Partitioning bonanza: Unix servers

Partitioning, a technology used in mainframe computers, is making its way into mid- and high-end Unix systems where corporate users employ it to isolate and protect applications from each other, combine processing power to run large applications or consolidate processing onto bigger machines.

In the next year, HP, IBM and Sun each will introduce more machines that can be divided into partitions for running different operating systems, applications and workloads. Lightweight Internet applications, such as Web serving, caching or load balancing, could be intermixed with heavier transaction-based applications such as Oracle on the same machine, saving IT the expense of buying two servers – one for each application.

In partitioning, a server’s resources – CPU, memory, I/O, interconnects and buses – are divvied up according to the needs of the applications running on the server. Applications are protected from the actions of other applications that could cause failures, and optimally, they can shift allocated resources on the fly without taking the system down.

In an economy where money for new equipment is becoming scarce, companies are saving by consolidating applications onto fewer larger, more powerful machines. Partitioning helps because it lets users run separate workloads on the same machine.

“Partitions are primarily used for segregating programs, data safeguarding and data recovery. [Without partitioning,] if you have one big partition and any part of it fails, or some of the critical operating system data or configuration becomes corrupt, the whole system is down and recovery is more time-consuming and difficult,” says Dan Gahlinger, senior network engineer and system administrator for Interlynx, an ISP in Hamilton, Ontario.

Gahlinger has a variety of Sun workstations and servers, including Sun’s entry-level Enterprise 450 Server, that are partitioned in what Sun calls Dynamic System Domains, meaning resources can be reallocated to other applications.

Partitioning is used not only to combine operations that formerly ran on different servers, but also to run applications that have become too large to run on one processor.

“We’ll be running PeopleSoft 8 with an Oracle database engine in four partitions – [the partitions] contain Web server, application server, database server and test and development,” says David Meacham, director of IT for Delaware North Companies, a concessions and hospitality company in Buffalo, N.Y.

“We are replacing an HP V-Class Enterprise Server V2250 and an HP K360 server with [HP’s] Superdome,” Meacham says. “We looked at buying several machines, but because of the size of our database, we needed to have 24 processors for PeopleSoft alone. We didn’t feel comfortable bringing in systems where capacity had already hit the ceiling.” Superdome is HP’s high-end PA-RISC based server.

Analysts say one promise of partitioning is its ability to adjust workloads across processors as they change.

“The more that workloads are Internet-driven and harder to predict, the more dynamic they need to be,” says Jean Bozman, research vice president for IDC. “You need to have resources that can be tapped and available, rather than going out and building tremendous data centers with unlimited spare capacity.”

Partitioning evolved from IBM mainframe environments, where it was used to balance the workload of a server and protect applications from harm. Because it was too expensive to buy several mainframes, IBM settled on partitioning as an answer for dividing up the Big Iron so it could run many applications, each protected from the other. IBM’s partitioning was by logical partition, commonly called LPAR. In 1996, Sun introduced physical partitioning with the Sun Enterprise 10K. On the low-end Intel server side, partitioning or software virtualization capability is offered in software from several vendors, including Connectix, Ensim, SW-Soft and VMware.

There are three types of partitioning: physical, logical and virtual.

In physical partitioning – the most common variety – the partitions are divided along hardware boundaries. Each partition might run a different version of the same operating system. Sun and HP servers deploy physical partitioning – the number of partitions relies on the hardware. Physical partitions have the advantage of allowing complete isolation of operations from operations running on other processors, thus ensuring their availability and uptime. Processors, I/O boards, memory and interconnects are not shared, allowing applications that are business-critical or for which there are security concerns.

“One of my clients is running a billing application on an HP AlphaServer GS160 with two partitions – each contains eight processors and 64G bytes of memory,” says Gordon Dixon, a consultant with Cybertech Resources. The GS160 is a midrange 16-processor system, which runs Tru64 Unix and OpenVMS and can be partitioned into as many as four partitions, one for every four CPUs.

“Each partition runs the OpenVMS operating system and is handled as a completely separate system, such that there is not a single point of failure,” Dixon says

The disadvantage of physical partitioning, analysts say, is that machines cannot be divided into as many partitions as those that use logical partitioning, and users can’t consolidate many lightweight applications on one machine.

HP will introduce a version of its SuperDome server midyear that would let different operating systems run in separate partitions. By year-end, the company is expected to launch a 32- and 64-way version of its EV7 processor-based AlphaServer, code-named Marvel. Later this year, the company also plans to introduce a new version of its HP9000, which uses the PA-8800 processor – this will support dual-core technology – doubling Superdome’s processors to 128. Superdome presently supports 16 hardware partitions using nPars, and 64 uniprocessor partitions with an HP partitioning method called virtual partitions.

In logical partitioning, supervisory software overlays the hardware so a machine can be divided along processors, memory, a bus or an I/O slot. IBM and Sun have software partitioning capability – the number of partitions each vendor allows in their servers varies from as few as 16 in IBM’s p690 “Regatta” server to an unlimited quantity in Sun’s newest Sun Fire v1280 or Sun Fire 12K or 15K servers. Servers that use software partitioning also have advantages and disadvantages, analysts say.

Because logical partitioning does not have the electrical isolation of hardware partitioning, it is not immune to failures.

In the first half of 2004, IBM will introduce a 64-processor p690, code-named Armada. Armada will use IBM’s Power5 processor, which deploys a technology called simultaneous multithreading. In simultaneous multithreading, each processor can undertake the actions of two concurrent threads, with full access to system resources, thus making a 64-processor machine look as if it has 128 processors. IBM says with the next version of AIX, Version 5.3, due in the first half of 2004, that improved partitioning capability will let users run as many as 10 operating systems per processor.

HP also plans 128-processor servers using PA-RISC and Itanium processors late this year and next.

Virtual or software partitioning, in which processors arbitrarily divide physical resources, operating systems or time segments, applies the least to Unix machines. Users will deploy it primarily in x86-based servers or in IBM mainframes using Linux. Because it is software-based, complete fault isolation is impossible. Illuminata analyst Gordon Haff says that as hardware becomes more reliable and the need to isolate applications from each other decreases, vendors will focus on letting users partition their servers more finely.