The Power Behind Google

By Michael Feldman

January 6, 2006

“Few Web services require as much computation per request as search engines. On average, a single query on Google reads hundreds of megabytes of data and consumes tens of billions of CPU cycles. Supporting a peak request stream of thousands of queries per second requires an infrastructure comparable in size to that of the largest supercomputer installations.” So begins the description of the computational power required for Google's Web search engine in “Web Search For A Planet: The Google Cluster Architecture”, a publication of the IEEE Computer Society.

Though we most often associate high performance computing with high-end mathematical applications such as climate modeling, fusion reaction simulations or quantum chromodynamics, one of the most common forms of data intensive computing is Web searching. It is used by millions of people worldwide on a daily basis. Today, Web searching is so ubiquitous that most of us now take it for granted that we can find the answer to just about anything on the Internet. But for the average person, it's hard to imagine the computational resources required to search and analyze petabytes of data millions of times per day.

So what is the nature of Google's computing infrastructure? How is it able to process thousands of queries per second from all over the world? And how will its infrastructure scale as the Web continues to grow? HPCwire recently spoke with Jeffrey Dean, Google Fellow in the Systems Infrastructure Group, to get the answers to these and other questions.

HPCwire: Tell us a little bit about your background and why you came to Google.

Dean: I received a B.S. in computer science and economics from the University of Minnesota. I then worked for a year and a half for the World Health Organization's Global Programme on AIDS, developing software for modeling the impact of the AIDS pandemic, before going to grad school. I received a Ph.D. in computer science from the University of Washington, doing research in high-level compiler optimizations for object-oriented languages. After graduating, I went to work for Digital Equipment Corporation's Western Research Lab (DEC WRL), where I worked on a variety of projects, including low-overhead profiling systems, some microprocessor architecture work, and some work on web-based information retrieval.

I joined Google in mid-1999 because my work at DEC WRL on information retrieval whet my appetite for working in that area. I knew a couple of people at Google, and figured it would be a fun place to work. I've always enjoyed working on problems that span a pretty wide range of computer science disciplines, and working on large-scale search engines is one of the best ways of doing that, because it requires solving problems across a really broad range of topics, including low-level system design, distributed systems, data compression, information retrieval, machine learning, user interfaces, with lots of general algorithmic problems thrown in at every turn.

HPCwire: Could you briefly describe the Google computing infrastructure and its rationale?

Dean: When designing our computing clusters, we place a great deal of emphasis on what sort of systems will give us the best price/performance. Search applications are relatively easy to parallelize, both within processing of a single query (by partitioning the index across machines), and across queries (by replicating each piece of the index across multiple machines and having each replica serve a fraction of the total traffic). Given this easy parallelism, the price/performance argument leads towards using clusters of large numbers of commodity PCs, that is, x86 processors, inexpensive hard drives, commodity Ethernet networking, etc.

Our clusters typically are composed of several thousands of these commodity machines, all connected via commodity Ethernet. The individual machines typically have gigabit NICs, and groups of around 40 machines are connected to commodity gigabit Ethernet switches. These switches are then connected into a large-scale core switch for the cluster, using a small number of Gigabit connections per group of 40 machines. So, our bisection bandwidth is considerably less than a gigabit per machine. More bandwidth would be great, but it's obviously considerably more expensive, and our current configuration is the current sweet spot given our applications.

We don't disclose the exact number of clusters that we have, but we do have many of them around the world at various locations. There are a couple of reasons for this. First, we can given our users faster response times by using software that tries to direct user queries towards a cluster that is located nearby to that user in terms of network latency. Second, it makes our search service considerably more robust. We try to have a bit of extra capacity at all times, so that we can turn off clusters at various points, either for planned events like hardware or network upgrades, or because we need to quickly turn off serving from a particular cluster — for example if one of the core network switches fails.

Having lots of relatively small machines means that you get a lot more bang for the computing dollar, but it also means that the machines are less reliable than more expensive machines, and because there are so many of them, higher-level software has to be designed to tolerate failures — with thousands of machines, machine failures happen many times per day. Our software is designed to assume that the hardware can fail.  Once you do that, it becomes fairly simple to deal with a lot of failures. Our serving systems generally have multiple replicas for each piece of the system to provide fault tolerance to individual machine failures.

We've also designed our own file system, the Google File System (GFS), to reliably store large amounts of data on large clusters of machines. [For information on GFS visit http://labs.google.com/papers/gfs.html.]

Finally, when you're doing large-scale data processing, it's important not to separate the storage from where you're going to do the computation. You don't need really high-end storage arrays with massive amounts of bandwidth to process a large amount of data. If you do the scheduling right, you can read the data from local disks on thousands of machines, simultaneously. By doing this, you can attain really good bandwidth from low-end storage systems with slightly clever software.  So rather than moving the data to the machine, we try to move the computation to the data.

HPCwire: What kinds of new infrastructure is Google looking at, in the near-term, to improve its price/performance?

Dean: Given the fact that our applications can be easily parallelized, chip multiprocessors (CMPs) look very attractive to us, compared with processors that go to great lengths to extract the highest single-thread performance possible.

My colleague, Luiz Barroso, has written up a nice article describing why CMP processors look very attractive for our applications [see http://labs.google.com/papers/priceofperformance.html].

As always, we're continually evaluating and refining our hardware infrastructure to explore which solutions provide the most attractive price/performance for our applications, but we are excited about the initial CMP processors coming out, as we feel their emphasis on high throughput for parallel applications rather than single-thread performance is a good match for our applications.

HPCwire: Could you describe the MapReduce model and implementation and how it is being used within Google?

Dean: MapReduce is a system originally developed by myself and my colleague, Sanjay Ghemawat, as a way of describing computations that want to process input data to compute some derived data. The general programming model is to break the computation down into two distinct phases: a Map phase, and a Reduce phase. Users specify a Map function that processes a key/value pair to generate a set of intermediate key/value pairs, and a Reduce function that merges all intermediate values associated with the same intermediate key. The basic idea is similar to the Map and Reduce primitives found in LISP and many other functional languages.

What makes it interesting is that we've developed a MapReduce library that is able to take programs written in this style and make them run on clusters of hundreds or thousands of machines. The underlying library takes care of lots of the messy details that arise when running very large-scale parallel computations, including automatically parallelizing the computation, deciding which machines should work on which pieces (including biasing those scheduling decisions to consider data locality), and handling what happens when machines fail (with long-running jobs that execute on thousands of machines, machines failures happen with some regularity). It also does things like scheduling multiple copies of the same pieces of work towards the end of the computation, to minimize the job completion time (whichever copy finishes first “wins”), and collects progress information on a centralized status page to make it easy to monitor a MapReduce computation.

We've been pleasantly surprised at how applicable the general MapReduce model has been to a wide variety of problems: it's being used internally at Google in areas as diverse as our core crawling and indexing system, data mining, statistical machine translation, our advertising systems, processing of satellite imagery, etc. It makes it relatively easy for people within Google to write relatively simple code and have that code run reasonably efficiently on thousands of machines. In a typical day at Google, thousands of different MapReduce computations are run with hundreds of distinct Map and Reduce functions across our various computational clusters.

HPCwire: Is MapReduce something you would make publicly available?

Dean: We don't currently make it available. We've had thoughts about it. It would be a moderate amount of effort on our part to divorce it from other pieces of our software, such as our cluster scheduling system. Also, we put a fair amount of effort into it and it's not clear that we would want to make it available to our competitors. At the same time, we feel like there are a lot of academic projects that would benefit from having access to something like this. So, in the future, I wouldn't be surprised if we did make it available in some form. [For more information about MapReduce visit http://labs.google.com/papers/mapreduce.html.]

HPCwire: As the capacity of the Web grows, how is the Google infrastructure going to change? Is the current model scalable for the foreseeable future? And besides Web growth, what other kinds of changes do you foresee that will create infrastructure challenges for Google?

Dean: Our software infrastructure undergoes fairly rapid evolution in response to changes in the underlying hardware platform and also our desire to scale our system in a variety of dimensions. For example, for Web search, the major dimensions are the number of documents searched, the number of user queries we need to handle, and the speed with which the index is updated. We're also usually looking to introduce new capabilities in our infrastructure, for example, the ability to examine more information about documents when making ranking decisions, the ability to quickly try out new ideas for improving our ranking algorithms, etc. In designing a system, one tries to anticipate scaling in these various dimensions, but a given design really works well only when the design parameters are within one or two orders of magnitude of the original design goals. Beyond that, the level of scaling can change the design; solutions that weren't feasible originally suddenly become very attractive, and this often leads to significant redesigns of pieces of the system. As a consequence of this, in the six and a half years that I've been at Google, our query serving [software] infrastructure has undergone fairly radical changes at least five times. I expect this to continue to be true in the future.

In terms of general infrastructure, we place quite a bit of emphasis on developing tools and infrastructure to make it easier to develop new and interesting products. GFS and MapReduce are a couple of examples. We also have a number of internal tools that help with understanding performance bottlenecks in large-scale distributed systems.

Many of our newer products, like Gmail and Google Earth, have fairly different characteristics than Web searching, and it's important to have infrastructure and system building blocks that meet the needs of a diverse set of products that we want to offer, and to make it easy to develop new applications and services. One example is that we saw a need in many of our products to manage large amounts of semi-structured mutable data in interesting ways, and to help with that, we're developing BigTable, a large-scale distributed storage system for managing semi-structured data.

In the future, a combination of both new products and our goals for pushing our current products in new and interesting directions will guide our decisions about the right software tools and infrastructure to build. These are very exciting times to be working on large-scale systems and products at Google.

Subscribe to HPCwire's Weekly Update!

Be the most informed person in the room! Stay ahead of the tech trends with industy updates delivered to you every week!

At SC19: What Is UrgentHPC and Why Is It Needed?

November 14, 2019

The UrgentHPC workshop, taking place Sunday (Nov. 17) at SC19, is focused on using HPC and real-time data for urgent decision making in response to disasters such as wildfires, flooding, health emergencies, and accidents. We chat with organizer Nick Brown, research fellow at EPCC, University of Edinburgh, to learn more. Read more…

By Tiffany Trader

China’s Tencent Server Design Will Use AMD Rome

November 13, 2019

Tencent, the Chinese cloud giant, said it would use AMD’s newest Epyc processor in its internally-designed server. The design win adds further momentum to AMD’s bid to erode rival Intel Corp.’s dominance of the glo Read more…

By George Leopold

NCSA Industry Conference Recap – Part 1

November 13, 2019

Industry Program Director Brendan McGinty welcomed guests to the annual National Center for Supercomputing Applications (NCSA) Industry Conference, October 8-10, on the University of Illinois campus in Urbana (UIUC). One hundred seventy from 40 organizations attended the invitation-only, two-day event. Read more…

By Elizabeth Leake, STEM-Trek

Cray, Fujitsu Both Bringing Fujitsu A64FX-based Supercomputers to Market in 2020

November 12, 2019

The number of top-tier HPC systems makers has shrunk due to a steady march of M&A activity, but there is increased diversity and choice of processing components with Intel Xeon, AMD Epyc, IBM Power, and Arm server ch Read more…

By Tiffany Trader

Intel AI Summit: New ‘Keem Bay’ Edge VPU, AI Product Roadmap

November 12, 2019

At its AI Summit today in San Francisco, Intel touted a raft of AI training and inference hardware for deployments ranging from cloud to edge and designed to support organizations at various points of their AI journeys. The company revealed its Movidius Myriad Vision Processing Unit (VPU)... Read more…

By Doug Black

AWS Solution Channel

Making High Performance Computing Affordable and Accessible for Small and Medium Businesses with HPC on AWS

High performance computing (HPC) brings a powerful set of tools to a broad range of industries, helping to drive innovation and boost revenue in finance, genomics, oil and gas extraction, and other fields. Read more…

IBM Accelerated Insights

Help HPC Work Smarter and Accelerate Time to Insight

 

[Attend the IBM LSF & HPC User Group Meeting at SC19 in Denver on November 19]

To recklessly misquote Jane Austen, it is a truth, universally acknowledged, that a company in possession of a highly complex problem must be in want of a massive technical computing cluster. Read more…

SIA Recognizes Robert Dennard with 2019 Noyce Award

November 12, 2019

If you don’t know what Dennard Scaling is, the chances are strong you don’t labor in electronics. Robert Dennard, longtime IBM researcher, inventor of the DRAM and the fellow for whom Dennard Scaling was named, is th Read more…

By John Russell

Cray, Fujitsu Both Bringing Fujitsu A64FX-based Supercomputers to Market in 2020

November 12, 2019

The number of top-tier HPC systems makers has shrunk due to a steady march of M&A activity, but there is increased diversity and choice of processing compon Read more…

By Tiffany Trader

Intel AI Summit: New ‘Keem Bay’ Edge VPU, AI Product Roadmap

November 12, 2019

At its AI Summit today in San Francisco, Intel touted a raft of AI training and inference hardware for deployments ranging from cloud to edge and designed to support organizations at various points of their AI journeys. The company revealed its Movidius Myriad Vision Processing Unit (VPU)... Read more…

By Doug Black

IBM Adds Support for Ion Trap Quantum Technology to Qiskit

November 11, 2019

After years of percolating in the shadow of quantum computing research based on superconducting semiconductors – think IBM, Rigetti, Google, and D-Wave (quant Read more…

By John Russell

Tackling HPC’s Memory and I/O Bottlenecks with On-Node, Non-Volatile RAM

November 8, 2019

On-node, non-volatile memory (NVRAM) is a game-changing technology that can remove many I/O and memory bottlenecks and provide a key enabler for exascale. That’s the conclusion drawn by the scientists and researchers of Europe’s NEXTGenIO project, an initiative funded by the European Commission’s Horizon 2020 program to explore this new... Read more…

By Jan Rowell

MLPerf Releases First Inference Benchmark Results; Nvidia Touts its Showing

November 6, 2019

MLPerf.org, the young AI-benchmarking consortium, today issued the first round of results for its inference test suite. Among organizations with submissions wer Read more…

By John Russell

Azure Cloud First with AMD Epyc Rome Processors

November 6, 2019

At Ignite 2019 this week, Microsoft's Azure cloud team and AMD announced an expansion of their partnership that began in 2017 when Azure debuted Epyc-backed instances for storage workloads. The fourth-generation Azure D-series and E-series virtual machines previewed at the Rome launch in August are now generally available. Read more…

By Tiffany Trader

Nvidia Launches Credit Card-Sized 21 TOPS Jetson System for Edge Devices

November 6, 2019

Nvidia has launched a new addition to its Jetson product line: a credit card-sized (70x45mm) form factor delivering up to 21 trillion operations/second (TOPS) o Read more…

By Doug Black

In Memoriam: Steve Tuecke, Globus Co-founder

November 4, 2019

HPCwire is deeply saddened to report that Steve Tuecke, longtime scientist at Argonne National Lab and University of Chicago, has passed away at age 52. Tuecke Read more…

By Tiffany Trader

Supercomputer-Powered AI Tackles a Key Fusion Energy Challenge

August 7, 2019

Fusion energy is the Holy Grail of the energy world: low-radioactivity, low-waste, zero-carbon, high-output nuclear power that can run on hydrogen or lithium. T Read more…

By Oliver Peckham

Using AI to Solve One of the Most Prevailing Problems in CFD

October 17, 2019

How can artificial intelligence (AI) and high-performance computing (HPC) solve mesh generation, one of the most commonly referenced problems in computational engineering? A new study has set out to answer this question and create an industry-first AI-mesh application... Read more…

By James Sharpe

Cray Wins NNSA-Livermore ‘El Capitan’ Exascale Contract

August 13, 2019

Cray has won the bid to build the first exascale supercomputer for the National Nuclear Security Administration (NNSA) and Lawrence Livermore National Laborator Read more…

By Tiffany Trader

DARPA Looks to Propel Parallelism

September 4, 2019

As Moore’s law runs out of steam, new programming approaches are being pursued with the goal of greater hardware performance with less coding. The Defense Advanced Projects Research Agency is launching a new programming effort aimed at leveraging the benefits of massive distributed parallelism with less sweat. Read more…

By George Leopold

AMD Launches Epyc Rome, First 7nm CPU

August 8, 2019

From a gala event at the Palace of Fine Arts in San Francisco yesterday (Aug. 7), AMD launched its second-generation Epyc Rome x86 chips, based on its 7nm proce Read more…

By Tiffany Trader

D-Wave’s Path to 5000 Qubits; Google’s Quantum Supremacy Claim

September 24, 2019

On the heels of IBM’s quantum news last week come two more quantum items. D-Wave Systems today announced the name of its forthcoming 5000-qubit system, Advantage (yes the name choice isn’t serendipity), at its user conference being held this week in Newport, RI. Read more…

By John Russell

Ayar Labs to Demo Photonics Chiplet in FPGA Package at Hot Chips

August 19, 2019

Silicon startup Ayar Labs continues to gain momentum with its DARPA-backed optical chiplet technology that puts advanced electronics and optics on the same chip Read more…

By Tiffany Trader

Crystal Ball Gazing: IBM’s Vision for the Future of Computing

October 14, 2019

Dario Gil, IBM’s relatively new director of research, painted a intriguing portrait of the future of computing along with a rough idea of how IBM thinks we’ Read more…

By John Russell

Leading Solution Providers

ISC 2019 Virtual Booth Video Tour

CRAY
CRAY
DDN
DDN
DELL EMC
DELL EMC
GOOGLE
GOOGLE
ONE STOP SYSTEMS
ONE STOP SYSTEMS
PANASAS
PANASAS
VERNE GLOBAL
VERNE GLOBAL

Intel Confirms Retreat on Omni-Path

August 1, 2019

Intel Corp.’s plans to make a big splash in the network fabric market for linking HPC and other workloads has apparently belly-flopped. The chipmaker confirmed to us the outlines of an earlier report by the website CRN that it has jettisoned plans for a second-generation version of its Omni-Path interconnect... Read more…

By Staff report

Kubernetes, Containers and HPC

September 19, 2019

Software containers and Kubernetes are important tools for building, deploying, running and managing modern enterprise applications at scale and delivering enterprise software faster and more reliably to the end user — while using resources more efficiently and reducing costs. Read more…

By Daniel Gruber, Burak Yenier and Wolfgang Gentzsch, UberCloud

Dell Ramps Up HPC Testing of AMD Rome Processors

October 21, 2019

Dell Technologies is wading deeper into the AMD-based systems market with a growing evaluation program for the latest Epyc (Rome) microprocessors from AMD. In a Read more…

By John Russell

Rise of NIH’s Biowulf Mirrors the Rise of Computational Biology

July 29, 2019

The story of NIH’s supercomputer Biowulf is fascinating, important, and in many ways representative of the transformation of life sciences and biomedical res Read more…

By John Russell

Xilinx vs. Intel: FPGA Market Leaders Launch Server Accelerator Cards

August 6, 2019

The two FPGA market leaders, Intel and Xilinx, both announced new accelerator cards this week designed to handle specialized, compute-intensive workloads and un Read more…

By Doug Black

When Dense Matrix Representations Beat Sparse

September 9, 2019

In our world filled with unintended consequences, it turns out that saving memory space to help deal with GPU limitations, knowing it introduces performance pen Read more…

By James Reinders

With the Help of HPC, Astronomers Prepare to Deflect a Real Asteroid

September 26, 2019

For years, NASA has been running simulations of asteroid impacts to understand the risks (and likelihoods) of asteroids colliding with Earth. Now, NASA and the European Space Agency (ESA) are preparing for the next, crucial step in planetary defense against asteroid impacts: physically deflecting a real asteroid. Read more…

By Oliver Peckham

Cerebras to Supply DOE with Wafer-Scale AI Supercomputing Technology

September 17, 2019

Cerebras Systems, which debuted its wafer-scale AI silicon at Hot Chips last month, has entered into a multi-year partnership with Argonne National Laboratory and Lawrence Livermore National Laboratory as part of a larger collaboration with the U.S. Department of Energy... Read more…

By Tiffany Trader

  • arrow
  • Click Here for More Headlines
  • arrow
Do NOT follow this link or you will be banned from the site!
Share This