Hewlett-Packard NonStop Enterprise Server

Summary
NonStop servers, the leading continuously available offering, are the culmination of almost 30 years of development by Tandem Computers, Compaq and now Hewlett-Packard.

– By Ann Katan

Analytical Source: Ed Cowger

Note:
In May 2002, Hewlett-Packard (HP) completed the acquisition of Compaq Computer Corporation and with it acquired the NonStop Himalaya servers. Since that time, HP has upgraded NonStop server technology, reaffirmed the roadmap for transitioning the NonStop servers to the Itanium processor family in 2004 and changed the product name to HP NonStop servers.

Corporate Headquarters
Hewlett-Packard Corporation
3000 Hanover Street
Palo Alto, CA 94304, U.S.A.
Tel: +1 800 533 1333
Fax: +1 800 333 1919
Internet: www.hp.com

Table of Contents

Overview
Analysis
Pricing
Competitors
Strengths
Limitations
Insight

Table 1: HP S-series PMF Units
Table 2: Processor (Speed)
Table 3: Price List: HP NonStop Enterprise Server

Overview
[Return to Table of Contents]

Tandem Computers was the first company to introduce general-purpose commercial fault-tolerant servers to the market. Today's HP NonStop S-series servers are the direct descendents of the original design that Tandem brought to market over 25 years ago, which based on a massively parallel system hardware architecture with no single point of hardware failure, and process and data replication. At its simplest, the hardware components of today's S-series server consist of a single system with two logical processors (each logical processor comprised of two lock-stepped physical processors) and extensive built-in error-checking and fault-isolation features connected through a low-cost, high-bandwidth, low-latency network to each other and other resources such as disk controllers and communications adapters.

Despite generations of changing technology and a vast expansion of its initial product offering into a robust product line, the NonStop Enterprise Division has maintained application compatibility across all its systems. The NonStop servers are targeted toward high-profile e-business, business intelligence, message switching and online transaction processing in the financial (banking, brokerage houses, securities exchanges), telecommunications (including cellular), travel, healthcare, government, U.S. emergency 911 services and retail industries, where reliability and scalability are crucial.

The entire culture of the HP's NonStop Enterprise Division (NED) is geared toward preventing any downtime for its customer's applications. The focus on providing continuous application availability has remained unchanged since the first Tandem computer was introduced. Since then, the NonStop Enterprise Division has continued to ride the technology curve, resulting in today's highly scalable ServerNet interconnect technology that allows data to be transferred directly between any processor and input/output (I/O) device.

The building block of the NonStop S-series server is the Processor MultiFunction (PMF) card. The PMF has memory, dual-ported ServerNet access, Small Computer Systems Interface (SCSI) ports, Ethernet controller, service processor and two processors, with each processor having its own cache. Each PMF runs its own copy of the NonStop Kernel (NSK) operating system. The two processors on each PMF behave as a "logical uniprocessor." They run in lock step, executing the same instruction stream out of their local cache. The output of the processors is continuously compared. If the outputs ever disagree, the operating system will immediately shut down the PMF and record the error. The operating system will then automatically shift the workload to another PMF that contains the backup process.

The loosely coupled, shared-nothing architecture of the NonStop S-series allows system performance to scale virtually linearly as additional resources are added to the system. ServerNet routers route packets of data directly between devices in the system: disk to processor, processor to processor and communications line to processor. This switched network eliminates performance bottlenecks caused by shared resource conflicts. With the ServerNet capability to directly connect data sources and destinations, any processor has direct access to system resources of other processors. The NonStop Kernel operating system permits the processors to operate independently but in a cooperative manner.

The NonStop Kernel microkernel-based operating system has two user/administrative interfaces or personalities: one interface supports the Guardian operating system interfaces used in older NonStop products, and the other interface provides a Unix environment. These interfaces are pure application programming interfaces (APIs), as opposed to any kind of translators, gateways or interpreters. Both "personalities" coexist on the same server.

The NonStop S-series servers have a minimum of two PMF units (four physical processors). Within a single system, PMF units can vary in processor type, amount of memory and type of SCSI adapters. This allows users to cost-effectively configure the server for different applications and protects customer investments.

The newest high-end members of the S-series are the S86000 and S76000 powered by MIPS Technologies R14000 processors. The midrange server is the S7600 based on the MIPS Technologies R12000 processor. Both the high-end and midrange servers can be expanded in increments of two PMFs to a maximum size of 16 PMFs in a single node. With the use of the NonStop Cluster Switch, up to 24 S-series nodes can be connected to create a virtual NonStop system with up to 384 PMFs. A NonStop system can expand to 255 nodes with 16 processors each for a total of 4,080 processors. HP also offers entry-level, nonexpandable, 2-PMF configurations using the MIPS R14000 and R12000 microprocessors. These entry-level systems can be upgraded to larger servers with a higher license charge.

[Return to List of Tables]

Table 1
HP S-series PMF Units

Models	S86000	S76000	S7600	S76	S760
Processor Type/(MHz)	MIPS R14000 (550)	MIPS R14000 (500)	MIPS R12000 (300)	MIPS R14000 (550)	MIPS R12000
Cache Size (Bytes)	8M	4M	2M	8M	2M
Memory Sizes (Bytes)	1G, 2G, 4G, 16G	1G, 2G, 4G, 16G	1G	1G, 2G, 4G, 16G	1G
SCSI Type	UltraSCSI	UltraSCSI	UltraSCSI	UltraSCSI	UltraSCSI
Date Delivered	June 2002	June 2002	October 2002	June 2002	June 2002
Entry Selling Price	$76,000	$60,000	$32,000	$60,000	$32,000

Future Direction
The real news for NonStop systems is the impending conversion from MIPS-based processors to the Intel Itanium processor family. The current plans of the NonStop Enterprise Division are to allow complex instruction set computer (CISC)-based application object modules to execute in a "compatibility-mode" environment and to provide source-level compatibility for reduced instruction set (RISC)-based applications. The NonStop Enterprise Division plans to provide several other initiatives to protect investments that its customers have already made:

• Since the Itanium microprocessor uses a different instruction set at the hardware level, it requires new versions of native-mode compilers and tools to create native application code files. Nonprivileged applications running on MIPS-based NonStop servers will run on the new Itanium-based NonStop servers following a recompile.

• CISC programs on Itanium-based NonStop servers employ the same source code and object code as on MIPS-based and earlier CISC-based NonStop servers. These code files can be executed directly in the Itanium environment with the CISC Object Code Interpreter for Itanium (OCI/I). Thus CISC-based application object code can be run directly on Itanium. However, applications that use the Open System Services (OSS) API must be recompiled with current native-mode compilers and OSS compiler tools.
• Itanium-based NonStop servers will use the same high-speed low-latency system interconnect architecture and NonStop Kernel operating system.

• NonStop S-series I/O cabinets and peripherals can be connected to Itanium-based NonStop servers.

• Itanium-based NonStop servers will support industry-standard disks (current MIPS-based S-series models have a proprietary disk interface).

• MIPS-based NonStop servers can interoperate with Itanium-based NonStop servers by belonging to the same ServerNet Cluster or Expand network. Individual nodes must contain either all MIPS-based PMFs or all Itanium-based PMFs. Investments in hardware, software and training will be preserved and optimally leveraged.

The major changes for application developers include:

• Build scripts will have to be redone since there are different names for the toolsets.

• On Itanium-based NonStop servers, all standard or customer-added native-mode libraries will be packaged as dynamic link libraries (DLLs) rather than as shared runtime libraries (SRLs) or components of the NonStop Kernel operating system's image that have had system generation (SYSGEN) applied.

• There are no cross compilers for MIPS engines – applications will have to be recompiled on a native Itanium platform.

• Most privileged code is subject to significant source-level changes to run on Itanium-based NonStop servers.

• All native COBOL, C, C++ and pTAL programs must be recompiled for the Itanium microprocessor from the source code.

Beyond the initial introduction of the NonStop/Itanium systems, both MIPS-based and Itanium-based NonStop servers will have concurrent releases of the NonStop Kernel operating system. However, some new functionality, such as 64-bit addressing, will only be implemented for the Itanium-based NonStop servers.

Analysis
[Return to Table of Contents]

To guard against failures caused by corrupted or inadvertently changed data, NonStop servers employ a variety of methods to ensure data integrity. The "logical uniprocessor" use of lock-stepped dual processors in the S-series modules provides a checking mechanism to ensure that a malfunctioning processor cannot corrupt any data. This feature has become more important over time, as increasing chip densities result in increasing frequencies of transient errors. In recent years, other vendors of high-end systems have moved to a lock-stepped approach in recognition of this problem, including IBM for its zSeries systems.

In addition, data is checked and verified whenever it is moved. Data on all buses is parity-protected. Parity checks, packet sequence numbers and checksums protect I/O controllers. Before data is written to disk, a checksum is calculated that is appended to the disk data. When data is read back, the checksum is recalculated and compared to the checksum that was stored on disk. If the checksums match, the data has not been altered. Of course, if the data has been altered, there is a mirrored disk drive from which the information can be retrieved.

ServerNet itself is capable of providing a fault-tolerant communications network. Since there are multiple paths to all the components of the servers, and since stored data is duplicated, no single failure can block communications between a process and its data. All processors, I/O devices and communications adapters are monitored by the service management subsystem, which detects and isolates failures. If a component fails, data is routed over a different path to the same destination. Each packet passed through the ServerNet network includes cyclic redundancy checks to ensure its integrity. Each ServerNet 2 link runs at 1.25G bps, full-duplex, for an aggregate router bandwidth of 1.5GB/second. ServerNet uses wormhole routing where an incoming packet is forwarded as soon as the destination address in its header is decoded. (In more conventional networks, a packet is not routed until it has been received in its entirety.) As a result, ServerNet can achieve latencies of only 300 nanoseconds in each router. ServerNet II delivers five times the throughput of the previous generation of ServerNet routers.

The NonStop environment provides the capability for any process to run on at least two different PMFs. Using this capability of the NonStop platform, the operating system implements the concept of "process pairs." A process pair consists of a primary process and a backup process that run on separate PMFs. The primary process sends checkpoint messages to its backup process. These messages ensure that the backup process has all of the process state information that it needs to take over the processing of the application. Thus, a NonStop server can provide a fault-tolerant platform where users can continue interacting with an application and will continue to be serviced without any noticeable interruptions, despite the failure of any component, including the system software. In contrast, high-availability clustering solutions ensure that a server is always available to run an application, but the application itself will have to be restarted. Restarting an application can cause delays in service to users.

The newest PMF used in the S86000, featuring the 550MHz MIPS R14000 processor combined with other hardware and software improvements, can provide a performance increase of up to 1.9 times over the previous high-end generation processor used in the S74000. The 500MHz R14000 processor used in the S76000 with its smaller cache size can provide a 1.5 times performance increase over the S74000. Despite these performance improvements, the MIPS R14000 processor significantly trails the performance of the leading Intel and RISC processors available from the major system vendors, as can be seen in the table below. [Note that the CINT benchmarks in the table below are an indication of individual processor performance only and are not a measure of system throughput.]

[Return to List of Tables] Table 2
Processor (Speed)

Benchmark	CINT2000	CFP2000
Compaq Alpha 21364C (1150MHz)	877	1482
HP PA-RISC 8700 (750MHz)	549	522
IBM POWER 4 (1,300MHz)	848	1281
MIPS R14000 (500MHz)	412	463
Sun UltraSPARC III (1050MHz)	610	817
Intel Pentium 4 (3,060MHz)	1107	1091
Intel Itanium 2 (1,000MHz)	Information not available	1431

The retrofit of the HP NonStop servers with Intel Itanium processors will provide a significant performance boost for NonStop customers. However, there will be steps that users will have to perform to upgrade to the Itanium-based servers. The Itanium-based S-series will be able to execute original CISC-based application binaries at roughly the same speed as they execute on today's MIPS-based processors. However, there will be dramatic performance improvements if these applications are recompiled on the Itanium platform from source code. MIPS-based binaries must be recompiled from source code since there will be no MIPS interpreter on Itanium. MIPS object code will not run in the Itanium environment.

HP has committed to providing the necessary compilers and other porting tools for this migration. The same languages will be supported on Itanium as are supported on MIPS. S-series infrastructure investments in storage and I/O adapters will be preserved during this transition, since the Itanium-based NonStop servers will continue to use ServerNet as the system communications backbone.

The NonStop Enterprise Division continues to provide new software interfaces and features to ensure that its servers support the latest computing paradigms. NonStop servers support many industry-standard development tools and application environments including Java, Java Database Connectivity (JDBC), Common Object Request Broker Architecture (CORBA), BEA Tuxedo and WebLogic Server (WLS), Extensible Markup Language (XML), Simple Object Access Protocol (SOAP) and Open Database Connectivity (ODBC). Several enterprise management frameworks such as CA-Unicenter, BMC Patrol and HP OpenView are also supported.

Pricing
[Return to Table of Contents]

[Return to List of Tables] Table 3
Price List: HP NonStop Enterprise Server

Component	Description	Estimated Base Selling Price
S86000 PMF	1GB memory 2GB memory 4GB memory 16GB memory	$76,000 $82,000 $100,000 $228,000
S76000 PMF	1GB memory 2GB memory 4GB memory 16GB memory	$60,000 $66,000 $84,000 $212,000
Power Shelf	Power for two PMFs	$4,000
S-series Chassis	Holds two PMFs	$24,200
Modular ServerNet Expansion Board	Extends ServerNet fabric across enclosures	$7,500
Disk Drives	18GB at 15,000 revolutions per minute (rpm) 36GB at 10,000 rpm	$2,500 $3,000

GSA Pricing
Yes.

Competitors
[Return to Table of Contents]

• IBM zSeries with Parallel Sysplex. HP targets data-intensive applications in addition to those requiring continuous operations. In terms of reliability and scalability, very few server platforms can match the capabilities of the NonStop servers. Parallel Sysplex, IBM's mainframe clustering approach, is emerging as the standard for enterprise-class, scalable servers for applications that cannot tolerate any downtime. However, Parallel Sysplex must be built and customized for continuous availability at each installation. This can be complex, due to the large number of options and parameters that must be adjusted for specific application environments. The NonStop servers are designed from the ground up for continuous operation without modification. The IBM zSeries is the primary competitor of the NonStop servers.
• Unix Systems Suppliers Sun Microsystems and IBM. These vendors offer clustering solutions for high availability instead of full fault tolerance. They are much less expensive to acquire than NonStop systems. However, the Unix clustering products introduce new levels of complexity in terms of day-to-day operations, thus adding to their total cost of ownership (TCO). More to the point, Unix vendors market aggressively in the decision support and online transaction processing (OLTP) applications markets that HP is targeting with the NonStop line. But even with clustering, Unix platforms cannot achieve the levels of application availability and scalability possible with NonStop.
• NCR WorldMark. With its Teradata database, NCR is a leading vendor in the use of massively parallel systems for data-intensive, decision support applications.

Strengths
[Return to Table of Contents]

Hardware/Software-Based Fault Tolerance
HP's hardware/software-based fault tolerance can be viewed as either a strength or a limitation, depending on IT management philosophy and application needs. NonStop provides total data integrity and continuous application availability through a unique combination of hardware and software features. The design ensures that there is no single point of failure, and the system automatically and transparently recovers from most failures.

System Scalability
System expandability is another result of the parallel architecture of the NonStop servers. HP's ServerNet architecture enables highly scalable processing and I/O. The company claims less than 2 percent degradation from totally linear performance improvement as additional processors are added, up to the 100- to 200-processor range.

Long Experience Building Parallel Systems
The NonStop Division, since its inception almost 30 years ago, has used a hardware/software approach to building parallel systems for high-end performance and high availability. This design is extremely well suited to today's emerging customer relationship and decision support applications that need continuous and rapid access to massive amounts of data. The entire culture of the NonStop Enterprise Division centers around assisting its customers to prevent any type of downtime, not to just fixing problems when they occur.

Database "Mixed Workload" Capability
Because the NonStop SQL database is tightly integrated with the software and shared-nothing hardware, it inherits the scalability and availability of the platform without the usual problems of cache coherency associated with merchant databases. In addition, it has the capability of allowing simultaneous update and querying of the same data without compromise to either workload – a feature NED calls "mixed workload."

Capacity On Demand
HP's capacity-on-demand program for the NonStop servers includes processors, memory and disks.

Limitations
[Return to Table of Contents]

Limited Application Set
The HP NonStop servers are a niche product in a high-profile but relatively small market. The NonStop Division supplies the hardware, operating system and application software for high-volume OLTP environments in which it has many years of experience and a strong, clearly defined offering. More recently, NSD has supplied a Java development and deployment environment and support for emerging standards to support Web services for its NonStop servers. However, most of the major independent software vendors (ISVs) are not developing or porting software for NonStop systems.

Proprietary Disk Interface
MIPS-based S-series NonStop servers have a modified interface to disk drives that does not allow the use of industry-standard storage devices. All such disks must be acquired from the NonStop division at higher prices than what would be possible in a competitive bidding situation. NED plans to eliminate this restriction on MIPS-based S-series servers prior to the introduction of the Itanium models (which will support industry-standard storage devices).

Future Processor Technology
In retrospect, the NonStop Enterprise Division made a poor decision when it chose to migrate its proprietary CISC technology to MIPS microprocessors. Silicon Graphics Inc. (SGI), the previous parent of MIPS, is basing new products on Intel processors, and MIPS Technologies, now an independent company, is focusing on embedded applications. While a transition to a new processor architecture involves some difficulties, the NonStop Enterprise Division understands the tools that will be required to make the transition as painless as possible, having already made the transition to a new processor architecture in the past.

Insight
[Return to Table of Contents]

HP continues to maintain the NonStop Enterprise Division's image as the leading fault-tolerant provider and has delivered enhanced products to satisfy its very loyal customer base. With HP's backing, the NonStop Enterprise Division should continue to enjoy its lead in mission-critical fault tolerance – even as it changes processor technology. The issue here is how much pain and suffering its customers will incur in the migration to the Intel Itanium processor family.

High-end, clustered Unix systems with partitioning or resource management software, robust database management systems and transaction processing (TP) monitors are becoming more reliable every year and today provide high-availability alternatives to NonStop systems. The same goes for IBM's zSeries and other S/390-compatable mainframes with Parallel Sysplex.

It should be noted that when the Unix server vendors offer availability guarantees, even at 99.999 percent, the guarantee is only for an operating system prompt. Restarting an application on another node in a cluster can require a substantial amount of time. The NonStop server customer must decide whether the additional costs are worth the reduction in user-perceived downtime. We should also point out that while the acquisition cost of a NonStop server may be higher than the alternatives, the TCO may be more in line with clustered high-availability solutions, which are difficult and expensive to implement and manage. The NonStop server advantage is that it has been designed from the ground up for continuous availability.

Gartner Datpro Document DPRO-90075, 27 February 2003.