New breed of servers deliver energy savings

Energy costs and infrastructure complexity are rising, right along with the pressure to go “green”. These factors are contributing to higher operating costs—while computing requirements continue to grow unabated—which favor technologies that can rein in costs and simplify server management. Moreover, many IT organizations will opt for solutions that improve the cost-efficiency and performance of their existing systems.



Organizations striving for greater data center efficiency are evaluating options to achieve the following infrastructure goals: first, the transition to “greener” computing—companies are asking IT departments to dramatically raise data center energy efficiency and reduce power and cooling costs, and second, getting more from strategic technologies—the latest advances in server clustering, software parallelization and virtualization are creating more value to data centers, further lowering cost and increasing performance without “forklift” changes.



COOLING SOLUTIONS, POWERFUL PROCESSORS

Some IT facilities are beginning to run out of cooling capacity and power because air conditioning systems are maxed out and power distribution infrastructure is completely utilized. Combining hardware and software strategies, data center cooling solutions use various approaches such as optimizing airflow, scheduling server workloads off-peak and sharing power supplies between servers, with each providing incremental improvements.



One technique is to position server enclosures back-to-back, so exhaust air is carried by a dedicated channel to the cooling units. Since newly cooled air and hot exhaust air never mix, the cold air remains effective as it enters the enclosures. Using similar innovative methods, HP’s Thermal Zone Mapping and Dynamic Smart Cooling solution can decrease energy costs by up to 45 percent by delivering cooling where and when it is needed the most.



Responding to pressures to improve performance per watt, CPU vendors have turned to multicore processor architecture. By running more cores at lower voltage and frequency, these processors are delivering greater computing performance, within a similar thermal envelope, than their predecessors. Launched in mid-2006, the Dual-Core Intel Itanium processor has twice the performance of previous, single-core Itanium processors, yet consumes 20 percent less power. In late 2008, Intel is expected to release the world’s first 2-billion transistor microprocessor, codenamed Tukwila, which will be a quad-core Itanium processor and will approximately double performance with only a 25 percent increase in energy consumption, for the planned 130W SKU offering.



Processors and operating systems are also working together to save energy during periods of lower data center demand, allowing the platform to run slower and consume less power. For example, the new Dual-Core Intel Itanium processor 9100 series released in late 2007 supports a demand-based switching feature on Microsoft Windows Server 2008 which throttles down the processor’s clock speed and voltage during idle periods in order to reduce power consumption.



Contrary to skeptics who think DC power is more marketing hype than reality, many industry leaders, including Intel, HP and Sun Microsystems, are advocates of this approach. Roughly 50 percent of the power delivered from a wall socket to a PC never actually performs any work, according to Urs Hölzle, Google fellow and senior vice president of operations. Half the energy gets converted to heat or is dissipated in some other manner in the AC-to-DC conversion. Supporting this claim, telecom operators widely deploy -48V DC power distribution in central offices, where it reduces power losses by 40 percent to 50 percent. The contrast between AC and DC power efficiency was also highlighted in a recent Intel study. It showed the combination of 380V DC power distribution and best-in-class components could achieve approximately 75 percent efficiency; this is a significant improvement over the 50 percent power efficiency of a typical data center. While converting existing data centers to DC power can be prohibitively expensive, it is an option worthy of consideration for new facilities.



ENRICHING STRATEGIC TECHNOLOGIES

A number of proven strategic technologies, such as server clustering, parallel processing software, and virtualization, are moving forward with new improvements and enhancements. Server cluster management software is becoming more user-friendly, paving the way for more companies to deploy clusters. Some independent software vendors are rewriting applications to take advantage of parallel computing platforms and increasing performance density. And virtualization solutions are adopting I/O acceleration technologies and dynamic provisioning that will speed up packet-processing applications and improve server utilization. An increasing number of companies are investing in server clustering to enhance system availability and paying for this investment with savings from their server consolidation programs (using virtualization). Some of the money IT saved on hardware, software licenses and utilities is being reinvested to boost system availability. Companies are adding more server redundancy and failover capabilities, which prevent a virtualized server failure from bringing down multiple applications at one time.



Before virtualization, it was not uncommon to have an application running on a unique server configuration. And it was often too costly to implement high availability strategies for these specialized systems since IT had to buy a duplicate system for failover. Server clusters, supporting virtualized applications on standards-based servers, provide an affordable path to consolidation with the added benefit of high availability.



Server cluster management can be very complex. It typically requires IT experts with an intimate knowledge of all the system components (hardware, software and interfaces) to setup failover clusters that correctly comprehend the system, storage and network configurations. Microsoft Windows Server 2008 features enhancements aimed at simplifying clusters, making them more secure and enhancing cluster stability. One enhancement is a “Validate Tool” that analyzes cluster subsystems, such as server nodes, networks and storage, and determines whether they can be used as redundant resources. Windows Server 2008 for Itanium-based systems supports 64-way symmetric multiprocessing and one terabyte of memory, and runs native 64-bit applications.



Software developers are looking for new ways to scale software performance and take advantage of the increasing number of CPU cores in high-performance computing platforms. Unfortunately, writing efficient multithreaded code is hard, and it’s not generally taught in schools. However, there are a number of software techniques and technologies that can unleash the performance of parallel computing platforms without requiring painful code-threading practices. Applications which process packets, such as virus detection, benefit from “flow-pinning”, which directs all the packets of a message to a single CPU core. This easy-to-use technique keeps all the cores busy and optimizes CPU cache memory utilization. Intel found the performance improvement associated with flow-pinning, using an open-source intrusion detection software called Snort, was over six times greater than distributing the workload symmetrically (round robin).



Parallelism is also inherent to service-oriented architecture (SOA), which breaks up a business application into a collection of services. Using an intelligent scheduler, these services can be dispatched across multiple cores so they run simultaneously, greatly decreasing execution time.



Check Point Software Technologies recently released a technology called CoreXL—for their VPN-1 integrated firewall, VPN and intrusion prevention solution—developed specifically for parallel processing platforms. Its load balancing technology prevents underutilized cores and sub-optimal performance. The result is a 600 percent throughput increase when CoreXL is activated.



VIRTUALIZATION ADDRESSES I/O AND DYNAMIC PROVISIONING

Many server applications are I/O intensive, especially networking and storage. When these applications run in a virtualized environment, they are either assigned a dedicated Ethernet network adapter card (NIC) or they share a NIC which requires special support from the virtual machine monitor (VMM). The VMM must perform address translation and housekeeping for interrupts and memory access (DMA), all of which can significantly increase latency.



Helping to relieve the burden on VMMs, the PCI-SIG defined two extensions that allow multiple operating systems to share PCI Express devices with less VMM intervention. Future processors and chipsets will perform even more of these time-consuming tasks in hardware, which will make I/O virtualization faster and more efficient in the future. This functionality is expected to be available on the next-generation Itanium processor, codenamed Tukwila, toward the end of 2008 with the introduction of Intel Virtualization Technology for Directed I/O (Intel VT-d).



Some virtualization environments are doing away with virtual machines (VM) altogether. Instead of running applications within VMs, each with its own operating system, some systems are implementing a shared services model. It runs many applications on one OS while a software scheduler provisions system resources (e.g. CPU cycles and memory) to their various software applications based on their workload and priority. Using dynamic provisioning, IT can rapidly meet the needs of the business and satisfy service level agreements (SLAs). The shared services model supports consolidation without the complexity of a VMM controlling multiple VMs, but it does limit the platform to one operating system.



Also offering VM-less virtualization, NEC Express5800/1320 Series Servers allow administrators to divide a multiprocessor system into several physically distinct servers, all running their own instance of Windows Server 2008. This platform, equipped with Intel Itanium processors, implements “dynamic hardware partitioning,” which permits the administrator to set thresholds associated with processor utilization for each server. When utilization levels exceed the threshold, more processors and memory (idle and unallocated) are automatically added to the partition per predetermined policies. There are also thresholds for processor and memory faults which initiate the transfer of processing from failing partitions to stable partitions without rebooting the operating system or restarting the application.



Dynamic provisioning can also be controlled at the virtual machine level. Operation management software from Fujitsu allows each VM to expand or contract dynamically and respond to real-time fluctuations in demand. This approach optimizes the performance of each VM by allocating small increments of system resources, such as CPU, memory and I/O devices. This level of granularity is particularly important to systems supporting a large number of VMs, like the Fujitsu PRIMEQUEST virtual machine function with up to 60 VMs, each running a dedicated operating system (Red Hat Enterprise Linux 5 and Windows). The system uses up to 32 high-performance 64-bit Dual-Core Intel Itanium processors, with a Fujitsu-developed chipset that maximizes the reliability and scalability of the processor.



Environmental concerns and global competition are forcing CIOs to execute strategies that deliver operational efficiency and look for ways to get more from less. Fortunately, some incumbent technologies can still deliver greater savings and performance without adding undue complexity.

Click here for the illustration:

Figure 1