2017 Oil & Gas HPC Conference

Seismic data-sets generated during Oil & Gas exploration are extremely large and are broken
into data files which can be 100s of GiBs to 10s of TiBs and larger. From a software
development perspective, parallel I/O with seismic files is complex and challenging and thus
involves a significant percentage of developer productivity. The fundamental factor for
complexity is the nature of the I/O-access patterns which tend to be varied and involve large
amounts of data, resulting in the sub-optimal exploitation of current and emerging large-scale
HPC platforms.

The primary file formats in use within petroleum seismology have also accumulated 40 years
of legacy. For example, the de-facto standard, SEG-Y, requires fixed byte positions for the
storage of pre-defined trace parameters, uses the lower-precision IBM floating point format
for trace data and also uses EBCDIC for text. SEG-Y also interleaves both data and metadata
in the same file due to its origins in tape storage. The format does not support many common
use cases that geophysicists require since the nature of seismic workflows makes this
difficult in a static file format. Processing the peculiarities of the SEG-Y format can further
increase the complexity of the code base as a result.

In this presentation, we provide an overview of a new high-performance parallel I/O library,
ExSeisPIOL, dedicated to seismic processing workflows for petroleum seismology. Developed
to target SEG-Y, ExSeisPIOL is optimised for the structure of seismic files to deliver significant
improvements in both productivity and performance on large-scale HPC systems. Using
modern OO design, the library is developed on top of an interface abstraction which targets
MPI-I/O and other I/O interfaces. The library also targets the latest in burst buffer technology
with appropriate access semantics or through the direct exploitation of the low-level
interfaces, to deliver a significant improvement to I/O performance over standalone parallel
filesystems such as Lustre and GPFS.

We will describe the ExSeisPIOL in more detail by providing an overview of the multi-layered
structure as well as the two primary APIs which provide a high-level and low-level abstraction.
We will discuss real-world Oil & Gas use cases on large-scale HPC systems using burst-buffer
technology as well as a port to an existing seismic migration code (>15% code reduction).

The pressure to reduce both operating and capital costs in seismic data analysis drives an on-going demand for efficiency improvements in computational processing facilities. This is against a background of dramatically increasing seismic data volumes. Wide/Multi/Rich-azimuth methods using multi-sensor arrays and sophisticated acquisition techniques are producing higher-fidelity subsurface images, and modern analytics techniques are enabling continued advancement in the interpretation of seismic data for both newly acquired data and historical oil field data. As a result of the volume and scale of the seismic data required for modern HPC-based seismic processing and imaging, the performance of the associated storage subsystem can be a source of the greatest overall efficiency improvements.

Seismic analysis is particularly challenging for today’s file systems due to a tendency towards large random IO and share-file IO. Therefore, improving IO efficiencies for complex seismic workloads is key. The benefit of a true parallel file system is the very high single-client performance that can be delivered and sustained even when many hundreds of clients are working concurrently

This session will share the results experimental benchmarks to attain optimal IO rates for Paradigm’s Echos application workloads.

An effort began three years ago to profile various types of equipment in CGG’s Houston data center to better understand the composite workload. This paper gives an overview of the effort as well as specific examples of how it has led to improved equipment selection. The paper begins with a sketch of the Houston data center and the main types of equipment. Two types of storage system profiles are covered followed by two types of GPU system profiles. The study concludes with two types of CPU workload profiles. In each of the subsystem categories the profiling effort has resulted in improved equipment selection (HW), improved utilization, or deployment of new technologies. In most of these cases profiling has resulted in an increase of Performance/$ by 25% or more. A common theme of these improvements is to shift the data center design point from “+3 standard deviations” to “+1 standard deviation” with respect to workload requirements.