The Incredible Morphing Server

The inextricable march away from dedicated servers toward more flexible approaches has provided data center architects with new ways to satisfy application loads and more economically meet business needs. While most machine-room floors still house some refrigerator-sized proprietary architecture servers, their days are clearly numbered. Blade systems with high-powered CPUs and lightning-fast network interfaces are enabling IT architects to nimbly meet application demands.

The Great Giveaway

IBM's BladeCenter has been influential since its release. While other vendors packed lots of low-powered CPUs into early, comparatively inflexible blade systems, IBM realized that a more flexible approach would meet greater acceptance. The company not only incorporated high-end CPUs into its blade offerings, but it also partnered with the industry's top-tier networking and storage vendors to bring specialized blades to the BladeCenter.

As the rest of the industry rethought its approach to blades, IBM did something perhaps only it could. The company opened the design of the BladeCenter, making its specifications freely available. IBM's stated goal is to remove barriers for vendors wishing to make special-purpose blades for the BladeCenter, but could this also lead to the creation of a new market for BladeCenter clones? We think it might, once again giving IBM a reach beyond its considerable grasp.

CPUs Breathe EasierAs servers morph away from their monolithic roots toward blades, high-speed, low-latency interprocessor communication becomes critical. The industry is currently exploring ways to speed up IP and Ethernet to meet that need. While 10Gbps Ethernet satisfies the speed criterion, the latency and processing overhead of IP and associated higher-level protocols are a problem, often swamping CPUs with protocol processing duties.

Chelsio Communications has shown that both issues can be overcome with its T110-CX TCP Offload Engine (TOE). With a total throughput of almost 8Gbps, Chelsio demonstrates that IP and Ethernet can still provide a very-high-performance interconnect for both iSCSI storage and Remote Direct Memory Access (RDMA) transfers as defined by the IETF. Chelsio was the first to support RDMA and iSCSI at 10Gbps, the first to use the CX copper cabling, and the first to drop end-to-end latency below the magic 10-microsecond threshold over Ethernet. Chelsio has already released its second generation of TOE cards.

A Dual Challenge

While blade servers and TOEs permit packing a big processing wallop into a small space, last year's announcement of dual-core x86 CPUs promises even more profound performance gains. Both Intel and AMD announced that future performance improvements would come mostly by putting more than one CPU core on a chip. AMD demonstrated its first dual-core 64-bit Opteron chip last August, and those chips will show up in servers in the second half of this year.

Although AMD did indeed beat arch rival Intel to the punch in demonstrating a dual-core x86 chip, multicore CPUs have been around for a while. However, AMD's elegant memory management architecture, 64-bit extensions, and superior power management were enough to convince us that its dual-core Opteron is indeed a breakthrough in addressing the needs of the ever-more performance-hungry data center.Editor-in-Chief Art Wittmann can be reached at [email protected].

servers :::