Página 1 de 2 12 ÚltimoÚltimo
Resultados 1 a 10 de 14
  1. #1
    WHT-BR Top Member
    Data de Ingresso
    Dec 2010

    [EN] A Rare Tour Of Microsoft’s Hyperscale Datacenters

    The shiny new Open Cloud Server racks going into the Generation 5 datacenters in Quincy

    Timothy Prickett Morgan
    September 26, 2016

    If you want to study how datacenter design has changed over the past two decades, a good place to visit is Quincy, Washington. There are five different datacenter operators in this small farming community of around 7,000 people, including Microsoft, Yahoo, Intuit, Sabey Data Centers, and Vantage Data Centers, and they have located there thanks to the proximity of Quincy to hydroelectric power generated from the Columbia River and the relatively cool and arid climate, which can be used to great advantage to keep servers, storage, and switches cool.

    All of the datacenter operators are pretty secretive about their glass houses, but every once in a while, just to prove how smart they are about infrastructure, one of them opens up the doors to let selected people inside. Ahead of the launch of Windows Server at its Ignite conference, Microsoft invited The Next Platform to visit its Quincy facilities and a history lesson of sorts in datacenter design, demonstrating how Microsoft has innovated and become one of the biggest of the hyperscalers in the world, rivaling Google and Amazon Web Services – companies that are its main competition in the public cloud business.

    Rick Bakken, senior director of datacenter evangelism for Microsoft’s Cloud Infrastructure and Operations team, was our tour guide, and as you might imagine, we were not allowed to take any pictures or make any recordings inside the several generations of datacenters at the facility, which is undergoing a massive expansion to keep up with the explosive growth of the online services that Microsoft supplies and that are increasingly deployed on the same infrastructure that is used to underpin the Azure public cloud. But Microsoft did tell us a lot about how it is thinking about the harmony between servers, storage, and networking and the configuration of the facilities that keep them up and running.

    From MSN Web Service To Azure Hyperscale

    Each of the hyperscalers has been pushed to the extremes of scale by different customer sets and application demands, and over different time horizons. The Google search engine was arguably the first hyperscale application in the world, if you do not count some of the signal processing and data analytics performed by US intelligence agencies and the military. While search continues to dominate at Google, it is now building out a slew of services and moving aggressively into the public cloud to take on Amazon Web Services and Microsoft. Amazon has to support more than 1 million users of its public cloud as well as the Amazon retail business, and Facebook has to support the 1.7 billion users on its social network and the expanding number of services and data types they make use of as they connect.

    For Microsoft, the buildout of its Azure cloud is driven by growth in its online services such as the Bing search engine, the Skype communications, the Xbox 360 gaming, and the Office 365 office automation tools, and others that add up to over 200 cloud-based services that are used by more than 1 billion users at more than 20 million businesses worldwide.

    Those are, as The Next Platform has said many times, some big numbers and ones that make Microsoft a player, by default, in the cloud. And those numbers not only drive big infrastructure investments at Microsoft, which has well north of 1 million servers deployed across its datacenters today (including Azure and other Microsoft services that have not yet been moved to Azure), but they drive Microsoft to continuously innovate on the designs of its systems, storage, networking and the more than 100 datacenters that wrap around them in its infrastructure.

    This move from enterprise scale to hyperscale all started a little more than a decade ago for Microsoft.

    “I worked for Steve Ballmer back in 2005 and I was on his capacity planning team,” Bakken explained to The Next Platform during the Quincy tour. “We ran two datacenters then, one in Quincy, Washington and the other in Chicago, Illinois. We looked at capex costs of these two facilities and they were dramatically different. Let’s say $15 million per megawatt for Quincy and $20 million per megawatt for Chicago, just as a bogie. What was really interesting is that we looked at power costs and there was quite a difference in what we paid in Quincy with hydroelectric and what we paid in Chicago. That variance in power cost drove us to question what we were doing building raised floor datacenters. What came out of this was a realization that we were really building large air conditioners, that we were not in the IT business but in the industrial air conditioning business when you build a facility like this. And this drove us to deeper questions as to why we were doing that, and the answer was heat.”

    So rather than take the heat, Microsoft’s engineers went after it.

    To backstep a little bit, between 1989 and 2005, with the first generation of co-located datacenters that Microsoft paid for, the company bought generic servers from the big tier one suppliers and racked them and stacked them like every other enterprise company of the client/server and dot-com eras did. With these facilities, which had traditional raised floors used in datacenters for decades, the power usage effectiveness of the datacenter – a measure of efficiency that is the total power of the datacenter divided by the total power of the IT gear – was around 2.0. So, that meant for every 1 megawatt of juice consumed by the IT gear doing actual work there was another 1 megawatt of electricity being consumed by the datacenter as voltages were stepped down for battery backup and the gear in the racks and water was chilled and air blown around the datacenter.

    Starting in 2005 with the Columbia 1 datacenter in Quincy and then adding on with the Columbia 2 facility in 2007, which together are the size of a dozen American football fields, Microsoft started using rack-dense system configurations. Each of the datacenters has five rooms with 18 rows of racked up gear. Each room is rated at 2.5 megawatts, and across those ten rooms with the networking and other gear thrown in, the whole thing draws 27.6 megawatts. With the compute density and improved cooling, Microsoft was able to get the PUE of its Generation 2 datacenters down to between 1.4 and 1.6, which means the power distribution and cooling were only adding 40 percent to 60 percent on top of the juice burned by the IT gear itself. This was a big improvement.

    With the Generation 3 datacenters that Microsoft rolled out in 2008, it took on the issue of containing the heat in the datacenter. The containment came literally from containers – in this case, shipping containers, and with some of the Microsoft locations (such as those used in its Chicago facilities), the containers were double stacked and packed with gear. (See here for more details on this.) This was the first time that Microsoft used air and water economization technologies in its datacenters, and it was able to drop the PUE from between 1.2 and 1.5.

    There was not a lot of Generation 3 gear in Quincy, and that was because Microsoft was working on the next phase of technology, its Generation 4 modular systems, for its other flagship facility. But hot aisle containment was retrofitted onto the Columbia 1 and 2 datacenters, which are mirror images of each other, using clear plastic barriers from floor to ceiling.

    Back in the mid-2000s, when Columbia 1 and 2 were built, server compute, storage, and power density was not what it is today, and so now, with the Microsoft’s Open Compute Server platforms installed in these facilities, only eight of the 18 rows have gear in them now, and because Microsoft has substantially virtualized its networking inside the datacenter, load balancers, switches, and routers are no longer in the facilities, either. (It is a pity that the datacenter can’t have its power distribution upped to 5.7 megawatts and be filled to the brim again.)


  2. #2
    WHT-BR Top Member
    Data de Ingresso
    Dec 2010

    Buying A Datacenter In A Box

    By 2011, Microsoft’s IT needs were growing fast as it made acquisitions and started preparing for a massive public cloud buildout, and the datacenter architecture changed again.

    “In Generation 4, we learned two things,” said Bakken. “First, buy the datacenter as a box with a single part number, which we call ITPACs. But at the same time we moved away from chiller environments and into chiller-less environments that use adiabatic cooling and run at above 85 degrees Fahrenheit. We filter the air that is coming in to remove particulates and then I have a separate barrier of filters that I can add water to above 85 degrees that will lower the temperature inside of the box and lower the heat. In the winter time, I can seal the box up and all of the heat stays inside. The question is, how often is it hotter than 85 degrees in a place like Quincy? It was 21 days last year. So I am going from a chiller environment where I am running chilled water loops for the full length of the datacenter to an ITPAC when I am only turning on the water when it gets warm outside. My CapEx comes way down, and my OpEx comes way down, and look at the efficiency. I can run in the winter at about 1.02 PUE and right now, in the late summer, they are running at about 1.2 PUE. That is really impressive compared to our Gen 2 datacenter, which was running at about 1.6 PUE.”

    With the ITPACs, which were manufactured by Hewlett Packard Enterprise and Dell Technologies (if we go by their new names), Microsoft told them that it wanted containerized units of datacenter that would take 440 volt power in and have a plug for the network and just run 2,000 servers in a pod. Bakken did not care how HPE and Dell did it, so long as it met the specs and could be rolled into the datacenter. As it turns out, they had very different designs, with HPE creating a pod that had a shared hot aisle in the center and two rows of gear on the outside and Dell creating a pod that had separate hot and cold aisles. These pods were lifted into a datacenter with a crane and then the roof was put onto the building afterwards, and interestingly, the first day that they went in it snowed like crazy in Quincy and the snow blew in through the side walls and surrounded the pods. (Microsoft eventually put up baffle screens to keep the rain and the snow out but let the wind in.)

    The outside view of the Generation 4.6 datacenter in Quincy

    Over the years, Microsoft has done a few tweaks and tunings with the Generation 4 datacenters, including getting rid of large-scale uninterriptible power supplies and putting small batteries inside servers instead. After realizing what a hassle it was to have a building with walls and a roof and that it took too long to get the ITPAC pods into the facility, Microsoft went all the way and got rid of the building. With the Generation 4.6 facility that was built in Quincy, the pods stand in the open air and the network and power are buried in the concrete pad they sit upon. And with the Generation 4.7 facility in Boydton, Virginia, Microsoft shifted from outside air cooling to adiabatic cooling, where air is blown over screens soaked with water in the walls of the datacenter to create cool air through evaporation of the water.

    All told, the four generations of datacenter gear in Quincy take up about 180 acres today. But the Generation 5 facility that is being built across the main drag on the outskirts of town has 300 acres in total, and is a testament to the growth that Microsoft is seeing in its cloudy infrastructure from the Azure public cloud and its own consumer-facing services.

    “We have taken all of the learnings from prior generations and put them into Generation 5,” said Bakken. “I am standing up four versions of infrastructure that we call maintenance zones. They are identical to each other and I am going to stripe some fiber between them to have great communication. I am on cement floors, I am using adiabatic cooling, and I don’t have chillers.”

    While Bakken was not at liberty to say precisely how the IT gear will be organized inside of the datacenter – and the facility was still under construction so we were not permitted to enter – he did say that Microsoft was growing so fast now that it needed to scale faster than was feasible using the “white box datacenter” ITPACs it had designed. Presumably Microsoft is using some kind of prefabbed rack.

    The racks that Microsoft is using are obviously compatible with its 19-inch Open Cloud Server chassis, and Bakken said that the Generation 5 datacenters would use 96-inch tall racks (compared to 74 inches for a standard server rack), and that should mean Microsoft can get racks to go floor to ceiling. That 55U rack is odd in that it is not divisible evenly by the 12U height of the Open Cloud Server. It leaves room for four Open Cloud Server enclosures plus 7U left over for other gear – or for a taller or shorter future Open Cloud Server that will fit the racks.

    The maintenance zones of the Generation 5 datacenters will be ganged up in quads, called theaters, which are mirrors of each other, and they can be georeplicated around the globe to other similar datacenters at some point and provide 32-way replication (four-way datacenters in a zone and four zones in the theater times two theaters with for geo-replication) for absolute high availability for strategic applications. The first Generation 5 maintenance zone has been released to engineering to be populated by servers, but it doesn’t have gear yet.

    Microsoft wants to get more time out of its datacenters, and has designed them to have longer lives than its prior datacenters; 30 years instead of 20 years, to be precise. And the Generation 5 datacenters have lots of power, at 10.5 megawatts each, 42 megawatts for a maintenance zone, and 168 megawatts for the theater, to take some dense gear if need be.


  3. #3
    WHT-BR Top Member
    Data de Ingresso
    Dec 2010

    Aerial view of the Generation 5 datacenters in a maintenance zone in Quincy under construction

    “We have put some provisions in place to allow new ways and new technologies relating to power generation and those types of things to be integrated into our environment, Bakken explained. “We like DC, and in the new Generation 5 datacenter I am putting AC on the bus down one side and DC on the other, so when thin film and batteries and all of that is viable, or we put solar panels on the roof, we are good to go. Let servers get small or big, let networks do SDN things. Whatever. We have thought that through. We can still run within the same datacenter infrastructure. And the way that I am doing that is that I am putting in a common signal backplane for my entire server environment, which shares power, cooling, and network.”

    Software defined networking is a key component of the Generation 5 facilities, too, and Microsoft has been catching up with Google and keeping pace with Facebook (and we presume with AWS) in this area.

    “SDN is a very big deal,” said Bakken. “We are already doing load balancing as a service within Azure, and what I really want to be able to do is cluster-based networking, which means I want to eliminate static load balancers, switches, and routers as devices and go to a standard industry interconnect and do those services in silicon-based software. When I can do that, I have a completely different setup for my datacenters because I don’t have to bring in a main distribution frame and replicate that in the intermediate distribution frame and put it in the middle of the co-lo and then bring my carriers up with all of that dedicated equipment. What we are doing with Generation 5 is dramatically different. This is a big thing in my environment because networking makes up about 20 percent of the capital expenses because I have to replicate those frames in multiple locations.”

    Microsoft is not bragging about PUE for the Generation 5 facilities yet, but Bakken has indicated in past presentations that it will be in the same ballpark as the best Generation 4 facilities (see above), and we presume Microsoft is going to be able to push it down some. The other metric Microsoft is watching very closely as it warps higher into hyperscale is cost. Ultimately, that is what all of these changes in datacenter designs over the past decade are really all about. We can see the progression in datacenter efficiency over time, and so we asked Bakken about the change in costs over time?

    “It comes down exponentially,” Bakken told us referring to the costs moving from Generation 1 through the phases of Generation 4 datacenters. “You are buying a box, you are not standing up a building, so it changes everything. My OpEx with the new datacenters is that I have to change the filters, and that is really the only maintenance I have. And we have moved to a resiliency configuration where I put more servers in each box than I need and if one breaks, I just turn it off and wait for the next refresh cycle. The whole OpEx changes with the delivery model of the white box. So we learned quite a bit there, but now we have got to really scale.”

    And that is not a technology problem so much as a supply chain and technology problem. And what seems clear is that racks beat ITPACs when speed matters. Bakken said that he had to grow the Microsoft network – meaning literally the network and the devices hanging off it – exponentially between now and 2020. “I have got to expand, and it is almost a capacity plan that is not a hockey stick but a flag pole when you look at certain regions,” he said. “The problem I have right now? It is supply chain. I am not so worried about technology. We have our Open Cloud Server, which I think is very compelling in that it offers some real economic capabilities. But I have got to nurture my supply chain because traditionally we bought from OEMs and now we are designing with ODMs so we can take advantage of prices and lower our overall costs. So I am moving very, very quickly to build out new capacity, and I want to do it in a very efficient and effective way and it is really about the commoditization of the infrastructure.”

    This further and relentless commoditization is key to the future of the public cloud – all public clouds, at least those that are going to survive. The way that Bakken and his colleagues at Microsoft see it, the cost of using cloud is going to keep coming down, and presumably fall below the cost of on-premises infrastructure, because the big clouds will do it better, cheaper, and faster. But, as Bakken tells IT managers, with so many connected devices and cross-connected services, the amount of capacity IT shops buy is going to go up in the aggregate and hence net spending will rise. That is the big bet, and it is why Microsoft, Amazon, and Google are spending their many billions most quarters. Time will tell if they are right.


  4. #4
    WHT-BR Top Member
    Data de Ingresso
    Dec 2010

    Microsoft Bets Its Future on a Reprogrammable Computer Chip

    Project Catapult V1, the hardware that Doug Burger and team tested in a data center on Microsoft’s Seattle campus

    High-end, custom-built "field programmable gate arrays" will run Bing, Office 365, and Azure

    Cade Metz

    It was December 2012, and Doug Burger was standing in front of Steve Ballmer, trying to predict the future.

    Ballmer, the big, bald, boisterous CEO of Microsoft, sat in the lecture room on the ground floor of Building 99, home base for the company’s blue-sky R&D lab just outside Seattle. The tables curved around the outside of the room in a U-shape, and Ballmer was surrounded by his top lieutenants, his laptop open. Burger, a computer chip researcher who had joined the company four years earlier, was pitching a new idea to the execs. He called it Project Catapult.

    The tech world, Burger explained, was moving into a new orbit. In the future, a few giant Internet companies would operate a few giant Internet services so complex and so different from what came before that these companies would have to build a whole new architecture to run them. They would create not just the software driving these services, but the hardware, including servers and networking gear. Project Catapult would equip all of Microsoft’s servers—millions of them—with specialized chips that the company could reprogram for particular tasks.

    But before Burger could even get to the part about the chips, Ballmer looked up from his laptop. When he visited Microsoft Research, Ballmer said, he expected updates on R&D, not a strategy briefing. “He just started grilling me,” Burger says. Microsoft had spent 40 years building PC software like Windows, Word, and Excel. It was only just finding its feet on the Internet. And it certainly didn’t have the tools and the engineers needed to program computer chips—a task that’s difficult, time consuming, expensive, and kind of weird. Microsoft programming computer chips was like Coca Cola making shark fin soup.

    Burger—trim, only slightly bald, and calmly analytical, like so many good engineers—pushed back. He told Ballmer that companies like Google and Amazon were already moving in this direction. He said the world’s hardware makers wouldn’t provide what Microsoft needed to run its online services. He said that Microsoft would fall behind if it didn’t build its own hardware. Ballmer wasn’t buying it. But after awhile, another voice joined the discussion. This was Qi Lu, who runs Bing, Microsoft’s search engine. Lu’s team had been talking to Burger about reprogrammable computer chips for almost two years. Project Catapult was more than possible, Lu said: His team had already started.

    Today, the programmable chips that Burger and Lu believed would transform the world—called field programmable gate arrays—are here. FPGAs already underpin Bing, and in the coming weeks, they will drive new search algorithms based on deep neural networks—artificial intelligence modeled on the structure of the human brain—executing this AI several orders of magnitude faster than ordinary chips could. As in, 23 milliseconds instead of four seconds of nothing on your screen. FPGAs also drive Azure, the company’s cloud computing service. And in the coming years, almost every new Microsoft server will include an FPGA. That’s millions of machines across the globe. “This gives us massive capacity and enormous flexibility, and the economics work,” Burger says. “This is now Microsoft’s standard, worldwide architecture.”

    This isn’t just Bing playing catch-up with Google. Project Catapult signals a change in how global systems will operate in the future. From Amazon in the US to Baidu in China, all the Internet giants are supplementing their standard server chips—central processing units, or CPUs—with alternative silicon that can keep pace with the rapid changes in AI. Microsoft now spends between $5 and $6 billion a year for the hardware needed to run its online empire. So this kind of work is “no longer just research,” says Satya Nadella, who took over as Microsoft’s CEO in 2014. “It’s an essential priority.” That’s what Burger was trying to explain in Building 99. And it’s what drove him and his team to overcome years of setbacks, redesigns, and institutional entropy to deliver a new kind of global supercomputer.

    A Brand New, Very Old Kind of Computer Chip

    In December of 2010, Microsoft researcher Andrew Putnam had left Seattle for the holidays and returned home to Colorado Springs. Two days before Christmas, he still hadn’t started shopping. As he drove to the mall, his phone rang. It was Burger, his boss. Burger was going to meet with Bing execs right after the holiday, and he needed a design for hardware that could run Bing’s machine learning algorithms on FPGAs.

    Putnam pulled into the nearest Starbucks and drew up the plans. It took him about five hours, and he still had time for shopping.

    Burger, 47, and Putnam, 39, are both former academics. Burger spent nine years as a professor of computer science at the University of Texas, Austin, where he specialized in microprocessors and designed a new kind of chip called EDGE. Putnam had worked for five years as a researcher at the University of Washington, where he experimented with FPGAs, programmable chips that had been around for decades but were mostly used as a way of prototyping other processors. Burger brought Putnam to Microsoft in 2009, where they started exploring the idea that these chips could actually accelerate online services.

    Even their boss didn’t buy it. “Every two years, FGPAs are ‘finally going to arrive,’” says Microsoft Research vice president Peter Lee, who oversees Burger’s group. “So, like any reasonable person, I kind of rolled my eyes when this was pitched.” But Burger and his team believed this old idea’s time had come, and Bing was the perfect test case.

    Microsoft’s search engine is a single online service that runs across thousands of machines. Each machine is driven by a CPU, and though companies like Intel continue to improve them, these chips aren’t keeping pace with advances in software, in large part because of the new wave in artificial intelligence. Services like Bing have outstripped Moore’s Law, the canonical notion that the number of transistors in a processor doubles every 18 months. Turns out, you can’t just throw more CPUs at the problem.

    But on the other hand, it’s generally too expensive to create specialized, purpose-built chips for every new problem. FPGAs bridge the gap. They let engineers build chips that are faster and less energy-hungry than an assembly-line, general-purpose CPU, but customizable so they handle the new problems of ever-shifting technologies and business models.

    At that post-holiday meeting, Burger pitched Bing’s execs on FPGAs as a low-power way of accelerating searches. The execs were noncommittal. So over the next several months, Burger and team took Putnam’s Christmas sketch and built a prototype, showing that it could run Bing’s machine learning algorithms about 100 times faster. “That’s when they really got interested,” says Jim Larus, another member of the team back then who’s now a dean at Switzerland’s École Polytechnique Fédérale in Lausanne. “They also started giving us a really hard time.”

    The prototype was a dedicated box with six FPGAs, shared by a rack full of servers. If the box went on the frizz, or if the machines needed more than six FPGAs—increasingly likely given the complexity of the machine learning models—all those machines were out of luck. Bing’s engineers hated it. “They were right,” Larus says.

    So Burger’s team spent many more months building a second prototype. This one was a circuit board that plugged into each server and included only one FPGA. But it also connected to all the other FPGA boards on all the other servers, creating a giant pool of programmable chips that any Bing machine could tap into.

    That was the prototype that got Qi Lu on board. He gave Burger the money to build and test over 1,600 servers equipped with FPGAs. The team spent six months building the hardware with help from manufacturers in China and Taiwan, and they installed the first rack in an experimental data center on the Microsoft campus. Then, one night, the fire suppression system went off by accident. They spent three days getting the rack back in shape—but it still worked.

    Over several months in 2013 and 2014, the test showed that Bing’s “decision tree” machine-learning algorithms ran about 40 times faster with the new chips. By the summer of 2014, Microsoft was publicly saying it would soon move this hardware into its live Bing data centers. And then the company put the brakes on.


  5. #5
    WHT-BR Top Member
    Data de Ingresso
    Dec 2010

    A newer version of the final hardware, V2, a card that slots into the end of each Microsoft server and connects directly to the network.

    Searching for More Than Bing

    Bing dominated Microsoft’s online ambitions in the early part of the decade, but by 2015 the company had two other massive online services: the business productivity suite Office 365 and the cloud computing service Microsoft Azure. And like all of their competitors, Microsoft executives realized that the only efficient way of running a growing online empire is to run all services on the same foundation. If Project Catapult was going to transform Microsoft, it couldn’t be exclusive to Bing. It had to work inside Azure and Office 365, too.

    The problem was, Azure executives didn’t care about accelerating machine learning. They needed help with networking. The traffic bouncing around Azure’s data centers was growing so fast, the service’s CPUs couldn’t keep pace. Eventually, people like Mark Russinovich, the chief architect on Azure, saw that Catapult could help with this too—but not the way it was designed for Bing. His team needed programmable chips right where each server connected to the primary network, so they could process all that traffic before it even got to the server.

    So the FPGA gang had to rebuild the hardware again. With this third prototype, the chips would sit at the edge of each server, plugging directly into the network, while still creating pool of FPGAs that was available for any machine to tap into. That started to look like something that would work for Office 365, too. Project Catapult was ready to go live at last.

    Larus describes the many redesigns as an extended nightmare—not because they had to build a new hardware, but because they had to reprogram the FPGAs every time. “That is just horrible, much worse than programming software,” he says. “Much more difficult to write. Much more difficult to get correct.” It’s finicky work, like trying to change tiny logic gates on the chip.

    Now that the final hardware is in place, Microsoft faces that same challenge every time it reprograms these chips. “It’s a very different way of seeing the world, of thinking about the world,” Larus says. But the Catapult hardware costs less than 30 percent of everything else in the server, consumes less than 10 percent of the power, and processes data twice as fast as the company could without it.

    The rollout is massive. Microsoft Azure uses these programmable chips to route data. On Bing, which an estimated 20 percent of the worldwide search market on desktop machines and about 6 percent on mobile phones, the chips are facilitating the move to the new breed of AI: deep neural nets. And according to one Microsoft employee, Office 365 is moving toward using FPGAs for encryption and compression as well as machine learning—for all of its 23.1 million users. Eventually, Burger says, these chips will power all Microsoft services.

    Wait—This Actually Works?

    “It still stuns me,” says Peter Lee, “that we got the company to do this.” Lee oversees an organization inside Microsoft Research called NExT, short for New Experiences and Technologies. After taking over as CEO, Nadella personally pushed for the creation of this new organization, and it represents a significant shift from the 10-year reign of Ballmer. It aims to foster research that can see the light of day sooner rather than later—that can change the course of Microsoft now rather than years from now. Project Catapult is a prime example. And it is part of a much larger change across the industry. “The leaps ahead,” Burger says, “are coming from non-CPU technologies.”

    All the Internet giants, including Microsoft, now supplement their CPUs with graphics processing units, chips designed to render images for games and other highly visual applications. When these companies train their neural networks to, for example, recognize faces in photos—feeding in millions and millions of pictures—GPUs handle much of the calculation. Some giants like Microsoft are also using alternative silicon to execute their neural networks after training. And even though it’s crazily expensive to custom-build chips, Google has gone so far as to design its own processor for executing neural nets, the tensor processing unit.

    With its TPUs, Google sacrifices long-term flexibility for speed. It wants to, say, eliminate any delay when recognizing commands spoken into smartphones. The trouble is that if its neural networking models change, Google must build a new chip. But with FPGAs, Microsoft is playing a longer game. Though an FPGA isn’t as fast as Google’s custom build, Microsoft can reprogram the silicon as needs change. The company can reprogram not only for new AI models, but for just about any task. And if one of those designs seems likely to be useful for years to come, Microsoft can always take the FPGA programming and build a dedicated chip.

    Microsoft’s services are so large, and they use so many FPGAs, that they’re shifting the worldwide chip market. The FPGAs come from a company called Altera, and Intel executive vice president Diane Bryant tells me that Microsoft is why Intel acquired Altera last summer—a deal worth $16.7 billion, the largest acquisition in the history of the largest chipmaker on Earth. By 2020, she says, a third of all servers inside all the major cloud computing companies will include FPGAs.

    It’s a typical tangle of tech acronyms. CPUs. GPUs. TPUs. FPGAs. But it’s the subtext that matters. With cloud computing, companies like Microsoft and Google and Amazon are driving so much of the world’s technology that those alternative chips will drive the wider universe of apps and online services. Lee says that Project Catapult will allow Microsoft to continue expanding the powers of its global supercomputer until the year 2030. After that, he says, the company can move toward quantum computing.

    Later, when we talk on the phone, Nadella tells me much the same thing. They’re reading from the same Microsoft script, touting a quantum-enabled future of ultrafast computers. Considering how hard it is to build a quantum machine, this seems like a pipe dream. But just a few years ago, so did Project Catapult.


  6. #6
    WHT-BR Top Member
    Data de Ingresso
    Dec 2010

    Microsoft taps Intel Altera FPGAs for speedier Azure cloud services

    Tim Anderson
    26 Sep 2016

    Ignite Microsoft is using Intel Altera Field Programmable Gate Arrays (FPGA) chips to speed up Azure services, according to an announcement at the Ignite event under way in Atlanta.

    FPGA chips aim to combine the performance advantage of hardware with the flexibility of software. They are integrated circuits that can be reconfigured by downloading a new hardware configuration after manufacture, hence "Field Programmable".

    Initiated in 2010, Microsoft's Project Catapult is an effort to accelerate cloud computing through a network of FPGAs in the company's datacenters.

    Now the company has announced what it says is the world's largest deployment of FPGAs. For three years or so, Microsoft has been including an Altera FPGA in every server it installs. Altera, a specialist FPGA vendor, was acquired by Intel in 2015. This FPGA network will be used to accelerate artificial intelligence services, among other things.

    Microsoft says the current rollout is in 15 countries over five continents, though there is no detail available. Newer data centres, such as those recently opened in the UK, are likely to be suitably equipped.

    “The really important thing is how we’ve architected the system,” Microsoft Distinguished Engineer Doug Burger told The Reg. “The FPGA sits directly between the server and the network, so all the traffic goes through. The CPU can also talk to it over PCIe, but the FPGAs can talk to one another over the network as well. So in some sense it’s a new kind of computer that’s been inserted into our cloud. That layer can do networking, it can do AI, it can do other things. It is a major architectural change."

    The fact that all the network traffic goes through the FPGAs is what enables Microsoft to use the FPGAs somewhat independently of the servers, rather than always going via the host servers. At the same time, this design introduces new risks, since a bug or fault impacts the whole system. That, said Burger, has been the key challenge.

    "You are putting an alien technology into a very mature system. All of the network traffic runs through this thing. You screw it up, you can do some real damage. You think about reliability at scale, failure diagnostics, health monitoring, debugging, version management, package management, all of that needs to be built into the platform. No one has gone large scale like this."

    The benefit though is huge speed-up for certain specialist tasks. "When you have a successful FPGA deployment, the speed-ups you get tend to range between 10 and 1000 times. Usually it’s in the low 10s," said Burger.

    Might Microsoft allow developers to upload their own FPGA images to run on Azure? "That is a potential business," says Burger. "We haven't announced any plans or schedule to do that."


  7. #7
    WHT-BR Top Member
    Data de Ingresso
    Dec 2010

    A Closer Look at Intel FPGAs, which Microsoft is tapping for its Azure

    Intel begins rollout of Xeon CPUs packaged with FPGA chips to power the cloud, IoT

    By Rich Miller
    April 14, 2016

    Intel has taken a big step toward more powerful and customizable chips for servers, shipping a module that integrates its latest Xeon E5 processors with programmable chips acquired in its $16 billion deal for Altera. Combining the technologies will provide broader access to CPU acceleration techniques that have been common in high performance computing (HPC) and are making their way into the hyperscale computing ecosystem.

    The new chip combines traditional Intel CPUs with field programmable gate arrays (FPGAs), semiconductors that can be reprogrammed to perform specialized computing tasks. FPGAs allow users to tailor compute power to specific workloads or applications.

    Intel sees FPGAs as the key to designing a new generation of products to address emerging customer workloads in the data center sector, as well as the Internet of Things (IoT). FPGAs can serve as coprocessors to accelerate CPU workloads, an approach that is used in supercomputing and HPC, usually by teaming CPUs with NVIDIA GPUs or Intel’s x86-based Xeon Phi coprocessors. Altera was also a leading player in programmable logic devices (PLDs), which are widely used to automate industrial infrastructure.

    Focus on Mega Data Centers

    At the Intel Developers Forum Wednesday in Shenzhen, China, Intel executive Diane Bryant confirmed that Intel has begun a pilot program shipping development modules that package CPUs and Altera Arria 10 FPGAs, the first step on a roadmap that envisions an eventual single-chip integration. The new chipset is being shipped to existing customers for testing and development.

    The integration of FPGAs is a key strategy for Intel as it sharpens its focus on chips for large data centers. This is driven by the shift to cloud computing, which has shifted buying power from enterprises to large cloud computing platforms. Cloud infrastructure accounted for 32 percent of IT infrastructure products in the fourth quarter of 2015, according to IDC, up 4 percentage points from a year earlier. Meanwhile, revenue from equipment sales to traditional enterprises declined by 2.7 percent.

    Intel expects this trend to continue. “Seventy to 80 percent of systems will be deployed in large-scale data centers by 2025,” said Jason Waxman, vice president of the Data Center and Cloud Platforms Group at Intel.

    Hyperscale data center operators are seeking to optimize their infrastructures, and FPGAs offer more sophisticated processing architectures that can bring HPC-style parallel processing of workloads.

    “We think FPGAs are very strategic,” said Raejeanne Skillern, GM of the Cloud Service Provider Business at Intel. “A lot of the use cases fall into cloud, but also the Internet of Things. There are a lot of workloads where you have the ability to save cores and offload work.

    “We’re still at the beginning of the integration (with Altera),” she added. “We’re doing a lot of development with OEMs and customers, and continuing to implement (FPGAs) into our roadmap.”

    FPGA acceleration can be used in computing tasks related to artificial intelligence, encryption and compression, as Intel outlined in this slide from the Altera deal announcement:

    On the customer front, a particular focus is the “Super 7” group of cloud service providers that are driving hyperscale infrastructure innovation. This group includes Amazon, Facebook, Google and Microsoft, along with Chinese hyperscale companies Alibaba, Baidu and Tencent. At Intel’s investor meeting last November, Bryant said that four of the seven companies are expected to sample the new CPU/FPGA products.

    Courting the Open Compute Crowd

    That’s why Intel was a prominent presence at last month’s Open Compute Summit, where it noted that Facebook is an early development partner. Facebook recently worked closely with Intel to retool its server form factor, shifting from its traditional two-processor server to a system-on-chip (SoC) based on a single Xeon-D processor that uses less power and solves several architectural challenges. The new architecture also allows for the integration of coprocessors.

    “We wanted to build an infrastructure that consisted of a number of simple, modular building blocks, which would allow us to incorporate other compute logic (FPGAs or GPUs) in the future,” Facebook’s Vijay Rao and Edwin Smith wrote on the Facebook Engineering blog.

    At the Open Compute Summit, Waxman said Intel is developing libraries to help developers create applications that can take advantage of the additional processing capabilities.

    Dan McNamara, general manager of the new Programmable Solutions Group at Intel, said the group will “create market-leading programmable logic devices that deliver a wider range of capabilities than customers experience today … (enabling) customers to create the next generation of electronic systems with unmatched performance and power efficiency.”


  8. #8
    WHT-BR Top Member
    Data de Ingresso
    Dec 2010
    Mathew Lodge ‏@mathewlodge 5 hours ago

    “chips would sit at the edge of each server, plugging directly into the network” i.e. just like how networking vendors use FPGAs

    Mathew Lodge ‏@mathewlodge 5 hours ago

    Good grief. @wired relates how server people are discovering the FPGA about 20 years after everyone else

  9. #9
    WHT-BR Top Member
    Data de Ingresso
    Dec 2010
    Post #1

    Back in the mid-2000s, when Columbia 1 and 2 were built, server compute, storage, and power density was not what it is today, and so now, with the Microsoft’s Open Compute Server platforms installed in these facilities, only eight of the 18 rows have gear in them now, and because Microsoft has substantially virtualized its networking inside the datacenter, load balancers, switches, and routers are no longer in the facilities, either. (It is a pity that the datacenter can’t have its power distribution upped to 5.7 megawatts and be filled to the brim again.)

    How Cloud Computing Alters Data Center Design

    Many of the major data center providers in today’s market are currently inclined to center their design efforts around one big template — for instance, a 10,000 square foot, single-tenant hall with 1100 kW of UPS power, he said. Realistically, it isn’t practical for such a provider to make such a facility multi-tenant, Leonard argued.

    “So say you get a software-as-a-service company. They can only buy one thing: 10,000 square feet and 1100 kW. And on day one, that might fit their needs perfectly, or maybe they can architect their application to where that’s perfect. But what happens when they re-architect their application and their hardware, and now they consume double the watts per square foot?

    “Well, they’ve just stranded half of that space,” Leonard answers himself. “Who pays for that space that’s stranded? Well, they have to pay for it, because there’s no flexibility there.”

    Now, certain well-known data center customers — Leonard cites Akamai as one example — are moving from a 12-15 kW per rack power usage profile down to about 9 kW/rack. Service providers are capable of making such deliberate changes to their applications to enable this kind of energy efficiency.

    Suppose a hypothetical SP customer of this same data center is inspired by Akamai, re-architects its application, and lowers its power consumption. “Well, now they can’t use the power that’s in that space,” argues Leonard.

    “Creating space where power and cooling are irretrievably tied to the floor space that is being delivered on is a really bad idea. When the use of that floor space, power, and cooling changes over time — and there’s a dozen dimensions that can cause it to change — those data centers are rigid and inflexible in their ability to react to those changes.”
    Última edição por 5ms; 27-09-2016 às 12:21.

  10. #10
    WHT-BR Top Member
    Data de Ingresso
    Dec 2010

    Microsoft Has Been Secretly Sneaking AI Into Its Azure Cloud For Two Years

    Dan Rowinski
    September 27th, 2016

    For the last two years Microsoft has been sneaking artificial intelligence into its Azure cloud.

    We not just talking about Microsoft Cognitive Services, Office 365 or Cortana Intelligence. Those are just apps and services. No, Microsoft has been sneaking silicon hardware—field programmable gate arrays—into its Azure cloud without telling anyone.

    The result, Microsoft believes, is the world’s first artificial intelligence supercomputer. That’s a hefty claim … and Microsoft may be right.

    “We made a major investment in PGAs [programmable gate arrays],” said Microsoft’ Research engineer Doug Burger at the company’s Ignite conference in Atlanta this week. “PGAs are programmable hardware. You get the efficiency of hardware, but you also get the flexibility to change your functionality on the fly. This new architecture that we built effectively embeds an FPGA-based AI super computer into our old, hyper-scaled cloud. You get compute, scale and efficiency. It will change what is possible for AI.”

    “Over the past two years, quietly, we have deployed it across our global hyper-scaled data centers in 15 countries spanning five continents.”

    Field programmable gate arrays are integrated circuits, much like a CPU or a GPU is an integrated circuit. The difference with FPGAs are that they can be programmed after manufacturing to fit any variety of tasks on the fly (thus able to be programmed in the “field”). And they have the capability to run at a much higher capacity. As such, FPGAs are the perfect type of integrated circuits to add to globally scaled clouds because they don’t just offer raw computing (which they do) power, but can be attuned to any specific task that are thrown at them.

    Google also announced this year that it had built its own, dedicated, “AI chip.” Google announced at Google I/O the Tensor Processing Unit, an application specific integrated circuit (ASIC) that it started deploying in its own data centers this year. The difference between Google and Microsoft is that Microsoft, apparently, has been working on this for a couple of years. The Bing ranking engine is already running on the FPGA-powered Azure cloud.

    For instance, Burger showed off how fast the FPGAs in the Azure cloud could translate Leo Tolstoy’s War And Peace from Russian to English if only four FPGA boards (with 10 CPU units) in the cloud could handle the task. Reading text and translating it are some of the harder, more computational intensive tasks for neural networks, thus making Tolstoy’s masterpiece a lot to chew through, even for an intelligent cloud.

    The result? The Azure cloud running the full stack of FPGAs translated War And Peace in two and a half seconds.

    If the full weight (nearly every FPGA in the network) were to work to translate all of English-language Wikipedia (5 million articles, nearly 3 billion words), it would take less than a tenth of a second.

    “We now have FPGA support across every compute port of Azure,” said Microsoft CEO Satya Nadella. “It means we have the ability through the magic of the fabric of the network to distribute you machine learnings, your deep neural nets to all of the silicon that is available so you get back performance, that scale.”

    “The first AI super computer in action.”


Permissões de Postagem

  • Você não pode iniciar novos tópicos
  • Você não pode enviar respostas
  • Você não pode enviar anexos
  • Você não pode editar suas mensagens