Lei Xu HADOOP DISTRIBUTED FILE SYSTEM 1 Brief Introduction Hadoop An apache project for data-intensive applications Typical application: Map-Reduce (OSDI’04), a distributed algorithm for massive-data computation Crawl and index web pages (Y!) Analyze popular topics and trends (Twitter) Led by Yahoo!/Facebook/Cloudera 2 Brief Introduction (cont’d) Hadoop Distributed File System (HDFS) A scalable distributed file system to serve Hadoop MapReduce applications Borrow the essential ideas from the Google File System Sanjay Ghenawat, Howard Gobioff and Shun-Tak Leung. The Google File System. 19TH ACM Symposium on Operating System Principles (SOSP’03) Share same design assumptions 3 Google File System A scalable distributed file system designed for: Data-intensive applications (mainly MapReduce) Web page indexing Then it has spread to other applications E.g. Gmail, Big Table, App Engine Fault-tolerant Low-cost hardware High throughputs 4 Google File System (cont’d) Departure from other file system assumptions Run on top of the commodity hardware Component failures are common Files are huge Basic block size 64~128 MB 1~64KB in traditional file systems (Ext3/NTFS and etc.) Massive-data/data-intensive processing Large streaming read and small random read Large, sequential writes No (or bare) random writes 5 Hadoop DFS Assumptions Other than the assumptions in Google File System, HDFS assumes that: Simple Coherency Model Write-once-read-many Once a file was created, written and closed, it can not be changed anymore. Moving Computation Is Cheaper than Moving Data “Semi-Location-Aware” computation Try its best to assign computations closer to the related data Portability Across Heterogeneous Hardware and Software Platforms Is written in Java, multi-platform support Google File System was written in C++ and run on Linux Store data on top of existing file systems (NTFS/Ext4/Btrfs…) 6 HDFS Architecture Master/Slave Architecture NameNode Metadata Server File location ( file name -> the DataNode ) File attributions (atime/ctime/mtime, size, the number of replicas and etc.) DataNode Manages the storage attached to the nodes that they run on Client Producer and Consumers of data 7 HDFS Architecture (cont’d) 8 NameNode Metadata Server Only one NameNode in one cluster Single Point Failure Potential performance bottleneck Manage the file system namespace Traditional hierarchical namespace Keep all file metadata in memory for fast access The memory size of NameNode determines how many files can be supported Execute file system namespace operation: Open/close/rename/create/unlink… Return the location of data blocks 9 NameNode (cont’d) Maintains system-wide activities E.g. creating new replications of file data, garbage collection, load balancing and etc. Periodically communicates with DataNode to collect their statuses Is DataNode alive? Is DataNode overload? 10 DataNode Storage server Store fixed-size data blocks on local file systems ( ext4/zfs/btrfs ) Serve read/write operations from the clients Create, delete, replicate data blocks upon instruction from the NameNode Block size = 64MB 11 Client Application-level implementations Does not provide POSIX API Hadoop has a FUSE interface FUSE: Filesystem in Userspace Has limited functions (e.g, no random write supports) Query the NameNode for file locations and metadata Contact corresponding DataNodes for file I/Os 12 Data Replication Files are stored as a sequence of blocks The blocks (typically 64MB) are replicated for fault tolerance Replication factor is configurable per file Can be specified at creation time, and can be changed later The NameNode decides how to replicate blocks. It periodically receives: Heartbeat, which implies the DataNode is alive Blockreport, which contains a list of all blocks on a DataNode When a DataNode is down, the NameNode replicas all blocks on this DataNode to other active DataNode to achieve enough replications 13 Data Replication (cont’d) 14 Data Replication (cont’d) Rack Awareness Hadoop instance runs on a cluster of computers that spread across many racks: Nodes in same rack are connected by one switches Communications between two nodes in different racks go through switches Slower than nodes in same rack One rack may fail due to network/power issues. Improve data reliability, availability and network bandwidth utilization 15 Data Replications (cont’d) Rack Awareness (cont’d) For common case, the replication factor is three Two replicas are placed on two different nodes in same rack The third replica is placed on a node in a remote rack Improves write performance 2/3 writes are in same rack, faster Without compromising data reliability 16 Replica Selection For READ operation: Minimize the bandwidth consumption and latency Prefer nearer node: If there is a replica on the same node, it is preferred The cluster may span multiple data centers, replicas in same data centers are preferred 17 Filesystem Metadata The HDFS stores all file metadata on NameNode An EditLog Record every change that occurs to filesystem metadata For failure recovery Same as journaling file systems (Ext3/NTFS) An FSImage Stores mapping of blocks to files and file attributes EditLog and FSImage are stored on NameNode locally 18 Filesystem Metedata(cont’d) DataNode has no knowledge about HDFS files It only stores data blocks as regular files on local file systems With a checksum for data integrity It periodically reports a Blockreport that includes all blocks stored on this DataNode to NameNode Only the DataNode has knowledge about the availability of one block replica. 19 Filesystem Metadata(cont’d) When NameNode starts up Load FSImage and EditLog from the local file system Update FSImage with latest EditLogs Create a new FSImage for latest checkpoint and store on local file system permanently 20 Communication Protocol A Hadoop specific RPC on top of TCP/IP NameNode is simply a server that only responses to the requests issued by DataNodes or clients ClientProtocol.java – client protocol DatanodeProtoco.java – datanode protocol 21 Robustness Primary object of HDFS: Reliable with component failures In a typical large cluster (>1K nodes), component failures are common Three common types of failures: NameNode failures DataNode failures Network failures 22 Robustness (cont’d) Heartbeats Each DataNode sends heartbeats to NameNode periodically System status and block reports The NameNode marks DataNodes w/o recent heartbeats as dead Does not forward I/O to it Mark all data blocks on these DataNodes as unavailable Re-replicate these blocks if necessary (according to the replication factor). Can detect network failures and DataNode dies 23 Robustness (cont’d) Re-Balancing Automatically move the data on one DataNode to another one If the free space falls below a threshold Data-Integrity A block of data may be corrupted Disk faults, network faults, buggy software Client computes checksums for each block and stores them in a separate hidden file in HDFS namespace Verify data before read it 24 Robustness (cont’d) Metadata failures FSImage and EditLog are the central data structures Once corrupted, HDFS can not build namespace and access data NameNode can be configured to support multiple- copies of FSImage and EditLog E.g: one FSImage/EditLog on local machine, another one is stored on mounted remote NFS server. Reduce the update performances Once NameNode is down, it must to restart the cluster manually 25 Data Organization Data Blocks HDFS is designed to support very large files and streaming I/Os A File is chopped up into 64MB blocks Reduce the number of connection establishments and accelerate TCP transmissions If possible, each block of a file will reside on a different DataNode For future parallel I/O and computations (MapReduce) 26 Data Organization (cont’d) Staging When write a new file A client firstly caches the file data into temporary local file until this file worth over the HDFS block size Then the client contacts NameNode to assign a DataNode The client flushes the cached data to the chosen DataNode Fully utilized the bandwidth 27 Data Organization (cont’d) Replication Pipeline A client obtains a DataNode list to flush one block The client firstly flushes the data to the first DataNode The first DataNode starts to receive the data in small portions (4kB), writes that portions to local storage, and transfer it to the next DataNode in the list immediately The second DataNode acts as the first one The total transfer time for one block(64MB) is: T(64MB) + T(4kb) * 2 , for pipeline 3 * T(64MB), for non-pipeline 28 Replication Pipeline The client asks the NameNode where to put data The client push data to DataNode linearly to fully utilize network bandwidth The secondary replicas reply to the primary. Then the primary replies to the client for success. * This figure was in “The Google File System” paper 29 See also HBase – a BigTable implementation on Hadoop Key-value storage Pig – high-level language to run data analyze on Hadoop ZooKeeper “ZooKeeper: Wait-free Coordination for Internet-scale Systems”, ATC’10, Best Paper CloudStore (KFS, previously Kosmosfs) A C++ implementation of Google File System Parallels the Hadoop project 30 Google v.s Y!/Facebook/Amazon.. Google Hadoop • Google File System • MapReduce • BigTable • Hadoop DFS • Hadoop MapReduce • HBase 31 Known Issues and Research Interests NameNode is the single point failure Limits the total files supported in the HDFS as well RAM limitation Google has changed the one-master architecture to multiple-header cluster However, the details are unrevealed 32 Known Issues and Research Interests (cont’d) Use replications to provide data reliability Same problems to RAID-1 ? Apply RAID technologies to HDFS? “DiskReduce: RAID for Data-Intensive Scalable Computing”, PDSW’09 33 Known Issues and Research Interests (cont’d) Energy Efficiency DataNodes are alive for data availability However, there may be no MapReduce computations running on them. Waste of energy 34 Conclusion Hadoop Distributed File System is designed to serve MapReduce computations Provide high reliable storage Support mass of data Optimized data placement policies based on the topology of data centers Large companies build their core businesses on top of these infrastructures Google: GFS/MapReduce/BigTable Yahoo!/Facebook/Amazon/Twitter/NY Times: Hadoop/HBase/Pig 35 Reference HDFS Architecture Guide: http://hadoop.apache.org/hdfs/docs/current/ hdfs_design.html 36 Questions? Thank you ! 37