CAPÍTULO IV: RESULTADOS
4.2 Contraste de hipótesis
4.2.2 Hipótesis Secundaria
The subsections below address the performance, cost, and reliability impacts of energy- efficiency improvements to data center computing functionality.
5.1.1. Computing Functionality and Features
More efficient servers will not be less functional, nor will they have fewer needed features. Manufacturers may offer models, configurations, or options that lack specific features as a way to reduce power consumption, but those will only be purchased by customers who don’t need the omitted features.
5.1.2. Performance
In systems for which performance is paramount (e.g., high-performance computing),
performance will remain the preeminent goal even when attention is given to energy efficiency. Some of the technologies for accomplishing energy efficiency introduce higher latencies (delays) than is currently typical, either on an ongoing basis or when making the transition between performance states. These latencies are often small, infrequent, or irrelevant. Server operators will only adopt energy-saving features to the degree that such latencies are acceptable.
A likely trend is that systems will enable users to specify needed levels of responsiveness and for the system to automatically monitor responsiveness and adjust its behavior and performance to stay within the specified parameters. In contrast, current systems mostly operate at maximum performance even when this greatly exceeds actual requirements.
5.1.3. Overall Cost
Some methods to reduce power consumption do not increase product cost. Other methods do increase cost but will typically only be implemented when the savings are considerably larger and offset the increased cost. Savings can result from reduced hardware purchases, reduced maintenance costs, reduced IT energy costs, reduced power and infrastructure operating costs, and avoided or deferred facility upgrades and expansions.
More efficient hardware may cost modestly more than the alternatives because of extra,
sophisticated, higher-quality components or more rigorous testing. Virtualization software (and hardware) adds cost, and the complexity of power management could increase time to market for products and so indirectly increase costs and delay the onset of the savings. Other solutions (e.g., distribution of high-voltage DC current) could lower hardware costs, save large amounts of energy, and improve reliability.
An aspect of integrated circuit production that is not generally apparent is a distribution in efficiency among seemingly identical components. For example, one product off the production line might consume 10 percent less than average, and the next product could consume 10 percent more. Variation among production lines of “identical” products is often even larger. This variation could become an explicit part of the market with more efficient products costing more (similar to the current phenomenon of processors that run at faster-than-normal speeds carrying a price premium).
Some operational changes may increase costs by requiring more staff time to implement and monitor than the traditional methods of greatly over-provisioning capacity.
Some energy-efficiency measures actually reduce the cost of building a new data center. Examples are supplying DC power (avoiding some power conversion and cooling equipment), sizing for the right load (and providing ways to efficiently add on later), improving system design and layout, and using free cooling (to reduce or eliminate the need for chillers).
As noted above, all of these potential cost increases are expected be much less than the value of lifetime energy savings and avoided site infrastructure costs from more efficient products, often with a very short payback period, so the more efficient systems have a lower total cost of
ownership (TCO). In general, however, it is not possible to draw general conclusions about the additional cost of energy-efficient products because there is significant variation among product types, the efficiency strategies applied vary among products, and the rapid pace of change in technology makes it difficult to collect cost data.
5.1.4. Reliability
Manufacturers have reliability standards in their equipment designs, and it is safe to assume that these will be maintained at current levels for new equipment. However, the following strategies that can be applied to improve energy efficiency may also affect reliability:
Energy efficiency strategies that may increase reliability
• Lower interior temperatures,
• Lower data center temperatures overall, and
Energy efficiency strategies that may reduce reliability
• Increased system complexity,
• More dynamism in system states and activity, and
• Higher average utilization.
One example of the relationship between reliability and product design is the spinning down of disk storage drives during long periods of non-access to save energy. If drives were not designed for routine spin-up/spin-down, this could result in a reduction in reliability. However, many drives are designed with this usage in mind so that frequent spin-down will not reduce (and may even increase) overall reliability. Today, only lower-speed drives are available that can spin down, not “server-class” drives, so this strategy can only be used in situations where data access speeds are not critical.
Many of the quantitative data about reliability and its relation to performance are anecdotal and/or proprietary. An exception is the study by Pinheiro et al., (2007) on failure trends in disk drives. They noted that the literature on disk drive failure patterns is sparse and found that “temperature and activity levels were much less correlated with drive failures than previously reported.” Moreover, Guha and Ouderkirk (2006) found that actual drive failure rates in MAID systems were reported to be significantly lower than the rates experienced in traditional always- spinning drives. Similarly, Bodik et al. (2006) argue that efficiency improvements will actually increase reliability because the reliability gain from not running equipment continuously can more than offset the reliability reduction from on/off cycles. This is an indication that some feared reliability risks of efficiency measures may be illusory.
Server power supplies interact with reliability in several ways. For example, removing the AC/DC conversion and directly feeding DC could improve reliability and improve efficiency (fewer potential points of failure), or it could reduce reliability (if there are fewer redundant components used in the system).
An increasing number of operators utilize redundant servers and data centers so that eliminating dual power supplies could improve efficiency and decrease local — but not overall — reliability, assuming the DC system was designed to provide the same level of redundancy. Virtualization technology can facilitate easy movement of applications among servers to increase reliability. More efficient cooling inside the chassis of IT equipment can support both efficiency and reliability (especially in blade servers). This can be accomplished by use of efficient fans, variable-speed drives, aqueous or dielectric liquid cooling, and heat pipes. In addition, some of these solutions could be coupled with efficient building systems to effectively move heat out of the building, which might reduce building infrastructure cost as well.
Ultimately, however, equipment failure modes need to be analyzed as part of understanding the effect of energy efficiency on reliability. In practice, this analysis is very complex as there are many possible failure modes, and this analysis is specific to the particular product being
examined. For this reason, it is not possible to undertake a complete analysis here; instead, this report treats qualitatively some of the factors that affect reliability.