As Intel prepares to introduce its next-generation Nehalem-EX processor, IBM announced a new server architecture that is designed to take full advantage of the new processor's capabilities. IBM eX5 is the fifth generation of the Enterprise X-Architecture, its x86 server design that is optimized for hosting workloads in demanding enterprise environments. IBM revealed two new rack-mounted servers based on eX5, the x3690 X5 with two sockets and the x3850 X5 with four sockets. Also, for the first time, IBM is extending the Enterprise X-Architecture to a blade form factor with the BladeCenter HX5.
IBM is one of the few vendors of x86 servers to perform its own engineering of the major system components. While most other x86 server suppliers use Intel's standard chip sets as the building blocks for complete server systems, IBM has differentiated its x86 servers with unique technology since acquiring Sequent in the late 1990's. The first few generations of this technology, which IBM brands as Enterprise X-Architecture, offered moderate differentiation, at least enough for IBM to separate itself from the commodity orientation of most other x86 suppliers. The architecture hit its stride with eX4 in 2007, which introduced some unique reliability and scalability features, including the ability to protect against memory failure at three different levels, which can dramatically boost the uptime of servers.
These features have allowed IBM to target considerably more critical workloads with its x86 servers than platforms from other vendors. While eX5 still has more reliability features than lower-end Intel Xeon platforms, many of its earlier reliability features have been incorporated into the base Intel architecture. Moreover, scalability and reliability is now commonly achieved with clustering software rather than with hardware features. While IBM emphasized its ability to deliver large systems with earlier generations of the Enterprise X-Architecture, its sales are now increasingly skewed towards 1- and 2-socket boxes, which are the de-facto industry standard for deploying virtual machines.
Memory is one of the most precious resources for a virtual machine host, and after extensive discussions with customers, IBM found that IT managers now consider the memory capacity of a server to be more important than almost any other feature. Many servers utilize only a fraction of their CPU power, and when consolidated in virtual machines, large numbers of VMs can be accommodated by the computational power available in a single CPU socket on the host, especially as the number of cores in that socket increases. However, each individual VM still requires the same amount of working memory that it did as a physical machine, since every workload will have its own instance of an operating system and application software (although some of this can be reduced in software by transparently sharing identical memory pages between VMs, or compressing them during inactivity). As a result, modern servers have to be rebalanced with far larger memory ranges per processor than previous designs. IBM found that customers were willing to pay up to $50,000 or more per server to get enough memory for virtualization. That investment is considered small relative to the hardware cost savings that can be achieved by consolidating multiple physical servers into virtual machines on a single system.
A key feature of eX5 is therefore its MAX5 snap-on memory expansion unit ("memory drawer") that enables up to double the memory capacity of the standard Intel-supplied chip set for Nehalem-EX that most other x86 systems suppliers will use in their servers. Compared to server designs with smaller memory ranges, eX5 could deliver significant economic benefits in terms of virtualization software costs, which is often licensed per socket. Since virtualization deployments are typically memory bound, rather than processor bound, larger memory ranges mean that more virtual machines can be deployed per socket. The superior memory ranges in eX5 will also be valuable for customers who want to deploy other classes of workloads on x86 servers, such as very large databases.
IBM originally started working on scalable and reliable memory management in the X-Architecture with the intent of supporting large scale-up SMP servers for monolithic workloads. These investments are now ready to pay off with IBM's Nehalem-EX systems, which may allow as many as 128 cores (the maximum supported by the eX5 architecture) to be harnessed in a single server footprint. Even on individual 2-socket servers, the larger memory range granted by MAX5 drawers promises to deliver significantly better value due to virtualization software costs relative to other Nehalem-EX –based platforms (pricing for the new systems will not be announced until Nehalem-EX itself is announced). IBM's implementation of MAX5 relies on the standard scalability port built into Nehalem-EX, but it also uses some IBM eX5 technology for linking computing modules. While competitors can duplicate functions that use Intel's standard scalability ports, the other functions are unique to IBM, so it remains to be seen whether any other x86 platform will be able to quickly match the memory ranges of eX5. If not, IBM will be offering customers a unique cost advantage at a time when virtualization is becoming one of the central business priorities for IT managers.
Comments