Power-adds to High-end Servers

The once broad line between high-end servers and supercomputers is now diminishing with server applications demanding more and more raw performance from the machine underneath. 64 bit chips from Intel (Itanium) and more recently from AMD (Opetron), supported by 64 bit Windows have started encroaching the space traditionally held by RISC-based CPUs like Sun’s UltraSPARCs and Power4s. Market projections state that Intel architecture servers this year will for the first time generate more revenue than Unix servers, $20 billion, compared with $18 billion.

Multiway processor systems with 4 Itanium processors have become common, and are being pushed in the entry-level of the high-end server market. The players in this field are few but growing with IBM, HP and NEC leading the pack. HP was also involved in the development of Itanium processors with Intel. Going up the multiway system ladder, even the space for RISC-based multi-processor systems (of 32 or more processors) is under attack. HP’s Superdome with 64 Itanium processors heads the TPCC benchmark for non-cluster architectures, and costs a whopping 6.4 million dollars. This extreme end, known as big iron in the industry, has already started making ripples in the enterprise market with HP ready to roll out the 64 as well as 32 way Itanium-based Superdome systems. The applications for these systems could be server consolidation, transaction processing, data warehousing and mainframe alternatives. More importantly, such systems also have the ability to be partitioned, for example, an HP 28 way Superdome running 64-bit Windows on 20, HP-UX on four and Linux on the remaining four processors is possible. Some innovations on part of HP to promote their Itanium systems and partly to protect the interests of their existing PA-RISC customers are also in the pipeline–the MX2 daughter card for one. Projected for mid 2004, this will enable two Itanium processors to fit into a single slot on the motherboard. This will make upgrading to higher way systems (doubling) very easy, allowing scalability up to 128 (64 * 2) processors. The other is the sx1000 chipset, which is being used both in PA RISC, and Itanium systems, allowing users of PA-RISC to shift to Itanium processors with a single cell board swap.

In NEC systems, some of the components in their Itanium servers can be swapped by the customer themselves with no impact on live performance. These components include processors, motherboards and expansion I./O cards. One of the features present that lets the user do hot-swapping is the capability of CPUs to perform in lockstep. This means that all aspects of the system software (OS, drivers, BIOS, etc) are able to keep application threads without any loss of state regardless of the failure of any component. Not a single bit of data is lost this way in case of failure of an individual component.

Itanium supports 32 bit applications but they run slower when compared to running them on the current high-end 32 bit Intel processors. However, AMD’s Opteron processor allows 32-bit applications to run as fast as the 64 bit applications. To count this, Intel has introduced a software to provide 32-bit support, allowing 32 bit applications to run nearly as fast the 64-bit applications if not equally good.

Yet another old horse in the game, SGI has an extreme machine on their product catalogue, namely the Altix 3000 system. This has 64 processors and runs a hardened version of Linux. Considered to be a combination of cluster and multi-way processor systems, it is marketed by SGI as a Supercluster. It uses Intel Itanium 2 processors and is based on the SGI NUMAflex global shared-memory architecture, which is the company’s implementation of a NUMA (Non-Uniform Memory Access) architecture. NUMA is a computer memory architecture, used in multiprocessors, where the memory access time depends on the memory location. The time consideration becomes important as a processor can access its own local memory faster than non-local memory (memory that is local to another processor or shared between processors) given the size of a system having upto 64 processors. The proprietary NUMAflex design enables the CPU, memory, I/O, interconnect, graphics and storage to be packaged into modular components, or bricks. These bricks can then be combined and configured according to the customer’s needs.

Ankit Khare