Cover Story

Servers that sip power

The power consumed by a server over its lifetime is likely to cost as much as the server itself if energy costs continue to rise. This has led to the development of energy-efficient multi-core server processors that reduce the operational cost of a data centre. Akhtar Pasha reports.

As servers and other data centre equipment get even more densely packed, power and cooling problems are worsened. Servers have become the biggest issue as server farms are huge in relation to everything else. “Within the server, increasing power budgets for processors is one of the biggest problems,” points out Daya Prakash, Manager, IT, LG (CNS Global). For IT managers building centres with thousand of servers, the performance-per-watt criterion is critical. In the recent past, the numbers of servers is increasing every quarter be it for new applications (ERP), high availability or disaster recovery and business continuity. This has a direct impact on energy consumption per square foot in a data centre, and has noticeable impact on a data centre’s TCO. The latest servers bought by LG have been chosen keeping energy efficiency in mind. This is why businesses are looking at multi-core processor for their data centres.

"Because a x86 server consumes between 30-40 percent of its maximum power when it’s idle, running systems with very light workloads wastes power" - Richard Fichera Vice-President Forrester Research Inch

In a typical data centre, every watt of power consumed by IT equipment requires another watt of power for overhead, including losses in power distribution, cooling and lighting. Depending on efficiency, this burden typically ranges from 1.8 to 2.5 times. Assuming only a 1:1 ratio, a 3 MW data centre will require 6 MW to operate. Says Richard Fichera, Vice-president, Forrester Research Inc, “Looking at the energy budget for the entire data centre, users need to factor in the energy for cooling systems as well as the power for the equipment. In theory, the best cooling systems require about 0.3 watts to cool 1 watt of equipment, but real-world results from actual data centres, including all parts of the cooling cycle, run between 0.5 and 1 watt of cooling power for each watt of dissipation.” With data centre electricity costs averaging approximately $0.11 per kilowatt hour (KWH) in North America, the costs of running a 1,000-server computer room are becoming significant, which makes energy-efficient servers that much more attractive.

Comments Rajesh Dhar, Country Manager, ISS, Technology Solutions Group, HP India Sales, “Energy will not be cheaper anytime. At Rs 8 per KWH, a 42-U rack server populated with 42 servers will yield only 60 to 70 percent utilisation, and it will consume 5-8 kilowatt per rack of power—that’s a huge cost if you have a data centre with hundreds of servers.”

Adds Fichera, “Optimising cooling systems will considerably improve the energy efficiency of data centres. Using energy-efficient servers is just another way to bring down the running cost of a data centre. Because a typical x86 server consumes between 30 and 40 percent of its maximum power when it’s idle, running systems with very light workloads wastes power. Thus, increasing the average utilisation of the servers can yield significant benefits in overall operational efficiency.”

Multi-core processors

According to Dhar, “With x86 servers sporting multi-core, multi-threaded processors hitting the market, they offer the best price-to-performance because of which we have seen large data centres that traditionally bought Unix servers replacing them with x86 servers. The best performance per watt per dollar offered by multi-core processors is getting them into the mid-market which was earlier serviced by Unix servers.”

Texas Instruments has the challenging requirement of shortening its chip design cycle time and reducing simulation time every quarter. This results in more servers being added, which in turn is leading the company to look for alternatives which will allow them to speed up the design cycle time without increasing the power requirements. Additionally, as per some analysts, the number of servers in data centres is still increasing annually by 18 percent. Also, when users migrate applications from several smaller boxes onto larger machines capable of running multiple applications, the freed-up servers were often redeployed rather than being retired, thus increasing the number of servers in the data centre.

If we look at current transistor technology, it limits the ability to continue making single-core processors more powerful. As a transistor gets smaller, the gate, which switches the electricity on and off, gets thinner and less able to block the flow of electrons. Thus, small transistors tend to use electricity all the time, even when they aren’t switching, and hence waste power. Also, increasing clock speeds causes transistors to switch faster and therefore generate more heat and consume more power. These challenges have hurt manufacturers’ plan for faster single-core processors.

Different approaches

"The communication rack is built directly into the chip, meaning internal communication tasks have a shorter distance to travel across metal. This speeds up performance and reduces energy consumption" Arnab Roy National Sales Manager Data Centre Practice Sun Microsystems India

Simply making current architectures more power-efficient can yield differences in power efficiency up to 50 percent. In addition to those efforts, CPU architects are exploring different approaches to chip architectures. One example is Sun’s UltraSPARC T1 processor, which combines multiple small cores, multiple threads, and a lower-speed power-saving process to yield power efficiencies. The UltraSPARC T1 contains up to eight cores, and each core is capable of executing four threads. The processor can handle 32 threads simultaneously. Notes Arnab Roy, National Sales Manager, Data Centre Practice, Sun Microsystems India, “Memory latency is by far the worst problem, causing a typical server CPU to be idle for 75 percent of the time. The UltraSPARC T1 is the first microprocessor design to build four memory controllers, transferring data between memory and the processing cores to ensure that data is transmitted into the chip as fast as it can be processed.” The communication rack is also built directly into the chip, meaning internal communication tasks have a shorter distance to travel across metal. This speeds up performance and reduces energy consumption, the latter because it reduces the amount of memory and processing latency that has been typical in most chip designs in the past.

According to Sun, the UltraSPARC T1 typically consumes 73 W of electrical power. Roy says, “We have sold over 130 systems based on UltraSPARC T1 processors…it has been an instant hit.” The company is attempting to capitalise on the system’s power attributes by introducing a new benchmark, SWaP (space, wattage and performance), that focusses on a data centre’s macro-environment. Businesses can calculate the running cost of a data centre on Sun’s Web site, and find out the amount of power used to cool their data centre.

"There will be a 60 percent improvement in performance per watt in the quad-core AMD Opteron based on Rev F server" - Mukund Ramaratnam Director Marketing & Business Development AMD Far East (India)

AMD’s next-generation processor for servers, the dual-core Opteron, will have a new processor, the Rev F, which is from the current crop of 90-nanometer Opteron processors and supports Pacifica X86/X64 instruction set virtualisation electronics. These chips will also have unspecified improvements in AMD’s PowerNow thermal management features; the features de-activate parts of chips when they are not being used, which cuts down on electricity use and heat dissipation. Reveals Mukund Ramaratnam, Director, Marketing & Business Development, AMD Far East (India), “To get to quad-core processors, AMD will be moving to a 65-nanometer process in 2007.” The dual-core to quad-core transition strategy is similar to the approach the chipmaker took with its Socket 940, which accommodated both single-core and dual-core Opterons. Promises Ramaratnam: “There will be a 60 percent improvement in performance per watt in the quad-core AMD Opteron based on Rev F server.”

For some time, Intel has been working to deliver new levels of energy-efficiency through silicon, processor, platform and software innovation. The results of these efforts are now clearly evident in the new generation of servers based on the dual-core Intel Xeon processor 5100 series called Woodcrest. “These Woodcrest servers boost performance by 80 percent and energy-efficiency by 35 percent compared to the latest single-core Intel Xeon 2.8 GHz processor-based servers,” says Narendra Bhandari, Director, Asia Pacific, SSG Global Developer Relations Division, Intel Technology India. He explains that the current Xeon processor consumes 130 watts of power and is based on the Dempsey platform, while the Woodcrest will consume only 80 watts at the chip level. “Currently there are 10 customers (including ISVs) in India who are doing trials and testing using some parts of Woodcrest inside their data centre.” The commercial availability of Woodcrest is expected shortly, as reported by Intel.

IBM Power6 is a 65-nm processor and operates at a clock speed of 4.2 GHz, says Jyothi Satyanathan, Country Manager, eServer pSeries, IBM Systems & Technology. Built around silicon-on-insulator, the Power6 is the follow-on processor to the company’s current Power5 architecture. As with the Power4 and Power5 chips, buses between cache and main memory, and I/O interconnections to the outside world that the Power6 chip uses to get data, will scale up with the clock speed too. The Power6 chip is a dual-core processor.

Cooling techniques

Given that the majority of today’s applications are single-threaded, they cannot take advantage of multiple cores. Optimising software for multi-threaded throughput can accelerate performance and improve utilisation, which reduces the amount of energy required to accomplish the same work.

Says James Reinders, Director of Marketing and Business at Intel Corp, “We have software tools such as Wide Dynamic Execution, Intelligent Power Capability, Advanced Smart Cache, Smart Memory Access and Advanced Digital Media Boost that help improve performance with lower power consumption.”

There are other ways to cool servers. For example, HP used Thermal Imaging for data centres wherein it can identify hot/cold spots so that cooling can be planned. Additionally, it uses water-cooled racks and special-purpose fan technology that is cheaper than cooling the room.

Concludes Fichera, “Improving energy efficiency and solving the problems generated by increasingly dense server form factors is a community effort that requires contributions from semiconductor, system, software, and data centre operation vendors. Vendors, of whom some lead the charge and others are being dragged along, have collectively realised that there is money in efficiency; they will vie for energy-conscious customers over the next 24 months.”