《分布式计算分布式文件系统》由会员分享，可在线阅读，更多相关《分布式计算分布式文件系统（199页珍藏版）》请在金锄头文库上搜索。
1、Lecture 22.1 分布式系统简介2.1.1 分布式系统概述2.1.2 分布式系统中的关键性问题2.1.3 分布式系统举例 2.2 分布式文件系统2.2.1 NFS，AFS2.2.2 GFS 2.3 Hadoop 2.3.1 Hadoop概述2.3.2 HDFS 2.3.3 Hadoop的安装2.3.4 Hadoop程序的运行2.1.1 分布式系统概述 Introduction2.1.2 分布式系统中的关键性问题 Failure Lord of the Rings New Line Cinema Distributed ProblemsnIndexing the web (Google)
2、nSimulating an Internet-sized network for networking experiments (PlanetLab)nSpeeding up content delivery (Akamai)What is the key attribute that all these examples have in common?PlanetLabPlanetLab is a global research network that supports the development of new network services.PlanetLab currently
3、 consists of 809 nodes at 401 sites.CDN - AkamaiDistributed ProblemsnAll involve separable computationnMany involve data that necessarily must be stored in multiple locations.nFor a problem to be distributable, different components of the problem should be able to be handled independently.Taking A S
4、tep BacknBefore we talk more about distributed computing what does it mean to design “a computer?”nHow would a distributed or parallel system look different from a single-CPU machine?Flynns TaxonomynFour categories of computer architecturesnBroke down serial/parallel in terms of instructions and dat
5、aParallel vs. DistributednParallel computing can mean:nVector processing of data (SIMD)nMultiple CPUs in a single computer (MIMD)nDistributed computing is multiple CPUs across many computers (MIMD)What is Different in Distributed?nHigher inter-CPU communication latencynIndividual nodes need to act m
6、ore autonomouslynDifferent nodes can be heterogeneous (by function, location)nSystem reliability is much harder to maintain分布式系统具有不可预测性“A distributed system is one in which the failure of a computer you didnt even know existed can render your own computer unusable”- Leslie Lamport对分布式系统的考虑1、Reliabil
7、ity Demands 2、FailureReliability Demandsn（1）Support partial failurenTotal system must support graceful decline in application performance rather than a full haltReliability Demandsn（2）Data RecoverabilitynIf components fail, their workload must be picked up by still-functioning unitsReliability Deman
8、dsn（3）Individual RecoverabilitynNodes that fail and restart must be able to rejoin the group activity without a full group restartReliability Demandsn（4）ConsistencynConcurrent operations or partial internal failures should not cause externally visible nondeterminismReliability Demandsn（5）Scalability
9、nAdding increased load to a system should not cause outright failure, but a graceful declinenIncreasing resources should support a proportional increase in load capacityReliability Demandsn（6）SecuritynThe entire system should be impervious to unauthorized accessnRequires considering many more attack
10、 vectors than single-machine systemsKen Arnold, CORBA designer:“Failure is the defining difference between distributed and local programming”Failuren在分布式系统中，为保证系统的可靠性， 必须增强对“失效”的考虑。nFailure 都包括哪些？（1）Component FailurenIndividual nodes simply stop（2）Data FailurenPackets omitted by overtaxed routernOr
11、dropped by full receive-buffer in kernelnCorrupt data retrieved from disk or net（3）Network FailurenExternal did not use TCP streamsnFile locking, etc, implemented in higher-level protocolsnModern implementations use TCP/IP multi-GB files typicalnFiles are write-once, mostly appended tonPerhaps concu
12、rrentlynLarge streaming readsnHigh sustained throughput favored over low latencyGFS Design Decisions （1/2 ）nFiles stored as chunksnFixed size (64MB)nReliability through replicationnEach chunk replicated across 3+ chunkserversnSingle master to coordinate access, keep metadatanSimple centralized manag
13、ementGFS Design DecisionsnNo data cachingnLittle benefit due to large data sets, streaming readsnFamiliar interface, but customize the APInSimplify the problem; focus on Google appsnAdd snapshot (快照)and record append operationsnsnapshot (快照)快照操作几乎在瞬间构造一个文 件和目录树的副本，同时将正在进行的其他修改 操作对它的影响减至最小。 GFS Clien
14、t Block DiagramGFS ArchitecturenSingle masternMutiple chunkserversCan anyone see a potential weakness in this design?Single masternFrom distributed systems we know this is a:nSingle point of failurenScalability bottlenecknGFS solutions:nShadow mastersnMinimize master involvementnnever move data thro
15、ugh it, use only for metadatanand cache metadata at clientsnlarge chunk sizenmaster delegates authority to primary replicas in data mutations (chunk leases)nSimple, and good enough!Metadata (1/2)nGlobal metadata is stored on the master（ Master保存着三类元数据）nFile and chunk namespacesnMapping from files to
16、 chunksnLocations of each chunks replicasnAll in memory (64 bytes / chunk)nFastnEasily accessibleMetadata (2/2)nMaster has an operation log for persistent logging of critical metadata updatesnpersistent on local disknreplicatedncheckpoints for faster recovery（进行恢复仅需要最新 的检查点和相应的日志文件）operation lognmaster會定期的將元数据的信息記錄到硬 盘中及遠端的机器，以防master掛掉時可 以利用這些机器做短暫的服務，而這些被 备份的信息又稱作operation log。n當operation log超過一定量時，master便會 將整個系統狀態記錄成一份checkpoint。
