Ti verde seu proximo datacenter pode ser um container

Notas do Editor

• #4: (move the discussion now over to IT):Energy does matter, not only in IT. The reason is the associated cost with the energy needs. There are direct costs and indirect costs related to energy. We will focus on direct costs first: the energy costs needed to run the current IT environment. Ask if the customer already measure their energy needs and have the costs in their IT budget. Nearly no IT person will be able to answer the energy cost question, as most often energy costs are within the facility group of the customers. This fact often prevents saving efforts: what cannot be measured cannot be optimized. And whom should you reward for saving efforts if the savings cannot be related to someone?Ask your customer what they would do if they would need to cover the additional 7% in their IT budget. Ask what 7% actually would be in real currency to make the picture more realistic! Ask if they would THEN start thinking about new purchase criteria: energy consumption of IT components![regarding the 7%]:The direct costs of IT can be estimated quickly: assume 3,500KWh energy needs per server per year (multiply by number of servers) and 700KWh per client (multiply by number of clients). For a midsize company (30 servers, 500 PCs), this is around 455,000 KWh a year. At 0,2 CHF per KWh, this is around 91,000 CHF per year. This estimation does not include cooling, network, storage, printers, datacenter etc. so this is the minimum prize you would pay for the energy needs. For most midsize customers, this amount (91,000 CHF) is more than 7% of the budget (which would be 1,3 Mio CHF)(click)Show the example of Germany: the energy market was opened 6 years ago. During this new “open market” time, the costs have increased over 36%, although the competition between the energy suppliers was now possible. And with the increasing energy needs worldwide, it is more than likely that the energy costs will raise again. Some analysts thinks it will even double in the next 5 years. Would your customer love to see the raising energy costs on their IT p&l? [indirect costs of energy]So far we only spoke about the direct costs. But unfortunately, you cannot just consume more and more energy according to your energy needs. There is an upper limit, defined per regional energy supplier. You cannot put unlimited energy through a dedicated power line, or even produce this unlimited energy. And the end, this could mean in best case you have to purchase a new power line to your buildings / datacenters, or in worst case you have to move to a new location offering the required energy. Who will cover these indirect costs?(click)Most people will be surprised to hear that there are huge difference for servers in the energy needs and efficiency. Regarding the calculation: the average 1u (pizza box) server consumes 400W per hour: 400w x 24 x 365 = 3505 KWh / yearEnergy costs with (0,2 Rp per KWh): 700 CHF or over 3 years: 2,100 CHF This is close to the server cost: A HP Proliant DL 145 is about 1600 CHF, a DL 360 is about 2600 CHF.Now power consumption of the server is normally only 30% of the overall energy costs in a datacenter, so the running costs might be up to 6,000 CHF for 3 years. This might be higher than your acquisition costs! So if you purchase a model which is 10% more efficient in energy, it could cost even more than 10% more than the inefficient server as energy costs are similar or even higher than acquisition costs.
• #5: This IDC slide from 2006 shows where the IT budget is moving to: More and more money is spent on power and cooling year over year; while the amount for new server stays nearly flat. So each year, the operating costs will be a bigger part of the IT budget. And with increasing energy costs, this ratio will change again in the future.
• #6: Depending on your customer profile, the energy will go to different areas. Larger enterprise customers / hosting providers: the data center will take a big part of the overall energy. We will focus on datacenter in the next slides in more details All customers: due to the huge numbers, the PCs are normally the biggest part of the energy bill. And here you can save most with “common sense” behavior: makes sure the PC don’t consume energy if they are not needed (weekends, nights etc). Department servers are the hidden factor for your server calculations. They are different than your datacenter servers as you cannot “assign” cooling costs etc to them if they are “just in a closet” or below the desk. Nevertheless, the worldwide percentage of these servers is huge, and much higher than the number of servers going into datacenters. Printers & network devices are currently out of scope of this discussion, mainly because I don’t have any good numbers and Microsoft is not working in this area at all. If you like, mention that most network printers have power saving settings for years (the wake-up phase we all know) and that Microsoft embedded OS offerings (like XP embedded) has the same smart energy saving features as the full OS we will cover later. Gartner estimated that printers are responsible for 6% and network components for 7% of the overall energy needs, so it’s not the biggest part of the bill. Storage: in the recent past, new HD models dramatically decreased the amount of power needed. The new disks consume 9W instead of 18W, but still your current SAN might need a lot more energy and should be updated to newer products over time. We won’t focus on them in this presentation. There are studies listing energy needs and showing saving options for different archiving and storage products and vendors, like the “Green Archival Storage” whitepaper from Plasmon.com: 40 TB in a Netapp FAS 3020 is about 2,5 KWh (~ energy of 6 midsize server) 40 TB in an EMC Centera (mirrored) is about 5,2 KWh (~energy of 13 midsize server)While this is noticeable energy, it is small compared to the overall server energy in an average ratio of servers / storage units.
• #9: Datacenter consume a lot of energy worldwide. The percentage estimations vary, but 1,5% is a good estimation average. With this value, IT is one of the biggest industry player regarding to the energy needs. Especially as large commercial datacenter from google are not in this calculation due to their specific situation (google builds their own servers from standard components so they are not counted in the worldwide server sells).It’s estimated that google is running their services on 450,000 servers worldwide. And of course there are more internet providers, like yahoo, MSN, amazon etc. Some analysts estimates that in 2030 nearly 30% of the worldwide energy goes to datacenters. http://www.wired.com/wired/archive/14.10/cloudware_pr.html Datacenter servers “consume” more energy than standalone server due to the additional cooling and infrastructure energy needed for them, which we will see in the next slides. It’s estimated that there are 4M servers in the US in data centers. With 2M servers shipped worldwide Q2 2007, this is “just a percentage” of the overall servers. So we have to address both situations, and both in a different way.
• #10: You can see that only 30% of the energy is actually energy used for servers (and network components). The rest (70%) goes to infrastructure, mainly cooling etc. So any energy you save at IT equipment could be saved again twice in cooling (energy in = waste heat out). But this requires a dynamic cooling infrastructure, too! In many installations, the chillers run at 100% regardless of the cooling load needed. So in an environment like that there is no savings.  Which brings up an important note:  One of the best recommendations we can make to customers is to buy infrastructure that scales to demand.  This includes servers, power infrastructure (UPS, PDUs, etc), and cooling.In some regions, humidifier are not needed at all, some vendors allows 20-80%, so think if this is needed for your datacenter at all.This comes back to the beginning with the light bulb: any energy you can save with “shutting down” is better and makes most often bigger effects than switching to new, more efficient technologies. At least this will help you at the beginning.After you have implemented these easy steps, you can only achieve the next savings by moving to different / more efficient technology which includes upfront investments. Most often the ROI of these investments is achieved quickly! Interesting though, the energy distribution within the datacenters have changes over time as we will see in the next slide
• #21: On power capacity reliability cost generation sources
• #42: Increased power efficiency at data centers.Microsoft has many data center locations throughout the world. Given the costs expected from growth in these data centers, Microsoft is focused on maximizing efficiencies. The Data Center Services team closely monitors and manages current power consumption and the design of future data centers. The largest Microsoft data center to date, in Quincy, Washington, is an example of the way the company is building for the future. It was designed to have minimal or no carbon footprint, with hydropower as its primary source of energy. Hydropower is a clean-air and renewable resource from the Columbia River Basin. The data center will maximize its use of green building technology under the Leadership in Energy and Environmental Design (LEED) rating system. The LEED Green Building Rating System is a voluntary, consensus-based national standard for developing high-performance, sustainable buildings.Microsoft is also researching new energy options and technologies that will bring about even greater efficiencies. Working closely with energy companies, Microsoft is exploring a variety of power generation and transmission technologies—some that have long been used in large industrial processes but have never been applied to data centers.
• #45: The total load of the facility is 120MW and the size is 100k sq/mtrs
• #48: Key Points –Pre manufactured supply chain of Data Center Facilities for each class using the same kit of parts delivered at the same time as serversEliminate concrete – responsible for 5% of global green house gasesHow do we get to such low costs? Class A – no building, no UPS, no generators. E.g. VE in Colorado came in at ~$3.3M per MW – that was at very low density and included all site costs. High density reduces costs – 200 servers in a container costs more than 2000 servers.Graphs on rightIncremental deployment – for the same investment as a Mega DC split up and develop 10 sites for Next Gen deployments – choose Class of DC and where in a 3 month ttm (time to market)Split Design - leverage the industry – move design away from large fixed facility to Colo Modular vendor focused effort – independent of total capacity at a single sitePie Chart on construction costs - ~50% Total costs are in labor and ducts, pipe, conduit, and copper to connect major pieces of equipment
• #50: Gen 1, mid 90’s – adding capacity. Focus on uptime, reliability and redundancy. Look at the deployment scale unit. It was a serverGen 2, Quincy, St Antonio – 12 football fields under roof. Cost of the buildings is $250M… majority of the cost is cement and copper. without putting a $ of IT equipment. These are beautiful DCs. Cover 12 football fields. We wont build more like these again. Completely hydro-electric power in Quincy and recycled waste water used for cooling in St Antonio. Gen 3, Chicago. – 700K sq ft. 120 MW off the grid. own power sub station, water station. We were convinced containers are scale units. Upper is raised floor… Gen2. and lower floor is all containers. Gen4 – modularized DC. Just in time DC. We have done all the engineering there. We have to be highest productivity per watt. In case of Gen 1 – PUE of 1.9Gen 2 – 1.7Gen 3 – 1.4Gen 4 – 1.2Cost/ MW to build out capacity. Gen 1 – $25M/MWGen 4 - $10M/MW and Edge node is about $2M/MWContainers have been around a long time..
• #53: Traditional server nowadays needs in average around 400-500W per server. This is depending of course on the server components (processing power, storage, cards, memory) of the server. High-end servers (multi-core Unix & mainframe machines) consume around 9,000W and more an hour. Blades seems to be more efficient with energy, but you need to add the frame to them. If you only run 1 single blade, the frame would be a large addition, in a normal case with many blades in a single frame, this can be ignored for this discussion.(click)Some facts from IDC as a background for you: 2M server shipped in Q2 2007. Now, the worldwide energy used for servers is noticeable! And even the midsize SMB customer have 30servers and more, each running at 400W in average.
• #54: Again, smart hardware purchases reduced the energy needs directly from the beginning. What is not needed at the beginning must not be reduced later on!Especially at low costs servers, they are not very efficient (60-70% based on some estimations, though 75-85% are more realistic) due to cost pressure. Now, if customers would look at the lifetime costs, they might ask for other power supplies and spend 10$ more at the beginning to save >10$ over the next years. 85-90% efficiency for power supplies is affordable and the norm for highend servers! Modern blade racks or the latest high-end servers can even reach 95% with smart supply designs (shutting of unused supplies of redundant supply sets to reach higher utilization = efficiency). Remember though, that the efficiency of power supplies unfortunately is not a static value but a function of the hardware parts and the load of the power supply. So it will vary over a broader range; and an efficient supply for high loads might be inefficient for lower loads. Best practice is to not oversize the power supply and at the same time run the server at little load: this will be the most inefficient situation for any power supply. Try to achieve a higher utilization from the beginning: are you really planning to change the server over time 8add cards or disks) so you need an oversized power supply?(click)The results are clear: with a better power supply you save energy & costs. Each hour, each day, each year. 100W savings an hour is 876KWh a year. This can be 175 CHF (0.2 CHF per KWh) and could be more than the prize for a better power supply. So ask you hardware vendor for different options when you purchase your new servers. Changing existing power supplies might not have a positive ROI due to server outage and maintenance work, but you would have to do your own math.
• #55: This slide is valid for servers only!Beside the energy lost for conversion, the different components of the servers are using energy, too. Now, if you don’t need certain components, don’t put them in your servers but remove them during initial server assembling!Most of the energy in servers is used by the CPU, followed by PCI cards and motherboards. With Windows 2003 onwards, we can address the CPU issue, and with Windows Server 2008, we will address the PCI-E cards. This makes a big chunk out of the server energy. Power management can be controlled in several ways. The whitepapers at the end describes certain features in more details. Some hardware vendors implement the management in the firmware / bios of the systems. Microsoft’s point of view is that the OS should control the system settings defined in the ACPI specifications. Pro and cons are discussed in several technical resources, some of them are mentioned at the end and are worth to be read. (Some background information:There are some power savings features the OS simply can’t manage as quickly as hardware can, due to the nature of the hardware design.  For example, AMD is shipping processors today where they manage the deeper ACPI c-states (processor idle sleep states) in hardware, due to the way in which they’ve chosen to integrate their memory controllers with the CPU, and the interaction with UMA graphics.  They still need OS to enter the first level idle state, then they h/w can go deeper opportunistically.  There are many other examples of this device-specific behavior.)Now, with the new / upcoming “flash memory” hybrid hard disks introduced with Windows Vista or diskless servers booting from SAN (since Windows 2003), and new disks needing half the energy compares to traditional disks, this area is addressed, too. Unfortunately there are no test results available describing the achieved power savings (if any!) with this technology). In the moment, SAN drives requires more energy than local drives, especially when you add the needed network components, so this is not an option purely seen from an energy point of view!So let’s focus on the area where we can offer a solution today: the CPU.
• #58: Processors nowadays know different “states”. In the past there was the turbo switch on your 386 machine, remember? This switch changed the overall frequency of the whole system from 8 to 33 MHz (or something like this). Now, modern processors handle this differently due to the overall increased system complexity and the higher frequencies. You simply cannot run the whole system at 2+ GHz. Each processor model has its own boundaries of the highest (default) and lowest frequencies and the specific steps in between. Older processors knows a single state, newer up to 5. A good rule of thumb: as lower the frequency as lower the consumed energy. To find out the different states, the OS must be able to read the processor settings from the bios, specified in the ACPI specifications. In Windows 2008, we support Intel and AMD processors out of the box, in Windows Server 2003, you must install AMD add-ons. By default, Windows Server 2003 runs at the highest frequency level all the time, but you can enable dynamic frequency switching on certain systems. Enabling must be done in the bios as well as on the OS level. Switching today is then highly dynamic, done automatically of the OS if wished and might take a 1/10 of a second, so nothing people should be afraid about.(click)If you run Windows Server 2008 or have enabled this feature for Windows Server 2003 (SP1), Windows figures out the different states and changes between them automatically. In Windows Server 2008, you can even define the max and min states at certain systems from all of the available states if you want to set your own boundaries. The feature set for both server OS are described in much more details in the mentioned whitepapers in the resource list as we can give here just a highly simplified model for the short overview. (click)As you can see here in the AMD product information, each state is not only running at a different frequency, but uses different energy, too. Frequency and energy are related to each other; you cannot simply change the voltage / energy and run at certain frequencies. The physical models and more theory behind is described in the whitepapers and at the intel and AMD web pages. You see here by the AMD example that different processor states could be used and therefore the energy need of the processor changes. Clearly different frequencies are related to different processor performance, too: as higher the frequency, as more performance can be achieved. So it depends how much performance do you need at which frequency you can / want to run.
• #61: You can see the frequency changes (related with the different p-states of the processor) in Perfmon by monitoring the processor frequency over time; best together with the processor utilization: if the utilization raises, the frequency will go up, and then drop again when the workload is over.To show the perfmon example in real, simply open perfmon and ad the following counters:Processor Performance: % of max frequency (all instances, though one might be enough). Keep the % Processor time and show both curves over time: if there is no or little load on the server, the frequency will drop to the lower limit you have configured or will stay on the upper limit and waste energy if you didn’t. This can be done on your presentation laptop, too. You might need to start the Perfmon application as an admin under Vista to see the processor performance counter object!
• #62: There is a great whitepaper which describes this slide in details (s. link below). In summary: You just activate the “state switch” feature in Windows Server 2003 like on would do on your laptop: (Click)In “control panel /energy settings” simply choose a different energy scheme than (max power / always on). There is a VBScript in the whitepaper which checks your processors for all available states and another script to set the energy states on your servers. The whitepaper can be found at http://www.microsoft.com/whdc/system/pnppwr/powermgmt/w2k3_ProcPower.mspx(click)This feature is enabled at Windows Server 2008 by default (screenshot) remember though that this feature must be activated in the bios AND your hardware must actually allow the dynamic p-state switching.
• #63: Eliminate the mixing of hot and cold air….there is a host of industry best practices around that and some are listed belowHot Aisle/Cold AisleEliminate Gaps in RowsUse longer rowsUse Cabinet Blanking PanelsSeal cable cutoutsUse High and Low Density AreasConsider economizersBetter temperature control and airflow distributionAir-conditioning needs to Match Server Airflows to ensure that all servers get fresh conditioned airOrient AC units perpendicular to hot aislesUse 0.8m to 1.0m High Floors  to provide more uniform static pressure under the floorUse economizers - substantially reduce energy use with right outside conditions Use liquid side economizers which use outside air to cool chilled waterUse air side economizers which bring outside air directly into the data centerEngineer the Data CenterUse modeling tools Design the data center like BoeingTake Control of your data center – monitor and control everythingMeasure server temperatures realtime and at the inlet throughout the datacenterThis is the only temperature that matters in the data centerDevelop tools to capture performance, temperature and power throughout the data centerMine your data to develop of a deep understanding on how to improve operationsMeasure your Power Usage Effectiveness real time and all the timePUE = Facility Power / IT loadMeasuring this will improve your data centerEach new data center “better be better”Eliminate overprovisioning for maximum efficiency and productivity - The data center is a living organism that is changing all the timeScale the Infrastructure with your needsUse power management and virtualization whenever possible.Match the data center Availability with the application needsUse Power and Cooling Architectures that Don’t strand powerDevelop flexible designs that don’t lock up capacity where it is not neededOptimize holistically everything that goes in the data centerDevelop TCO models that optimize holistically Include CAPEX and OPEXInclude Energy, Infrastructure, network and ITDevelop tools for optimum site selectionMake the data center part of the culture Data Center metrics should be part of the day to day lingoDevelop tools to that help communicate operational health of your data centerShare and Learn from industry partners (The Green Grid, Climate Savers, EPA, LBNL, ASHRAE, etc)Work with the industry and share best practiceBe open minded and learn from others – nobody knows it all