect a typical user load. To anticipate future growth and technology trends, this number should be doubled. In short, allocate 40 MB/day per user to ensure daily needs can be met and that future needs can be handled adequately.
Determine Peak TimeOrganizations need to determine their peak time and ensure the system can support it. It is important to remember that peak time-the most critical time to support users-will vary from site to site. While no server can handle the simultaneous use of all of its services by the entire user community, steps can be taken to ensure it handles peak loads well. To do so, organizations must determine when the peak load will occur and for how long, whether that occurs for half an hour each morning and afternoon, or an hour in the middle of the day, or both. The system should be sized assuming the user load for the entire day will occur during this peak time. Once such patterns are analyzed and understood, choices can be made that help the system handle the load and provide the services users demand.
Measure the SystemTo evaluate server systems, such as those running the Solaris PC NetLink software, Sun utilizes the Ziff Davis benchmark Netbench-an industry accepted benchmark that measures server file throughput. During each test, PC clients make network file requests. As the operations are performed, Netbench determines how much data is moved with each request, as well as how quickly the task is performed. The throughput calculations for each client are combined to determine overall server throughput.
CPU UtilizationFor many years, a central theme in the development of commercial systems has been the creation of faster processors. This emphasis paid off handsomely as increasingly sophisticated technology allowed the development of faster and faster systems. However, applications have become more sophisticated, and now require faster architectures to get the job done. Indeed, many applications yield the best results when hosted on scalable, parallel processing systems.The first, and easiest, place to start is to analyze the CPUs involved. With very few exceptions, little can be gained from a system once the CPUs have become saturated. It is important, therefore, for administrators to determine how many CPUs are available, and whether or not they can handle expected loads. Furthermore, it is important to be able to understand how many processors can be added to the system before no further performance gains can be realized. To understand the effects of CPU utilization and its impact on system throughput, Sun tested a Sun Enterprise 450 server running the Solaris PC NetLink software with Netbench. The Sun Enterprise 450 server was configured with one to four 400þMHz UltraSPARC processors, 1 GB of memory, twenty 10,000 RPM SCSI drives, and a Gigabit Ethernet connection. Figure 2 depicts the effect of processor scaling in a Solaris PC NetLink server environment. The maximum relative performance of the Sun Enterprise 450 server when 1 to 4 processors were enabled on the system indicates that a 4 processor system delivers a maximum throughput of 15.76þMB/second.
Calculate Maximum Number of UsersOrganizations need to know how many users can be accommodated by any given system. Indeed, the level of service the server is expected to deliver during peak times typically determines the resources needed. Since the nature of peak loads varies from site to site, a conservative estimate is appropriate. By determining a reasonable peak time and applying the results of benchmarks, it is possible to determine the number of users the system can support.Figure 3 suggests a simple yet effective measurement technique that indicates that the maximum number of users capable of being supported on a system for a particular peak period is equal to the system throughput multiplied by the peak time divided by the load. [Number of Users = (System Throughput x Peak Time)/Load]. It is important to note that this suggests a theoretical maximum number of users that can be supported at the peak and should only be considered a baseline from which other server considerations and trade-offs should be made. The previous doubling of the user load and using only 80% the maximum measured throughput are conservative measures that endeavor to make this user count realistic.
 Calculation using example data listed above: [(12.61 MB/second) x (3000 seconds)]/(40 MB/user) = 946 users
Furthermore, it is assumed that the users place a peak on the server for 25 minutes in the morning as well as for 25 minutes in the afternoon, for a total peak time of 50 minutes. Consequently, the server will have more than sufficient resources during off-peak times. As a result, 12.61 MB/second throughput can handle 946 users each utilizing 40 MB per day during a combined peak time that lasts one hour. Again, this is a baseline by which further functionality and cost trade-offs may be required, and planning for user disk capacity and redundancy should be considered.
System ThroughputSimply measuring overall system performance is not sufficient. Indeed, system throughput is a critical factor, and organizations need to understand which subsystems are underperforming and ensure those problems are resolved. To ensure the entire system performs well, each subsystem must be evaluated, configured, and tuned to enable it to handle peak loads. In particular, administrators need to analyze the following:
MemoryDetermining the memory requirements for a Solaris PC NetLink server requires several considerations, including the needs of the Solaris Operating Environment as well as the Solaris PC NetLink software.
Disk ThroughputIt is not unusual for storage capacity requirements to grow by as much as 100% per year. Coincident with this growth is an increasing need to ensure ready access to data whenever it is needed. System administrators know that raw capacity and performance do not stand alone-both scalability and high availability are important practical requirements for ensuring economical and reliable operation and information access. Indeed, disk subsystems must be configured and tuned to match the performance, availability, and cost requirements of users, and that the user load be distributed across disk subsystems that have sufficient redundancy and throughput to match what CPUs can deliver. When sizing a server, organizations need to focus on both capacity (MB/user) and throughput (MB/second) as either one can be a limiting factor.Redundant Arrays of Inexpensive Disks, or RAID systems, were first formally defined by researchers at Berkeley in 1987. The most important part of their idea was to combine the power of small, inexpensive drives to achieve performance equal to a single large, expensive disk. Because RAID required multiple spindles, features were also needed that could protect against the data loss that could be caused by the failure of any one of the drives. These original safeguards subsequently became an important feature in their own right, with many users purchasing RAID systems just for their availability. Today, disk arrays are sold as an intelligently managed collection of disk drives organized to optimize performance for a particular set of tasks. All RAID systems achieve higher performance and/or reliability by replicating or spanning data across multiple disks. Exactly how this is done has profound effects on subsystem performance, cost, and reliability. Multiple RAID levels, numbered 0 through 5, offer varying degrees of performance, reliability, redundancy, and availability. Sun suggests the following guidelines for enabling sufficient disk throughput on Solaris PC NetLink servers:
Disk CapacityWhen planning server system disk space, administrators must include additional space for Solaris PC NetLink system software and databases. Sun suggests the following items be considered:
Disk SizingDisk sizing requires two considerations for the sizing exercise, and one for RAS and cost trade-offs. Disk sizing requires disk subsystems meet throughput and capacity requirements. System planners must decide the RAID configurations required, and consider if software or hardware RAID implementations are best suited to meet performance or cost priorities. Administrators should note that software RAID, particularly RAID-5, can require significant CPU resources which were not part of the Solaris PC NetLink CPU sizing exercise.To determine the effects of disk subsystems in Solaris PC NetLink server configurations, Sun tested several Sun Enterprise servers running the Solaris PC NetLink software with multiple disk subsystem configurations. Table 2 describes the maximum throughput performance of these systems under various load conditions.
Table 2 Netbench disk subsystem performance comparison Consider a sample calculation using the previous CPU example as a basis, and determine the number and type of disks required to support the configuration during expected peak conditions. Previously it was determined that 12.61þMB/second was the peak throughput to support 946 users for a 50 minute peak duration. In such a scenario, a Sun Storedge SRC/P RAID Card with eight disks configured as a RAID-5 subsystem could meet the throughput requirement. But what about capacity requirements? Since 100 MB of disk space is needed per user, 946 users would require 92 GB of storage space. However, an 8-disk RAID-5 volume using 9 GB drives would only provide approximately 58 MB of storage capacity. Consequently, at least two such volumes would be needed to meet capacity needs, and would have the side effect of enabling throughput requirements to be met more comfortably. Administrators need to consider such issues to ensure the proper amount and type of disk storage is employed.
Network ThroughputNetworks must not only keep pace with technological advancements, they must ensure that sufficient resources are available to meet the demands of increasingly sophisticated applications.When a network server is unable to respond to a client's request, the client typically retransmits the request a number of times. Each retransmission induces additional system overhead and generates more network traffic. Organizations can mitigate the extent of excessive retransmissions by improvements in three areas: data integrity, system performance, and network congestion. Sun suggests the following guidelines for network loading:
Solaris PC NetLink Sizing GuidelinesDuring the course of its analysis and testing of the Solaris PC NetLink software, Sun has determined that small, medium, and large configurations require different resource allocations to ensure peak system efficiency. Table 3 describes the suggested configurations for Solaris PC NetLink server environments.The Sun blueprint website (http://www.sun.com/blueprints) contains documentation and other information related to the sizing of Solaris PC NetLink servers. In addition, Sun has created a spreadsheet that enables organizations to try out different configurations and scenarios and understand the impact on server sizing. Tools such as these are not a panacea, and cannot take into account the nuances of every environment. Consequently, use of this spreadsheet requires close attention to its underlying working assumptions. The sizing spreadsheet will be available on Sun's blueprint web site in the near future.
1. Add one processor if software RAID-5 is used. Table 3 Suggested configuration sizing for Solaris PC NetLink server environments
Performance SnapshotWhen sized appropriately, Solaris PC NetLink servers perform well and provide the throughput required for most needs. Figure 5 shows the results of the Netbench 5.01 benchmark executing on two Sun Enterprise servers running the Solaris PC NetLink software. The Netbench benchmark was executed using sixty 200þMHz PCs running Microsoft Windows 95 as Netbench clients. The PCs were attached to a network switch via 100 Mbit full-duplexed connections, and the server was attached to the switch via a Gigabit Ethernet connection. Both servers had 1 GB of memory.The results illustrate the performance of a Sun Enterprise 450 server with four 400 MHz UltraSPARC-II processors as well as that of a Sun Enterprise 6000 server with twelve 250 MHz UltraSPARC-II processors. It is important to note that as the benchmark adds more 100% duty cycle PC loads, server throughput responds to the load. Keep in mind that the loads placed on the servers by the PCs running the benchmark place a significantly greater load than a typical user attempting to access files via PC-based productivity applications. For example, one simulated Netbench client running the Netbench benchmark could simulate 10 to 100 actual PC clients (with the number varying greatly depending upon actual user conditions) because one Netbench client stresses the server in a manner far beyond a typical PC user.
The Sun Enterprise 6000 server maintained higher throughput with its twelve 250 Mhz processors, however the Sun Enterprise 450 server obtained more throughput for each of its 400 MHz processors. In addition, the Sun Enterprise 6000 server continued to increase slightly at the end of the benchmark, indicating that additional throughput might be possible. Note that the Solaris PC NetLink registry values were changed to permit the Sun Enterprise 6000 server to utilize all of its 12 processors during the test. Any slight perturbations in the benchmark are caused by Solaris PC NetLink lmx.srv processes coming on-line one at a time. As new processes are spawned by additional PC clients, the benchmark accelerates slightly as more processors are involved in the test.
Solaris PC NetLink -þProven Performance and Unparalleled ScalabilityWorkgroups are fast becoming a business-critical component of enterprise computing. This new reliance on the role of workgroups is causing organizations to revisit their implementations. Indeed, workgroups based on personal computers are coming up short, and organizations need more reliable and scalable platforms from which to deploy business-critical computing services.
Sun has provided foundation-level products for mission-critical computing for over 15 years, and stands prepared to deliver the solutions needed to advance the effectiveness of workgroup computing environments. The fact is, every environment is different, and some will achieve more throughput than The Solaris PC NetLink server software was designed to provide native Microsoft NT Server functionality in the robust Solaris Operating Environment. Like all server systems, those based on Solaris PC NetLink need to be sized and configured to match the needs of users. Indeed, only by measuring such systems, properly determining peak loads and times, and making configuration adjustments can organizations ensure they operate at peak efficiency. The combination of Solaris PC NetLink software running on scalable Sun servers and the Solaris Operating Environment finally gives departmental computing administrators the ability to configure workgroup servers for extremely high levels of reliability, availability, serviceability, and economy. Indeed, this combination enables enterprises to reach a long sought after goal-the ability to provide PC clients with the functionality they need, but consolidated onto fewer, highly reliable platforms running in a more robust, proven operating environment. More information on Solaris PC NetLink and related Sun products can be found at http://www.sun.com on the World Wide Web.
Sun Microsystems, Inc. posts product information in the form of data sheets, specifications, and white papers on its Internet World Wide Web Home page at: http://www.sun.com. |
| ||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||
 |
| ||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||