The Anatomy Of The Google Architecture Fina Lv1.1

A comprehensive overview of Google's architecture - starting from the search page and all the way

  1. Hasan Veldstra
    A comprehensive overview of Google's architecture - starting from the search page and all the way
    Transcript Header:
    The Anatomy Of The Google Architecture Fina Lv1.1
    Transcript Body:
    • 1. GOOGLE TALK
              • Ed Austin 12-09-09
    • 2. Pre Presentation The Google Philosophy (according to ed)
      • Jedis build their own lightsabres (the MS Eat your own Dog Food)
      • Parallelize Everything
      • Distribute Everything (to atomic level if possible)
      • Compress Everything (CPU cheaper than bandwidth)
      • Secure Everything (you can never be too paranoid)
      • Cache (almost) Everything
      • Redundantize Everything (in triplicate usually)
      • Latency is VERY evil
    • 3. The Anatomy of the Google Architecture “The unofficial Version“ V1.0 November 2009 Ed Austin {ed, edik}
    • 4. Section I – The Basic Glue 1. Exterior Network (Perimeter Architecture) 2. Data Centre 3. Rack Characteristics 4. Core Server Hardware 5. Operating System Implementation 6. Interior Network Architecture
    • 5. THE PERIMETER How does your data enter the Google empire?
    • 6. Perimeter Network Security (as known)
      • DNS Load Balanced splits traffic (country, .com multiple DNS, other X1) to FW
      • Firewall filters traffic (http/s, smtp,pop etc)
      • Netscalar Load Balancers take Request from FW blocks DOS attacks, ping floods (DOS) – blocks non IPv4/6 and none 80/443 ports and http multiplexes (limited caching capability)
      • User Request forwarded to Squid (Reverse Proxy) probably HUGE cache (Petabytes?)
      • If not in Cache forwarded to GWS (Custom C++ Web Server) – now not using Custom apache?
      • GWS sends the Request to appropriate internal ( Cell ) servers
      • Request is processed
      • exterior https via thawte certs
      • Dedicated Crawler Architecture separate from other infrastructure
      • Uses Squid Reverse Proxy
      • Perimeter Cache hit rates 30-60% = Huge!
      • - Dependent on search complexity/user preferences/traffic type
          • All Image Thumbnails caches, much Multimedia cached
          • Expensive common queries cached (common words i.e. ‘Obama‘, ‘edinburgh‘) as they require significant back-end processing.
          • On cache flush/update big latency spike and capacity drop
            • Index servers need to do significant work to rebuild cache
    • 8. THE DATA CENTRE Where do they store all that Data?
    • 9. Worldwide Data Centres Where is Google Located? Last estimated were 36 Data Centers, 300+ GFSII Clusters and upwards of 800K machines. US (#1) – Europe (#2) – Asia (#3) – South America/Russia (#4) Australia – on Hold Future : Taiwan, Malaysia, Lithuania, and Blythewood, South Carolina.
    • 10. The Modular Data Centre Standard Google Modular DC (Cell) holds 1160 Servers / 250KW Power Consumption in 30 racks (40U). This is the “Atomic“ Data Centre Building Block of Google. A Data Centre would consist of 100‘s of Modular Cells. DC architecture then being the aggregation of smaller Cell level infrastructures in their own container – some being pure GFS, other BT, other Map, some mixed etc. MDC‘s can also be deployed autonomously at the Perimeter (stand alone).
    • 11. THE RACK How is a server stored in the Data Centre?
    • 12. Google Rack (GOOG rack)
        • Why interesting?
          • The rack Implementation!
          • EVERYTHING custom!
        • Mini Server Size
          • Old Servers are Custom 1U
          • New Servers are 2U...
          • again a custom design
          • seem 1/3 width of a normal 2U Server
        • 40U/80U Custom Racks (50% each side)
          • Design
          • Huge Heating and Power Issues
          • Optimized Motherboards
          • Work closely with HW MB developers
          • Have their own HW builds
          • specified to component level
          • Servers expected to be expendable –
          • build redundancy on top of failure
        • Motherboard directly mounted into Rack
          • servers have no casing - just bare boards
          • – assist with heat dispersal issues
    • 13. THE HARDWARE Millions of exactly what?
    • 14. Server Hardware
      • 2U Low-Cost (but not slow) Commodity Servers
        • 2009 Currently 2-Way, Dual Core/16GB/1-2TB +- Standard
        • Both Intel/AMD Chipsets – 1 NIC – 2 USB
        • Looks like they RAID1/mirror the disks for better I/O - read performance
        • SATA 7.2K/10K/15K drives? 8 x 2GB DDR3 ECC
      • Standard HW Build (Several HW Build Versions at any one time)
        • Currently at 7Gen Build (1G 2005 was probably Dual Core/SMP )
        • Each Server 12V Battery Backup and can run autonomously without external power (lasts 20-30s?)
        • Work closely with chip manufacturers to improve design/reduce power – custom Intel chips that can withstand higher heat factors than generic versions
      YEAR Average Server Specification 1999/2000 PII/PIII 128MB+ 2003/2004 Celeron 533, PIII 1.4 SMP, 2-4GB DRAM, Dual XEON 2.0/1-4GB/40-160GB IDE - SATA Disks via Silicon Images SATA 3114/SATA 3124 2006 Dual Opteron/Working Set DRAM(4GB+)/2x400GB IDE (RAID0?) 2009 2-Way/Dual Core/16GB/1-2TB SATA
    • 15. THE OPERATING SYSTEM The Core Software on each of those servers
      • 100% Redhat Linux Based since 1998 inception
      • - RHEL (Why not CentOS?) - 2.6.X Kernel - PAE - Custom glibc.. rpc... ipvs... - Custom FS (GFS II) - Custom Kerberos - Custom NFS - Custom CUPS
      • - Custom gPXE bootloader
      • - Custom EVERYTHING.....
      • Kernel/Subsystem Modifications
      • tcmalloc – replaces glibc 2.3 malloc – much faster! works very well with threads... rpc – the rpc layer extensively modified to provide > perf increase < latency (52%/40%)
      • Significantly modified Kernel and Subsystems – all IPv6 enabled
      • Use Python as the primary scripting language Deploy Ubuntu internally (likely for the Desktop) – also Chrome OS base Easily the Worlds largest installed Linux base
    • 17. THE INTERIOR NETWORK How does your datatravel around the Google empire?
    • 18. INTERIOR NETWORK ROUTING PROTOCOL Internal network is IPv6 (exterior machines can be reached using IPv6) Heavily Modified Version of OSPF as the IRP Intra-rack network is 100baseT Inter-rack network is 1000baseT Inter-DC network pipes unknown but very fast Technology: Juniper, Cisco, Foundry, HP, routers and switches Software: ipvs (ip virtual server)
    • 19. THE MAJOR GLUE The three foundation blocks of Googles Secret Sauce
    • 20. Section II – Googles Major Glue 1. Google File System Architecture – GFS II 2. Google Database - Bigtable 3. Google Computation - Mapreduce 4. Google Scheduling - GWQ
    • 21. GOOGLE FILE SYSTEM Manages the underlying Data on behalf of the upper layers and ultimately the applications
    • 22. FILE SYSTEM I – GFS v1 The GFS II cell is Googles fundamental building block – everything can be layered on top of this Consists of (Highly distributed Linux based) Master Servers and Chunk Servers Chunk Servers serve the Data in 64MB Chunks to the client directly via Master arbitration DATA REDUNDANCY/FAULT TOLERANCE? Triplicate Copies of Chunks are kept often in other clusters / DC Chunks can be pulled from outside the DC! Expensive.... And try not to do! However apps built on top of GFS/BT do this on an ad-hoc basis (i.e. Gmail) On Chunk loss the Master handles the Recovery by sourcing a chunk copy Data is compressed using BMDiff/Zippy Chunk Server Fault-Tolerance achieved by Heart-beat to the Master (I am alive..) Master Failure was problematic for Google (finally down from 2 minutes to 10 seconds)
    • 23. FILE SYSTEM I – GFS II GFS II “Colossus“ Version 2 improves in many ways (is a complete rewrite) Elegant Master Failover (no more 2s delays...) Chunk Size is now 1MB – likely to improve latency for serving data other than Indexing – for example GMail – this was the rationale behind the change Master can store more Chunk Metadata (therefore more chunks addressable up to 100 million) = also more Chunk Servers However according to Google Engineer they have only ever lost one 64MB chunk (in GFS I) during its entire production deployment (2004 – 2008?) so assumed extremely reliable
    • 24. GOOGLE DATABASE Accesses the underlying Data on behalf of the upper layers and ultimately the applications
    • 25. Bigtable I - Introduction What is it? Googles Database Implementation since 1994 Used internally for all large scale (Search, Indexing, GMail etc) Similar to a sharded Database implemention GOALS Huge Scalability to many PB‘s (Web Database currently around 40 Billion URL‘s) Tight Latency Highly efficient scans over Textual Data Fault Tolerant Load Balancable Eliminate Googles dependency on an external provider
    • 26. Bigtable II How is Data Referenced? Distributed Multi-Dimensional Sparse Map Simple addressing model using a triple: (row, column, { timestamp } ) -> cell contents ROWS - Rows (arbitrary length usually 10-100 Bytes Max 27. Bigtable III – Table Structure
      • Studying contents: column shows three versions of contents of a page (current, cached and ?) – presumably all other columns are timestamped so could be used in a comparitive way (such as anchor increase/decrease) OTF in SERPS – alg must use a combo of TimeSt diff between n(=3 rest garbage collected) page Versions and crawled anchors - what else does table hold? Possibly PR (or OTF) and other search related weightings
      • Google keeps much more info for ranking purposes than it did in 1999
      • Webtable hinted at 100 columns+!
      • How do page units affect the URL reversal of the URL bigtable?
      • Does a Tables Tablets Cross a Clusters namespace (yes if unified else no?)
      C++ Bigtable Mutate of some Anchors //open table Table *T=OpenOrDie(“/ bigtable/web/bigtable “); //write new anchor and delete old anchor RowMutation r1(T,““); r1.Set (““,“CNN“); r1.Delete (““); Operation op; Apply (&op, &r1); //atomic mutate to the columns Other primitives such as DeleteCells(), DeleteRow(), Scanner (read arbitrary cells in a row)
    • 28. Bigtable IV
      • How tables are broken down in storage ?
      • For example Webtable is billions of pages!
      • Large Tables broken (split) into tablets at row boundaries
      • Tablets discontiguous (assists in fault-tolerance) – spread over DC but try to keep one copy in same rack
      • Tablet Size is approximately 100-200MB of compressed Data
      • Load Balanced – migrate tablets from heavily loaded machines to lightly loaded ones
      • Heavily used tablets probably stay in working set (cached)
    • 29. GOOGLE MAPREDUCE Computes the underlying Data on behalf of the applications
    • 30. Mapreduce I Map Reduction can be seen as a way to exploit massive parallelism by breaking a task down into constituent parts and executing on multiple processors The Major Functions are MAP & REDUCE (with a number of intermediatary steps) MAP Break task down into parallel steps REDUCE Combine results into final output Shown is a 2-pipeline Map Reduction (There are 24 Map Reductions in the indexing pipeline) Mappers & Reducers usually run on separate processors (90% loss of reducers job still completed!)
    • 31. Mapreduce II
      • C++, Java, Python, Sawzall
      • DEPLOYED
      • Implemented 2004 – before this MySQL?
      • In September 2009 Google ran 3,467,000 MR Jobs with an average 475 sec completion time averaging 488 machines per MR and utilising 25.5K Machine years
      • Technique extensively used by Yahoo with Hadoop (similar architecture to Google) and Facebook (since ‘06 multiple Hadoop clusters, one being 2500CPU/1PB with HBase).
    • 32. Chubby Lock
      • Googles Distributed File Locking Service for Bigtable
      • Provides Mutex Support for Data Access (atomic access to column data)
      • Used to synchronize access to shared resources
      • Consists of a Master and Slaves (designated by election)
      • Failover consists of a Slave replacing the functionality of a Master
      • - Also servers as an ultra-fast high availability File Server for small fines (100‘s bytes)
      • Provides an ACL for tablet authentication (row and column data)
    • 33. GOOGLE WORKQUEUE Provides Resource Management for the Computational Jobs
    • 34. GWQ – Google Workqueue
      • Batch Submission/Scheduler System
      • Software to submit Mapreduce Jobs to a Cell/Cluster
      • Arbitrates (process priorities) Schedules, Allocates Resources, process failover, Reports status, collects results - Often Workqueue overlaid on a GFS Cluster - i.e. GFS cluster not computational bound jobs – also seems to match co-locate tasks near data = just disk I/O not Network I/O (on the Chunk Server?)
      • Workqueue can manage many tens of thousands of machines Launched via API or command line (sawzall example shown) saw --program code.szl --workqueue testing --input_files /gfs/cluster1/2005-02-0[1-7]/submits.* --destination /gfs/cluster2/$USER/output@100
    • 35. Section III – Some more Glue 1. Languages employed 2. Development Environment 3. Google App Engine 4. Network Security 5. Future Google Architecture Advances 6. Odds n Sods 7. DIY Google
    • 36. DEVELOPMENT LANGUAGES - Initially Python, Java, C++ Usual Suspects - Sawzall (since 2006) - equivalent to Hadoops Pig Latin - written in C++ - interpreted bytecode output JIT‘d An internal Procedural language employed to solve map reduction problems. The few published Google papers employ Sawzall in the algorithm examples. Runs in the Map phase, Aggregators run in the Reduce phase (from each Sawzall Map instance) to get the final output. - Transparent Parallelization – no specialist Distrib Sys Knowledge Required (Good for developer) - Simple Datatypes 64-bit signed int, float, string, byte and a few unique such as time - Much STR regexp support - Compound Types arrays, tuples - typesafed (and declarations) similar to Pascal (Probably an LL(1) lang?) - similar to Algol, C Syntax (no pointers though!) - No Processing of exceptions (no exception handlers) - Shorter than corresponding C++ code by a factor of 10 Early versions could not write into Bigtable. Now implemented? Output sometimes pipelined into MySQL for further analysis
    • 37. GOOGLE APP ENGINE Using “Application Platform technology stack“
      • Allows a developer to leverage components of Google Technology (but not necessarily primary Infrastructure i.e. The usual business resources)
      • Supports Python, Java
      • - Bigtable support (via GQL)
      • Uses GFS as underlying FS – usual Fault-tolerance/Load-balancing
      • Task Queue similar to GWQ?
      • Code exposed to Google
      • - No support for subprocess spawning – more importantly none of the google mapreduce library made available - isolates computational aspects to single servers but the I/O is probably the google standard implementation underneath - therefore computationally intensive tasks more problematic = keeping your resource usage under control
    • 38. Security Rack Board Level (possible scenario) gPXE on the board goes through DHCP/tftp sequence to pull over an encrypted image (this is not expensive as is done once per boot and boots are not usual) Image is pulled from a Secure Image Distribution Server (and held encrypted on these) Once at the board end the image is OTF decrypted and booted as normal RHEL 02/09 Google Engineer didn‘t dispute this and seemed to concur adding that in-core encryption might be a possibility (R/T decryption might not be that expensive) – this possibily means cryptology is used throughout the lifetime of the image – including components outside the working-set but sensitive parts of the in-core OS (OTF decrypted) Enterprise Kerberos is used throughout the enterprise They have an Automated issuance system for SSL certificates, used by internal (secure) infrastructure to validate https/TLS and generic SSL connections . Complete internal network encryption unlikely due to latency introduced? Likely that one of the reasons failover between DC‘s problematic is the latency introduced due to the expense of Wide Area Encryption (essential)
    • 39. Google Future Architecture
      • - 99%ile latency for all data 48.
        • Internal Linux development and deployment
        • Served as technical lead of team responsible for customizing and deploying Linux to internal systems and workstations.
        • Fixed bugs and added enterprise features to several Linux components, including NFS, Kerberos, CUPS. All relevant patches were pushed to upstream maintainers, and most are in current released distributions.
        • Developed and maintained systems to automate installation, updates, and upgrades of Linux systems.
        • Developed IPv6 support for Linux load-balancing (ipvs).
        • Managed several interns and contractors.
        • loadbalancing user accounts within a datacenter, and coordinating with the global loadbalancer , which uses linear programming to optimally allocate users. In particular, this avoids "shared fate" risks and reduces latency and costs incurred due to excessive transatlantic data traffic. Learned Sketchup so as to document the four dimensional data structures effectively
        • The testing, evalulation, deployment, operations, and maintenance of Netscaler load balancers.
        • automated Apache configuration reloader
        • gPXE open-source network booting software
        • GWS – custom C++ webserver = not apache?
        • Google 02/09 talk example was a Cluster is 30 racks (I believe this refers to Google). At a 40U rack 40Ux30racks = 1200 = approximately a MDC – can assume each MDC is a Cluster/cell at architectural level
        • Google engineer stated a DC is a collection of Modular Units (MDC‘s?) – the picture (not above) illustrated suggested this.
      • 49. Some Pre Presentation Information
        • 1 Million GB = 1000TB = 1 PB (x 1000 = 1 EXABYTE)
          • Internet Archive is around 3PB (2009)
          • CLEAN UP BEFORE – all the poorly sourced stuff
          • Add lock service to bt to all slides
          • Google rack server on rack page
          • SSTable
          • Google PROFITS US $16M A DAY
      • 50. Pre Presentation Disclaimer
        • Put together in a week from knowing zero about Google
        • I am not associated with Google
        • Numbers are approximate but certainly are ball-park – Google often delivers contradictory figures and uses many terms for some items - cell /cluster – scheduler/ workqueue (obfuscation?)
        • Googles philosophy/paranoia of tell as little as possible (pausing presenters and sideways answers) makes it hard to fill in some (significant) gaps – inferences are sometimes drawn (in red)
        • Google seem to design absolutely EVERYTHING themselves – from HW MB build, Racks, Switches(?), Software... So its hard to find sources of information beyond broad concepts
      • 51. Bigtable VI
        • Latest (or at least since 2006..)
        • Increased Scalability (across Namespace/Datacenters)
          • i.e. Tablets spread over DC‘s for a table but expensive (both computationally and financially!)
        • Service Clusters (?)
        • Multiple Bigtable Clusters replicated throughout DC
        • Current Status
        • - Many Hundreds may be thousands of Bigtable Cells - Late 2009 stated 500 Bigtable clusters
        • - At minimum scaled to many thousands of machines per cell in production - Cells manage Managing 3-figure TB data (0.X PB)
    View More

Recent Reviews

  1. Shweta Murarka
    Shweta Murarka
    Version: 10/12/2009
    Good work:)i can say its worthy.thanx
  2. Andrii Vozniuk
    Andrii Vozniuk
    Version: 10/12/2009
    Very well structured presentation! Thanks for sharing!
  3. Goran Matic
    Goran Matic
    Version: 10/12/2009
    Very interesting - i was always curious about the Google technology stack, and how it is structured; thanks!
  4. Archana Thakkar
    Archana Thakkar
    Version: 10/12/2009
    Google is being run by Indians, managerially and technically. Even though Page and Schmidt are CEO and Executive Chairman of Big G, but still we can’t forget that it was Amit Singhal, an IIT Roorkey Graduate, who re-wrote the whole algorithm of Google Search Engine in 2000 which made Google the best in the industry. Then, Nikesh Arora of BHU-IT is the Chief Business Manager; Vic Goundotra is the man behind the whole Google Plus… and, many many more. Search FAMOUS INDIANS WORKING IN GOOGLE for more details.
  5. Shekh Morshed Akther
    Shekh Morshed Akther
    Version: 10/12/2009
    great job.
  6. kylexlau
    Version: 10/12/2009
    not a googler.
  7. oxgen
    Version: 10/12/2009
    Good work~:)3x
  8. Piergiorgio Borgogno
    Piergiorgio Borgogno
    Version: 10/12/2009
  9. liu feng
    liu feng
    Version: 10/12/2009
  10. Sakthivel Rajan
    Sakthivel Rajan
    Version: 10/12/2009
    Nice Presentation on Google's Architecture.