D.H. Brown Associates, Inc.
222 Grace Church Street
Port Chester, NY 10573
(914) 937-4302 ext. 272
FAX: (914) 937-2485
abbott@dhbrown.com

About D.H. Brown Associates, Inc.

D.H. Brown Associates, Inc., (DHBA) is a leading research and consulting firm, concentrating on the strategic analysis, assessment and evaluations of technologies, products and market trends in the "Information Industry." DHBA concentrates on thoroughly understanding and continuously updating its knowledge of the core technologies and trends underlying current and future products. DHBA provides its clients with the high level balanced perspective coupled with the technical depth required for tactical and strategic decisions to fully capitalize on existing and emerging technology. For additional information, please contact Fred Abbott, Executive Vice President Open Systems, at 914-937-4302, ext. 272.

This report was jointly prepared by the DHBA services: Departmental and Desktop Computing and Parallel Open Systems Hardware, with assistance from the Windows NT Pricing and Configuration Monitor and Solutions for Enterprise Management programs.

About Our Programs

Departmental and Desktop Computing: The Departmental and Desktop Computing Program (DDC) assesses the enterprise network requirements to support full-function operating systems on the desktop. DDC evaluates the shift of historical boundaries between PCs, workstations and network servers in terms of performance and functionality.

Parallel Open Systems Hardware: Parallel Open Systems Hardware (POSH) evaluates how architectural parallelism enables scalability -- from microprocessors to Massively Parallel. In between, POSH investigates how multiple processors work together in mid-range and Enterprise level computing -- SMPs: both bus and switch based, Clusters, NUMA, etc. POSH emphasizes analysis of architectural strengths and weaknesses of products of leading vendors, and explores tradeoffs in performance, price, and other factors that influence broad acceptance.

Windows NT Pricing and Configuration Monitor: The NT Pricing Monitor analyzes competitive pricing trends and market positioning of NT servers by comparing list pricing, product features, and performance characteristics available on leading supplier offerings. Fully-detailed comparisons are immediately updated for product announcements or pricing changes for each of the major suppliers.

Solutions for Enterprise Management: SEM examines the technologies, tools, and organizational requirements to harness the explosion in desktop computing to serve corporate enterprise goals. The program targets the technological changes taking place in the operation and support of the distributed client/server environment. DHBA’s analysis clarifies the issues, positions, and solutions for distributed software and network management, combined with disciplined polling of leading-edge users for best practices.

A summary of this report is available to all of our subscribers free of charge. Sponsors of our programs in Departmental and Desktop Computing, Parallel Open Systems Hardware and Windows NT Pricing and Configuration Monitor receive the full report as part of our comprehensive services. Those interested in the program should contact Fred Abbott, Executive Vice President, Open Systems, at 914-937-4302, ext. 272.

NT Servers Drive Into The Enterprise

SUMMARY   (contents)

The WorldMark 4300’s outstanding combination of scalability and high-availability features which extend NT’s range, wrapped by NCR’s comprehensive service capabilities, should draw the attention of enterprise users. Although none of the studied solutions support more than four processors yet in SMP configurations, the WorldMark 4300’s design will allow current 4300s to be upgraded to 8-way systems, targeted for 3Q97. The upgrade will also allow 4300s to support 8 GB of memory, enabling future support for Very Large Memory addressed in NT 5.0.

Digital’s AlphaServer 4000 and NCR’s WorldMark 4300 offer the greatest amount of file storage, with external disk configurations enabling capacity up to15 TB and 14.4 TB respectively. The next closest competitor, IBM’s 704/200, supports 4.4 TB with external configurations while Dell trails at sub-TB capabilities. Digital’s AlphaServer 4000 leads in terms of the number of slots, however, offering superior expandability for peripherals such as network adapters.

The studied systems all bundle systems management enhancements. In the context of this study, though, which focuses on enterprise NT servers that primarily run mission-critical applications, the alternatives differentiate very little with the exception of Dell, which lacks several features included by others.

NCR offers the strongest clustering options for NT today in advance of Microsoft’s standard Wolfpack clustering, with leading features such as three-node support and options for remote mirroring. Digital follows closely, with its Digital Clusters for NT option. The remaining vendors have chosen to wait for Microsoft to issue its standard Wolfpack clustering.

All studied systems fall along a wide spectrum in terms of pricing based on apples-to-apples comparisons of similarly configured systems. Typically, though, more aggressively priced alternatives, such as Dell and Compaq, trade off the availability of enterprise-grade NT support directly from the respective vendor.

Contents

• SUMMARY
• SMP SCALABILITY
• EXPANDABILITY
• SYSTEM MANAGEMENT
• CLUSTERING SUPPORT
• BUSINESS FACTORS
• APPENDIX I: PRICING CONFIGURATION DETAILS

SMP SCALABILITY    (contents)

The leading NT vendors listed in Figure 1 represent both those moving upwards from the PC space and those extending downward from a tradition of enterprise systems. Despite differing heritage, both groups face the same challenges in creating servers that deliver NT scalability while retaining the low prices associated with commodity components, especially Intel’s Standard High Volume (SHV) 4-way design. Among the systems of Figure 1, NCR’s WorldMark 4300 stands out in its ability to provide scalability and upgradeability while remaining on a commodity-like pricing curve.

Indeed, the challenge is as much meeting commodity-like price expectations as it is developing a large shared memory server. The engineering techniques are well understood; many of the WinNT server players have substantial experience designing large n-way RISC platforms for UNIX. Instead, the objective is to make the appropriate price-performance design tradeoffs to deliver extended scalability on an SHV-like price curve. Surrounding Intel processors with unique n-way design may add significant cost, not only for the additional chips themselves, but also for the amortized development expenses. In addition, with the short product cycles typical of Wintel competitors, any development delays in time-to-market may substantially impact meeting a window of opportunity. Thus, the challenge relates to discovering clever ways to leverage commodity components while creating demonstrable differentiation. This section outlines some of the alternative designs for greater than 4-way Pentium Pro, highlighting utilization of common building blocks.

Much of the exploration of innovative design points has been led by smaller players, such as Advanced Logic Research, Axil and Corollary. Mainline UNIX vendors with their own RISC SMP lines, Digital, Hewlett-Packard and IBM to name three, have felt less pressure to extend Intel platforms beyond 4-way, depending on their UNIX lines to satisfy performance needs of the midrange and above. For example, HP has indicated it will provide 8-way NT platforms in the future, but for the near term will depend upon increased cache sizes and Intelligent I/O (I2O) to improve overall performance. Digital, in the unique position as the only vendor offering NT on both Intel and RISC, positions its Alpha-based servers as extending well beyond the capabilities of 4-way Intel. While greater than 4-way NT on Alpha has been demo-ed, and an early adopter program begun, formal availability of AlphaServer 8200/8400 with NT is not expected until the Fall. Also, despite Digital’s early success with its FX!32 translator, many customers feel comfortable remaining with Intel on NT. UNIX vendors with Intel-driven systems, NCR and Unisys for instance, have an additional alternative. By leveraging their UNIX SMP design points, much of the product development cost has already been covered. Still, keeping up with the rapidly changing Intel roadmap favors those designs that best exploit Intel’s standard components. In this regard, NCR’s WorldMark 4300 offers an intriguing blend of common components and creative design that yields an attractively priced 4-way easily upgradeable to 8-way.

First to market with greater than 4-way Pentium Pro for NT, Advanced Logic Research’s (ALR) approach was to squeeze two more processors onto the P6 bus. Physically inserting two additional processors on a standard SHV bus was not actually possible due to P6 electrical specifications and Pentium Pro thermal limits. Instead, ALR employed two 3-way daughter cards perched on a common P6 bus. As for Pentium Pro’s built-in 4-way bus arbitration, ALR assigns one ID as a "phantom" processor on each daughter card. As illustrated in Figure 2, synchronizing the arbiters ensures that one daughter card holds in the "phantom" state during the period that the other daughter card authorizes access to the shared system bus. A key benefit of ALR’s approach ¾ leveraging SHV components, including P6 bus and associated Orion (450GX) chip set, minimizes re-design, thus speeding the product to market.

Figure 2: ALR 6-way

Note that since only one of ALR’s six processors can access the system bus at any time, overall performance may be limited by shared memory and I/O. That is, going from four to six processors may not gain appreciable performance if the application is memory bus constrained. Until performance measurements have been released, specific application suitability can not be easily projected. Still, since the cost increment to add two processors is relatively small, attractive price/performance should accrue for some codes. This simple, early-to-market approach has enticed both Data General and Unisys to OEM ALR’s design.

In addition to offering the ALR 6-way, Unisys has also announced its own 10-way NT server. Retrofitting the UNIX-driven SMP 61000 with Pentium Pro, the 10-way XR/6 replaces Intel’s P6 bus with Unisys’s own Synchronous Coherent Memory (SCM) bus. Supporting split transaction memory references, the SCM offers up to 533 MB/sec sustained data rate, only modestly higher than the P6 bus. Maximum memory doubles to 8 GB, a better balance for the larger number of processors.

To reduce shared system bus traffic, Unisys inserts Level 3 caches for each pair of processors, as seen in Figure 3. For codes with a high L3 cache hit rate, that is when desired data is often found in the L3 cache, then performance may well scale. On the other hand, applications with a large memory write content, or those sharing appreciable data across all processors, may stress the capacity of the shared memory and I/O bus. Once again, evidence of scalability awaits performance measurements, such as the TPC benchmarks. Although XR/6 hardware becomes available in June, full support of 10-way under NT is not scheduled until early 1998.

Figure 3: Unisys XR/6 10-way

Note that NCR also offers a similar upgrade to Pentium Pros for its installed WorldMark SMPs. Originally designed with Pentium chips, the 8-way 4100 and 16-way 4500/4550 run Windows NT as well as NCR’s own UNIX SVR4 MP-RAS. Similar to Unisys’s approach, an additional level of cache, in NCR’s case up to 16 MB per processor, reduces shared memory bus contention. Upgrades to 200 MHz Pentium Pros are available, however, greater than 4-way has not yet been shipped. NCR positions the 4100/45x0 processor upgrades as providing investment protection, and suggests that a system originally balanced for Pentium processors, be it their own or Unisys’s, may not achieve optimum Pentium Pro performance. As such, NCR has created its WorldMark 4300 system specifically for Pentium Pro.

Applications demanding more robust memory bandwidth, or greater I/O connectivity, beyond the alternatives covered above, could be addressed by SMP designs with multiple active P6 buses. Although SHV boards may serve as a building block, such SMPs require significant additional design ¾ in "glue" chips bridging the SHVs or in multi-ported memory subsystems. A number of vendors have indicated plans for 8-way systems including Axil, Corollary and NCR. Their approaches differ mainly in how the buses are bridged to each other, which also reflects differences in the business focus between Axil/Corollary and NCR.

Both Axil and Corollary seek to make their name through sophisticated design. Both firms seek others to help market their technology; neither is an established system supplier. Corollary Inc. is strictly a design house, licensing its technology to others and not attempting to sell servers under its own banner. Previous Corollary chip sets were utilized in Pentium SMPs marketed by a broad range of vendors, including Data General, Hitachi and Samsung who have now indicated their intention to ship Corollary Profusion 8-way Pentium Pro systems. Axil Computers, a subsidiary of Hyundai Electronics of America, does ship its own 8-way NT servers but is understood to be looking for other distribution partners.

The Axil and Corollary design points employ similar approaches in accessing a common memory. As depicted in Figure 4, central "glue" logic is employed to bridge two P6 buses, each with four Pentium Pro processors. Both sets of processors contend for the central memory, interleaved to provide adequate bandwidth. The "glue" logic provides arbitration between requests from two active P6 buses and assures that appropriate cache coherency addresses are reflected from one P6 bus to the other. Both vendors incorporate "snoop filtering" logic that determines which coherency requests must span both buses, removing unnecessary snoop addresses to reduce P6 traffic.

The Axil and Corollary designs differ with respect to I/O attachment. Corollary’s ProFusion adds a third P6 bus dedicated to I/O to reduce load on the processor buses. I/O traffic is estimated at 15% of P6 load in transaction processing environments and perhaps substantially higher with other applications. Corollary’s three bus design reduces processor bus contention, thus improving performance, but at the cost of additional "glue" logic for the separate I/O bus. The approach taken by Axil’s Northbridge system keeps I/O connections on the processor buses to retain use of Intel I/O chip sets. To compensate for the P6 bus I/O load, Axil optimizes performance through a sophisticated 16 bank memory subsystem that reorders memory requests to increase efficiency. That is, Axil retains SHV-like I/O but adds additional memory subsystem complexity.

Figure 4: Axil and Corollary 8-ways

Figure 5: NCR 8-way OctaSCALE

The differences between Figures 4 and 5 are subtle but important. The NCR design point emphasizes lower cost and ease of upgrade while accepting a potentially small performance compromise. While not obvious from the Figures, NCR’s "glue" logic is far less extensive than that of Axil/Corollary. NCR’s goal is to ship competitively priced platforms; design elegance is secondary. Splitting Figure 5 vertically in half would yield a 4-way system very similar to Intel’s commodity SHV. In fact, the same Intel Orion P6 bus controller chips are employed within NCR’s "glue" as are found in Intel’s SHV. NCR currently ships such a 4-way subset as a WorldMark 4300, competitively priced against platforms that are limited to 4-way. Connecting another four processors via a shared P6 I/O bus allows current 4300s to be upgraded to 8-way systems, targeted for 3Q97.

Note though, that joining 4-ways via an additional P6 bus does represent a price-performance trade-off. If the left hand set of processors requires data from the right hand memory, it must incur the additional latency of traversing the bus. Bridging via the P6 I/O bus should not represent a large contention point. Even if half the memory traffic from each set of processors was carried on that third bus, the aggregate load is still similar to P6 SHV design point. Furthermore, NCR has developed a mechanism, described as Intelligent Locality Management System (ILMS), that strives to keep most memory access to the "local" ("near") memory, avoiding the delay of accessing "remote" ("far") memory via the bridge bus. ILMS hardware monitors requests for each 4 KB page of memory. If requests frequently traverse the bridge bus, i.e., memory "remote" from the requester, then ILMS will migrate that 4 KB to locally attached memory. Note that ILMS relies on hardware only, not involving any modifications to the operating system or applications. Unlike full NUMA approaches, on the 4300 only minor access latency differences exist between "near/local" and "far/remote" memory. Thus, the penalty of accessing "remote" memory remains small. Consequently, hardware-only ILMS may adequately boost performance, avoiding changes to operating system or applications. Nonetheless, NCR complements ILMS with its SMP Utilization Manager (SUM), a performance tuning tool to designate affinity between processes and processors. By assigning associated processes that share data to the same memory subsystem, SUM can reduce ILMS page "ping ponging." That is, if four processors are to be allocated to SQL Server, then SUM can assign them to the same "side" so that shared data resides in the same memory subsystem. SUM can assign other applications to the other "side" to balance workload in the typical multi-application server environment.

NT’s kernel design satisfies functional requirements for a competitive SMP implementation, including fine-grained locking and kernel threads. Evidence has emerged showing that NT 4.0 scales competitively on SMP systems containing up to four processors on commercial benchmarks such as TPC-C, which measure a system’s ability to run single, heavy duty tasks such as databases that stress both compute resources and I/O (see Figure 6).

However, performance data based on industry-standard benchmarks for NT systems with more than four processors remains sparse. In particular, no NT-based TPC-C results have yet been published using more than four processors on any system, indicating that the NT kernel still requires tuning that optimizes for advanced SMP scalability. Indeed, Microsoft promises to introduce optimizations targeting improved SMP scalability, particularly on eight-processor systems, in the Enterprise Edition of NT 4.0 later this year, and in NT 5.0 in 1998.

Targeting opportunities to introduce more sophisticated SMP capabilities in advance of NT support for single-application scalability, NCR’s SUM allows system administrators to apply affinity between applications and groups of CPUs. By allowing administrators to control the distribution of workloads among system processors, the Utilization Manager effectively allows an eight-way server to be partitioned into multiple logical servers, which is useful for consolidating the functions of separate LAN servers into a single system.

Note that clusters of multiple four-way servers potentially enable similar capabilities, and introduce the additional benefit of high availability. However, most current clustering solutions for NT currently offer at best crude load-balancing capabilities.

*** Transaction performance estimate for Proliant 6000 based on 8.6% increase over Proliant 5000 observed comparing Proliant 5000 and Proliant 6000 systems in transaction tests running SCO UNIX.

**No NT-based TPC results yet available for Compaq Proliant 6000

TPC and tpmC are registered trademarks of the Transaction Processing Council.

EXPANDABILITY    (contents)

NCR’s WorldMark 4300, Compaq’s ProLiant 6000, Dell’s PowerEdge 6100, and Digital’s AlphaServer 4000 offer superior memory capacity, currently supporting up to 4 GB. We expect HP, Digital, and IBM to increase memory support to 4 GB for high-end PC Servers in the near future. NCR’s upgrade of the 4300 to eight processors, promised for 3Q97, will involve deployment of an additional motherboard which can support an extra 4 GB, driving up total memory capacity to 8 GB per system.

Note that support for greater than 4 GB is not truly valuable today. NT 4.0 remains a 32-bit operating system and can thus work with only 4 GB of physical memory. Further, applications in NT can currently access no more than 2 GB of memory, although Microsoft’s forthcoming Enterprise Edition of NT boosts the limit to 3 GB for key applications such as databases. Indeed, NT’s memory limitation may represent a crucial bottleneck preventing it from matching UNIX for high-end performance on benchmarks such as TPC-C. Full 64-bit addressing, which would allow applications to optimally use physical memory greater than 4 GB, awaits Intel’s 64-bit Merced processor or Digital-aided enhancements, which will allow NT to leverage 64-bit addressing in Alpha-based systems. However, Merced will not ship until 1999.

As an intermediate step, Microsoft plans to add some incremental Very Large Memory (VLM) capabilities to NT 5.0, which is expected in early 1998. Pentium Pro supports 36 bits of real addressability, allowing it to support physical memory of up to 64 GB, and Microsoft’s enhancements to NT may involve key applications to exploit Pentium Pro’s larger memories. However, Pentium Pro is still a 32-bit virtual chip. Thus, for applications such as Microsoft’s SQL Server to leverage Pentium Pro’s capabilities, they will have to use multiple address spaces, i.e. data spaces and program spaces, akin to Sun’s technique for exploiting large physical memories with the 32-bit Solaris.

Since NT doesn’t support more than 4 GB now, most vendors do not see the need to ship servers with unaddressable memory, claiming that more memory could become available whenever it became supported. Thus, NCR’s advantage derives primarily from its ability to provide an upgrade path with today’s platforms.

Figure 8: Maximum Slot Availability Comparison

When upgraded with the additional motherboard, NCR’s Worldmark 4300 will earn bragging rights for the greatest number of slots, offering up to 14 PCI and 4 EISA slots (6 PCI and 4 EISA standard). Currently, Digital’s Alpha 4000 leads with 13 PCI and 3 PCI/EISA slots using upgrades. Digital’s Prioris ZX currently includes the largest number of slots overall in standard configurations ¾ 7 PCI, 4 EISA, 1PCI/EISA. Compaq’s ProLiant 6000 offers the greatest number of PCI slots (9) in standard configurations. The remaining systems all have identical slot availability, offering 6 PCI and 4 EISA slots.

Figure 9: Maximum Storage Capacity Comparison

Digital’s Alpha 4000 and NCR’s WorldMark 4300 top the charts in total storage capacity at 15 TB and 14.4 TB respectively. Digital’s Alpha 4000 (pedestal mount) offers a leading 191 GB of internal disk storage. Most of other studied systems offer roughly equivalent internal storage capacity, at approximately 109 GB with only the Digital Prioris ZX and Dell PowerEdge falling short at 63 GB and 54 GB internal storage, respectively. All of the Intel Server vendors offer hot-pluggable internal disks with Compaq offering up to 18 hot-pluggable drives (6 standard), IBM, and NCR offering 12 standard hot-pluggable drives, HP’s LXe offering 6 standard and 6 optional hot-pluggable drives, with Digital and Dell offering 7 and 6 hot-pluggable disk drives respectively.

SYSTEM MANAGEMENT    (contents)

Systems management means different things to different users. Some managers need tools to manage the server, some need to include desktop management, and some also require network device management. This has lead to the varying bundled systems management product offerings ¾ NCR’s Server Manager focuses on server management. Similarly, Dell’s servers are managed with Intel’s LANDesk Server Manager Pro. IBM’s NetFinity and Compaq’s Insight Manager manage servers, desktops, portables and workstations, while Digital’s ServerWORKS and HP NetServer Assistant handle servers, desktops and SNMP network devices.

The focus of this paper is enterprise NT servers which would primarily run mission critical applications. Thus, the features surveyed are limited to server management requirements and do not include broader LAN management issues. However, some of the additional features are listed in Table 2 for informational purposes.

Central to any management is monitoring hardware inventory and configurations information such as firmware version, tracer number, and type of server components. Detailed information about what each system contains, such as type of server components and firmware versions, facilitates faster troubleshooting and financial asset management. Retrieval of manually entered information such as asset tag numbers allows administrators to customize the solution as the site requires. All vendors offer solid capabilities in this area, however, only NCR, IBM, Digital, and Compaq maintain manually entered information.

Event management involves the presentation of system event to the administrator in a timely fashion as well as assist in the troubleshooting process. The management product should monitor performance parameters for critical conditions (e.g. exceeding a user-defined threshold value) then notify administrators of the event by various methods (e.g. email, paging, or color coded icons on a network map). The manager should also allow customizable event filtering so that specialized administrators are notified only if the event is related to their work.

All vendors provide solid support for monitoring performance parameters and thresholds; alerting administrators through various means; and customizing automated corrective actions to common events, however, Dell cannot filter events at the console to customize administrator view of events. Compaq, NCR and IBM also monitor the NT Event Log enabling the administrator to monitor components, drivers, and applications that log information with this NT service. HP handles this with add-on Exposé product.

Remote server management allows centralized managers to perform administrative tasks on remote servers, a key requirement for IT departments with rapidly increasing numbers of resources to be managed. This functionality is available standard from all vendors. IBM and Compaq also ship Web-based interfaces allowing administrative access from any browser.

Emergency access to the server provides management capabilities for off-line systems. This is implemented with a server management board that has its own CPU, power and modem, thus, if the server shuts down due to a fault or loss of power, the administrator can use remote diagnostic tools or reboot the system, a crucial feature for remote management. Only NCR offers an integrated, hot-pluggable management board, other vendors offer the boards as an add-on option. All vendors offer remote reboot capabilities.

Performance tuning involves monitoring system resource activities and using the collected data to identify potential problems and to configure the system to provide optimal performance. The system should store performance data indefinitely, permitting its use for historical analysis of usage trends and fault analysis. All vendors offer similar capabilities in this area, with real-time graphs of performance data and exporting of historical data to spreadsheet or database applications for detailed trend analysis.

Systems cannot be managed in isolation in a distributed environment because the systems must work together for peak performance. Isolated optimization may run counter to larger goals of the computing environment. Thus, specialized systems management tools must be able to share information with enterprise management platforms by forwarding relevant event information, sharing collected inventory and configuration data, and accepting remote commands. All vendors forward event information to enterprise network management systems through SNMP traps. Compaq Insight Manager, Digital ServerWORKS, and HP NetServer Assistant automatically discovers and displays management systems on HP’s OpenView and Tivoli’s TME10 NetView IP maps as well as forwards SNMP traps and shares inventory information. Compaq has similar capabilities with CA, Cabletron, and BMC. NCR is working with CA for tighter integration with CA Unicenter TNG. Dell does not currently have tight integration with enterprise management platforms. Digital and HP also have published API’s that allow other third-party management applications to be launched from their console and share some data.

CLUSTERING SUPPORT    (contents)

Compaq currently offers On-Line Recovery Server which utilizes a combination of hardware and software to allow two ProLiant servers running Windows NT Server to provide back-up processing for each other in the event of a system outage on either system. On-Line Recovery Server supports Oracle and Microsoft SQL Server, but Compaq now focuses on introducing forthcoming Microsoft Wolfpack-based products in the future. IBM currently ships a third party failover product, Vinca Standby Server, and plans to introduce options based on Wolfpack as well as its own Phoenix technology in the future. Dell and HP do not currently offer clustering options for NT, planning instead to offer Microsoft’s standard Wolfpack-based clustering when it is introduced.

But while most NT cluster vendors chose to hesitate rather than innovate in advance of Microsoft’s standard Wolfpack clustering, Digital and NCR charged ahead with LifeKeeper for NT and Digital Clusters for NT, capturing the technology lead for delivering high-availability (HA) NT clustering solutions in advance of standard Wolfpack clustering, which they plan to support as well when it ships. NCR became the first UNIX vendor to focus on NT clustering, shipping a version of its UNIX-based LifeKeeper clustering solution for NT in late 1995. The UNIX version of LifeKeeper offers a very strong feature set, closely pacing the leaders for delivering high-availability UNIX clusters. NCR has tapped some of that technical strength in the NT version, offering a number of clustering features not matched by any of the competitors.

Hoping to extend its clustering leadership from the VAX domain into NT, Digital became involved with NT clusters relatively early on, and delivered the first version of its NT clustering product, Digital Clusters for NT, in mid-1996. The aggressive approach paid off well. Digital has shipped over 3,000 clusters for NT since then, earning it leading market share for NT clusters and far surpassing NCR’s LifeKeeper for NT, which began shipping more than six months earlier. To deal with the impact of commodity Wolfpack products, Digital had hoped to preserve a 9-12 month functional lead over the standard Wolfpack offerings, reconciling compatibility of its "enhanced" flavor with the standard functions when appropriate. Rather than drive its current offering, though, Digital now plans to essentially reset its core technology to match the standard NT clustering functions as soon as Wolfpack ships, migrating its users to the new base. Without any significant added value beyond generic NT clustering functions, Digital may soon become exposed to competitors such as Compaq, which can deliver, in greater volumes, clustering solutions based on the same technology.

NCR’s LifeKeeper for NT enjoys its strongest lead over competitors for clustering hardware features. Of all studied NT clustering solutions, only LifeKeeper supports clusters with more than two nodes, an option that can help reduce performance degradation at times of failure. Moreover, only NCR covers all bases for addressing shared storage. Besides supporting simple shared-SCSI configurations, the most basic technique for enabling application failover, NCR offers both dual-attach RAID subsystems for enhanced reliability, and remote mirroring options for disaster-recovery configurations. Finally, LifeKeeper supports the broadest choice of heartbeat media and network adapter types that can be recovered. Digital Clusters for NT benefits from support for SCSI-based as well as Ethernet-based heartbeat media, but remains limited to Ethernet adapter recovery and lacking support for dual-attach RAID storage.

LifeKeeper for NT keeps a strong lead in highly available services, offering the broadest selection of application "kits," i.e., prepackaged solutions for running off-the-shelf applications in failover configurations on a cluster. Uniquely, LifeKeeper also supports file sharing failover for all of the file systems types available in NT, a modest advantage. Digital Clusters for NT follows LifeKeeper, also offering a good choice of prepackaged application kits, but lacking any tools that allow users to detect failures in their own applications. Digital has decided to delay offering such tools until Microsoft ships its standardized detection and recovery mechanisms as part of its Wolfpack clustering technology.

LifeKeeper retains its lead in cluster operation and management facilities, but Digital Clusters significantly narrows the gap. Both Digital Clusters and LifeKeeper provide basic levels of a single system image with GUIs that can be used to maintain parameters affecting all nodes in the cluster, and all support cluster configuration while on-line. Further, both have configureable thresholds for key parameters such as heartbeat frequency, which helps prevent false failovers.

LifeKeeper’s unique support for three-way clusters provides the greatest flexibility for defining failover and recovery policies, enabling dual node redundancy and the ability to split failed-over loads evenly across available nodes, reducing performance degradation during failures. However, Digital Clusters for NT offers the broadest choice of network protocols to take over, including IPX/SPX in addition to TCP/IP, NetBIOS, and Named Pipes.

BUSINESS FACTORS    (contents)

Based on DHBA’s standard Departmental PCM configuration (see Pricing Configuration Details in Appendix I), Dell’s PowerEdge 6100 leads over all studied competitors in terms of absolute price. Digital’s Alpha-based 4000 is priced substantially higher than the Intel-powered systems, but includes 512 MB RAM. Estimated street prices listed are vendor-provided or reseller-provided. Digital’s 4000 estimated street price derived from hardware list price less 30% (TPC test discount).

While vendors such as Compaq and Dell benefit from aggressive pricing, they fall short on the ability to provide enterprise-grade service and support. Microsoft has defined a number of support requirements deriving from enterprise NT configurations, including:

32-bit Migration Services: the ability to move customer applications from 16-bit Windows 3.1 implementations to 32-bit NT implementations. While NT runs 16-bit programs, they perform far more efficiently in native 32-bit versions.

Custom Application Development: the ability to develop custom NT applications for enterprise users. While NT enjoys a growing base of off-the-shelf "shrink-wrapped" applications, enterprise users typically require auxiliary programs to implement particular solutions.

System Integration & Outsourcing: NT runs on a broad set of hardware platforms, and can be configured with a variety of middleware options. Enterprise users wishing to insure successful solutions often need vendors to take on direct responsibility for putting together and maintaining systems.

Premium Microsoft Product Support: Microsoft designates third parties with a variety of classifications for their ability to provide product support. The highest level is the Authorized Support Center (ASC), which involves taking on significant support responsibilities.

Of the studied vendors, only Digital and NCR meet all the above requirements (see Table 4). HP and IBM each lack particular NT-oriented services, while Compaq and Dell, focused on optimizing for cost and delivery, relegate much of NT’s service and support to third parties such as VARs and ISVs.

Table 4: NT Support Capabilities

Source: Microsoft

* Delivered primarily by third parties

APPENDIX I: PRICING CONFIGURATION DETAILS    (contents)

*Compaq Proliant 6000 TPC estimate based on 8.6% increase over audited Proliant 5000 NT result of 8070 tpmC. The 8.6% increase represents the difference in audited results between a Proliant 5000 and Proliant 6000 running SCO UNIX. With no NT results yet available for the 6000, this represents a reasonable approximation of performance.

**Digital Alpha 4000 5/466 TPC results were run using Digital UNIX OS, IBM’s 704 was tested with Solaris 2.5.1. Since no comparable NT results are available, we did not estimate an NT-based tpm although we believe that the raw performance would not vary dramatically. Generally, NT yields somewhat lower TPC results than UNIX run on comparable systems. Price/performance of UNIX-based systems is consistently higher than NT reflecting generally higher costs for database and operating system software.