High performance Kirchhoff pre-stack depth migration on Hadoop

Kirchhoff Pre-Stack Depth Migration (KPSDM), a widely used seismic imaging algorithm in petroleum industry, is a typical IO-bound application since a large amount of seismic data and travel timetable data needs to be read from the storage system iteratively during runtime. We present an optimized high performance KPSDM implementation called HKPSDM based on Hadoop where a large I/O aggregated bandwidth is offered cheaply to replace our previous implementation based on Network Attached Storage (NAS) appliances over NFS which have a limited I/O bandwidth when hundreds of processes participate in a migration job. In our implementation, MapReduce facilitates the travel timetable rearrangement and HDFS provides a stable and scalable storage system for seismic data as well as travel timetable data. Various optimizations are applied to HKPSDM to reduce the global bandwidth consumption of HDFS and overhead caused by some disadvantages of HDFS. Experimental results show that HKPSDM can scale better when the number of computing cores ranges from 160 to 800. HKPSDM performs more than 5 times better than NAS-based one when rearranging the travel timetables. And a 37% efficiency improvement is observed when 800 cores are used for migration.