Spark basic concepts:
1, RDD (resillient distributed dataset) 2, Task: shuffleMapTask and resultTask (simillar to map and reduce) 3, Job: a job can be made of multiple tasks 4, Stage: a job can have multiple stages 5, Partition: RDD can be partitioned into different machine 6, NarrowDependency: Base class for dependencies where each partition of the child RDD depends on a small number of partitions of the parent RDD. Narrow dependencies allow for pipelined execution. 7, ShuffleDependency: also called wideDependency, child RDD depend on all partitions of the parent RDD. 8: DAG: Directed Acycle graph, no parent depend on child RDD
Spark Core functions:
1, SparkContext: for driverApplication execution and output, we need to initiallize SparkContext before submit spark jobs SparkContext has: 1) communiation 2) distributed deployment 3) message 4) storage 5) computation 6) cashe 7) measurement system 8) file service 9) web service Application need use SparkContext API to create jobs, use DAGScheduler, plan RDDs in DAG to different stages and submit the stages. use TaskScheduler, apply resouces, submit jobs and requst cluster for scheduling 2, Storage System 1) Spark take memory as priority, if memory is not enough, then consider to use disk, Tachyon (distributed memory file system) 3, Computation Engine: 4, Deployment 1) Standalone 2) Yarn 3) Mesos
1, Data Serialization: 1) Java serializaion (object --> byte --> object) 2) Kyro serializaton (10x faster than Java serialization) (object --> object) val conf = new SparkConf().setMaster(...).setAppName(...) conf.registerKryoClasses(Array(classOf[MyClass1], classOf[MyClass2])) val sc = new SparkContext(conf) 2, Memory Tuning: 1) object header: 16 bytes 2) String header: 40 bytes 3) Common collection class: HashMap or LinkedList, 8 bytes 4) Collections of primitive types often store them as "boxed" object as java.lang.Integer 3, Memory management overview 1)Memory usage in Spark largely falls under one of two categories: execution and storage. a) Execution memory refers to that used for computation in shuffles, joins, sorts and aggregations b) Storage memory refers to that used for caching and propagating internal data across the cluster 2) M/R a) When no execution memory is used, storage can acquire all the available memory and vice versa. b) R describes a subregion within M where cached blocks are never evicted 3) This design ensures several desirable properties: a) First, applications that do not use caching can use the entire space for execution, obviating unnecessary disk spills. b) Second, applications that do use caching can reserve a minimum storage space (R) where their data blocks are immune to being evicted. c) Lastly, this approach provides reasonable out-of-the-box performance for a variety of workloads without requiring user expertise of how memory is divided internally.