关键词: 内存计算, 大数据运算, 分布式计算, 有向无环图调度, 容错恢复

Abstract:

Concerning the limitations of Spark, including immutable datasets and significant costs of code execution, memory management and data serialization/deserialization caused by running environment of Java Virtual Machine (JVM), a light-weight big data processing system, named Helius, was implemented in C/C++. Helius supports the basic operations of Spark, while allowing the data set to be modified as a whole. In Helius, the C/C++ is utilized to optimize the memory management and network communication, and a stateless worker mechanism is utilized to simplify the fault tolerance and recovery process of the distributed computing platform. The experimental results showed that in 5 iterations, the running time in Helius was only 25.12% to 53.14% of that in Spark when running PageRank iterative jobs, and the running time in Helius was only 57.37% of that in Spark when processing TPCH Q6. On the basis of one iteration of PageRank, the IP incoming and outcoming data sizes of master node in Helius were about 40% and 15% of those in Sparks, and the total memory consumed in the worker node in Helius was only 25% of that in Spark.Compared with Spark, Helius has the advantages of saving memory, eliminating the need for serialization and deserialization, reducing network interaction and simplifying fault tolerance.

Key words: in-memory computation, big data processing, distributed computation, Directed Acyclic Graph (DAG) scheduling, fault tolerance and recovery