Earth System Modelling as a Service

Nic Hannah

Double Precision

Very important: the models are of huge and growing importance to society - scientists, policy makers and the public require demonstrably high quality, credible models.

Very high complexity: millions of lines of code developed over decades. Code complexity on par with Linux kernel, Hubble space telescope, a military drone.

Low engineering quality: the models are developed in an ad-hoc manner without engineering processes/tools/metrics. e.g. climate models typically have 0 unit tests.

Very high risk! Just combine the points above and most engineering managers would start to worry.

Scientists are rewarded for doing science, publishing papers etc. not building systems.

Scientific work practice often focuses on short term and local - this make sense since long-term investment may not pay off.

For a variety of reasons software is viewed differently to other engineered scientific infrastructure such as telescopes, satellites, supercomputers.

Limited budgets are typically spent on more tangible or measurable things than engineering quality. e.g. NCI currently spending $$$ on model performance but very little on other (more important) aspects of software quality.

Scientists and the public are becoming more discerning - quality (and correctness) matter

Growing pressure from funding bodies and publishers particularly around scientific integrity and process, e.g. accessibility and reproducibility matter

Scientists in other disciplines are making amazing breakthroughs by embracing engineering, e.g. astronomy, plasma physics

Idea:

Employ more engineers (and fewer scientists) - hope that scientific output eventually increases. This is a long-term and risky strategy.

Idea:

Modelling as a service (MaaS)

Treat geophysical models like any other (engineered) infrastructure - such as a supercomputer or Internet connection.

Users pay for what they consume, e.g. per CPU hour or per simulation day
Very low risk for the user since no large investment in HR or machinery required
The service provider (Double Precision) is highly motivated to:
- increase the number of, and ease with which simulations are run
- invest in the quality and performance of the model and related infrastructure for the long term (it's our livelihood)
- work with scientists to extend the scientific reach of the model

Key points about MaaS:

as it stands model quality is of little interest or concern, under MaaS DP invests in a long-term, holistic approach to software quality. This includes such things as: testing, reliability, documentation.
restructuring hires to solve this problem within ARCCSS, BOM is expensive and risky, MaaS reduces risk
MaaS is customer focused and outward facing, this is an improvement on existing approaches model development in Australia.
the model becomes a tangible, visible asset much like a piece of hardware

We have interest from commercial (non-research) organisations who presently buy machines and hire staff to run WRF, using a service represents a saving for them.

We are building a relationship with Amazon scientific computing and have done preliminary testing of MOM and WRF using their HPC cloud