FACILITIES ef f icienc y
#dcsarticle
http://dcseurope.info/n/riis
Breaking new ground on data
centre efficiency Part 2 of a two part article explaining how eBay’s “Project Mercury” used PUE, TCO and DCMM best practices to drive the end-to-end data center ecosystem. By The Green Grid.
T he Project Mercury data center was designed with 1300
square meters of usable Tier II IT space. The data center needed to have two levels: one open raised-floor area for rack-and-roll deployments, and a second rooftop infrastructure capable of supporting up to twelve, 12-meter long containers. The building would be fortified to support over one million pounds of weight with an initial capacity of 4 MW of IT power and cooling and the ability to scale to 12 MW over time.
With these rough parameters established, eBay began Project Mercury by initiating two parallel processes: one for server procurement and one for data center design. The philosophy was based on the idea that the PUE and TCO metrics would be integral to the RFP process and would help to drive the supply chain to deliver the most cost-effective and energy-efficient infrastructure for eBay. This approach again illustrates eBay using current and future best practices from the DCMM by incorporating TCO in decisions involving both IT procurement and facility design.
Server Procurement For its server purchases, eBay issued RFPs that incorporated three key focus areas for the project: holistic total cost of ownership, volume packaging and delivery, and future technology implementations. The RFP process was designed so that open competition could fuel technological advancements, leading to the achievement of eBay’s business objectives: a lower TCO, and reduced energy consumption. By using a competitive RFP process with industry vendors, eBay was able to achieve similar efficiencies and optimizations compared to organizations building their own white-box solutions. eBay considers its vendor engineering teams as extensions of its own, with open competition between them stimulating innovation.
Simplifying Server Types The process of reducing the number of SKUs began with an internal assessment of server requirements. After an extensive asset audit and workload performance requirement evaluation, eBay found that it could reduce the diversity of its server population. eBay was able to consolidate 15 primary server SKUs down to only two server types: those systems suitable for its high-performance computing deployments, and those systems suitable for supporting big data applications. As described in Section 5.2 of the DCMM, reducing the number of server types simplifies the number of moving parts
20
www.dcseurope.info I May 2012
in the data center, increases workload portability, and increases agility through more rapid deployment. From the vendor perspective, responding to specific requirements in eBay’s RFP by fine-tuning server designs was more feasible for a small number of SKUs. The process enabled a high degree of optimization for the two server types, resulting in highly efficient products in eBay’s data centers, and arming the vendors with better products to sell to the rest of the industry.
Updated Search Engine
The first RFP eBay issued was for high-performance servers to support a new compute-intensive workload. This RFP was designed to encourage server suppliers to meet the specifications eBay laid out based on the total cost to procure and operate a server. This included server cost depreciation and total kilowatt hours projected to be consumed over three years running eBay’s standard workload. After an intense process with more than 50 participants, eBay’s selection was based on the vendor’s unit cost and ability to fine tune its server components to consume the smallest wattage possible while executing the eBay-specified workload modeled over three years. This resulted in a holistic TCO that assessed the real cost of operating the server over its useful life.
Supporting Big Data
The second public RFP was for servers to support new Hadoop clusters. The vendor selected for this RFP had fine-tuned its servers and container solutions to improve the total wattage consumed from what was reported in the first RFP response. The vendor’s engineering teams increased efficiency through improved airflow, lower-voltage DIMMs, low-load server BIOS optimization, and improved CPU heat sinks. The result was energy savings of up to 16 percent at full load and 41 percent while idle, a highly significant improvement that could benefit not just eBay, but other customers of the optimized product.
Profiling Energy Use to Workload
While improvements in server energy efficiency came from the bottom up as vendors optimized their servers to perform against eBay’s RFP, other innovations came from the top down as specific RFP requirements. One of the innovative elements of the RFP required hooks to allow eBay software to dynamically change the server CPU clock frequency so that it could slow down a server (and reduce its energy consumption) when workload conditions permit, and ramp
Page 1 |
Page 2 |
Page 3 |
Page 4 |
Page 5 |
Page 6 |
Page 7 |
Page 8 |
Page 9 |
Page 10 |
Page 11 |
Page 12 |
Page 13 |
Page 14 |
Page 15 |
Page 16 |
Page 17 |
Page 18 |
Page 19 |
Page 20 |
Page 21 |
Page 22 |
Page 23 |
Page 24 |
Page 25 |
Page 26 |
Page 27 |
Page 28 |
Page 29 |
Page 30 |
Page 31 |
Page 32 |
Page 33 |
Page 34 |
Page 35 |
Page 36 |
Page 37 |
Page 38 |
Page 39 |
Page 40 |
Page 41 |
Page 42 |
Page 43 |
Page 44 |
Page 45 |
Page 46 |
Page 47 |
Page 48 |
Page 49 |
Page 50 |
Page 51 |
Page 52