Sizing a seismic processing cluster: from 128 to 4,096 cores

01What drives the size of a cluster

Three things decide how much compute a processing project needs: the volume of data, the imaging or inversion method, and the turnaround the client needs. A 2D programme and a high-density 3D survey can differ in data volume by orders of magnitude, and depth imaging and full-waveform inversion repeat their most expensive steps many times.

We size the cluster from those three inputs first, then choose the interconnect and storage that keep the cores busy. A cluster with too little storage bandwidth or a slow network spends its time waiting on data, so extra cores add cost without adding speed.

02Core tiers and their typical use

Capacity scales in doubling steps. The tiers below are a starting guide; the final choice follows from the survey and the schedule.

• 128 cores: 2D lines, parameter testing and quality control.
• 256 cores: regional 2D programmes and smaller 3D volumes.
• 512 cores: pre-stack time migration (PSTM) on full 3D surveys.
• 1,024 cores: large 3D volumes and pre-stack depth migration (PSDM) with velocity model building.
• 2,048 cores: high-density 3D and time-lapse 4D with iterative depth imaging.
• 4,096 cores and beyond: full-waveform inversion (FWI) and large merged surveys.

03Interconnect and storage

Migration and inversion exchange large amounts of data between nodes. A low-latency InfiniBand interconnect keeps that exchange from becoming the bottleneck as the core count grows.

Pre-stack data, intermediate gathers and velocity models need fast shared storage. We specify SAN or NAS capacity and throughput together with the compute, so that reading and writing gathers keeps pace with the processing flow.

04Delivering a turnkey centre

A turnkey processing centre is delivered in five stages, each signed off before the next begins:

• Engineering plan: workload sizing, power, cooling and rack layout.
• Compute supply: servers sized to the chosen core tier.
• Network: InfiniBand interconnect between compute nodes, storage and head nodes.
• Storage: SAN/NAS sized for the survey data and intermediate products.
• Cluster and software: operating system, scheduler, cluster management and the GeoVision 1.3 processing environment.

05Planning to grow

Because tiers double, a centre can start at the size today's projects need and expand later. We plan rack space, power, cooling, switch ports and storage expansion at the start, so moving up a tier is an addition, not a rebuild.