Release notes v2.9

Several improvements entered in this release, most notably about the event based calculator. Over 150 issues were closed. For the complete list of changes, please see the changelog: https://github.com/gem/oq-engine/blob/engine-2.9/debian/changelog .

Hazard

The problem of slow tasks has been an issue with the engine from the very beginning. This release does not solve the problem, but features the biggest improvement ever made. The task distribution has been improved so much that we can afford to generate 40% fewer tasks than in the previous release and yet have fewer slow tasks. This is very important in large scale computations, where data transfer is an issue and producing fewer tasks improves the performance.

Moreover, we solved the issue of the engine not producing adequate tasks in some situations, for instance classical PSHA calculations with few sites, or event based hazard calculations with few sources. The “inadequate number of tasks” issue meant wasting of computation power, i.e. not using all of the available cores.

The improvement of the task distribution was made possible by refining the source weighting algorithm and by splitting the area sources before the prefiltering, thus getting better upfront estimates of the computational size of a task.

The prefiltering mechanism itself has been improved and we fixed a bug affecting sites around the international date line: the prefiltering with rtree was incorrectly discarding relevant sources. The bug was introduced several releases ago and discovered by our users from New Zealand.

We substantially reduced (in a SHARE calculation by a factor of 170!) the data transfer when calculating hazard statistics in postprocessing. Thanks to that, it is now possible to compute mean/max/quantile hazard curves in calculations as large as the 5th Generation Seismic Hazard Model of Canada (SHMC) (with over 200,000 sites and 13,122 realizations).

We fixed a few bugs in the disaggregation calculator breaking some corner cases and we paved the way for implementing disaggregation by source.

We improved the support for nonparametric sources with gridded surfaces, a feature introduced several releases ago. Now nonparametric sources can be split properly, thus producing a good task distribution and they can be used in a disaggregation calculation. The support is still not complete, for instance you cannot export ruptures with gridded surfaces, but we are working on it.

We introduced a minimum distance concept for Ground Motion Prediction Equations. The GMPEs have a range of validity, and you cannot trust them on sites too close to the seismic source. By specifying the minimum distance it is possible to cut off the value of the generated ground motion field: for sites below the minimum distance the GMF is cut off at the maximum value. It is possible to specify a minimum distance (in km) for each GSIM by updating the gsim logic tree file. Here is an example of how it would look like:

              <logicTreeBranch branchID="b1">
                <uncertaintyModel minimum_distance="10">
                  LinLee2008SSlab
                </uncertaintyModel>
                <uncertaintyWeight>0.60</uncertaintyWeight>
              </logicTreeBranch>
                
              <logicTreeBranch branchID="b2">
                <uncertaintyModel minimum_distance="5">
                  YoungsEtAl1997SSlab
                </uncertaintyModel>
                <uncertaintyWeight>0.40</uncertaintyWeight>
              </logicTreeBranch>

Hazardlib/HMTK

Graeme Weatherill contributed the GMPEs of Bommer et al. (2009) and Afshari & Stewart (2016) as well as a fix to the ASK14 GMPE, which was failing in the case of small magnitudes (< 4.4) and long periods (> 5). The fix is the same used by the original authors of the GMPE.

Graeme also contributed a comprehensive suite of NGA East ground motion models for central and eastern United States.

There is a small fix for the Derras 2014 GMPR, for distances below 0.1 km.

We added the Yu et al. (2013) GMPEs for China together with Changlong Li.

The visualization capabilities of the Hazard Modeller Toolkit (HMTK) have been enhanced thanks to a contribution by Nick Ackerley.

We changed the ordering of the IMTs in hazardlib: before it was purely lexicographic, now it is lexicographic for non-SA IMTs, while SA IMTs are ordered according to their period, as requested by Mike Hearne of USGS.

The hazardlib classes FilteredSiteCollection and SiteCollection have been unified. This should not affect users, except people calling the internals of the SiteCollection class.

Risk

The biggest improvement on the risk side is in the event based calculator, which is now able to use precomputed ground motion fields, a feature that before was available only in the experimental gmf_ebrisk calculator. The net effect is a huge improvement in the performance of the calculator, so much that now the suggested strategy to run large event based risk calculation is to precompute the ground motion fields, provided you have enough disk space. There were also a few fixes, in particular in the management of the epsilons, now controlled by the master_seed parameter in all situations, and in the generation of the aggregate loss curves, which is now a lot faster.

It is now possible to read the ruptures directly from the workers without sending them through the wire. This is relevant for large calculations producing millions of events, since it reduces substantially the data transfer. Also, the way the ruptures are stored in the datastore has been optimized. The consequence is that you cannot read ruptures stored with engine 2.8 with engine 2.9.

On the other hand, hazard curves, hazard maps, and uniform hazard spectra produced with older versions of the engine are still readable by this version of the engine and are viewable by the latest QGIS plugin. We will make any effort to maintain this feature even in future releases.

The other big improvement on the risk side was the support for CSV exposures. The support is not complete yet, and we are not supporting all the features of XML exposures, however we already support the most common cases. The performance is of course better than for XML and we were able to import 3 million assets for an exposure in California in just 4 minutes. We had to change the internal storage of the exposure to work around a bug in h5py/HDF5 making it impossible to store such large exposures. We are now able to run full risk calculations for exposures of this size in a couple of hours on a workstation, a feat previously impossible.

We fixed a bug in vulnerability functions with PMFs: the random seed for the sampling method of the discrete probability distribution was set uncorrectly and in some cases the same value of the loss ratio could be sampled multiple times.

We improved the display names for the risk outputs, i.e. the names of the outputs as seen from the engine, the WebUI, and the QGIS plugin.

WebAPI/WebUI/QGIS plugin

We enhanced significantly the /extract/ web API, allowing more data to be extracted from the datastore in a proper way, “proper” meaning future proofed against changes in the internal structure of the datastore.

The /extract/ API is based on an underlying extract Python API, that can be used directly in scripts, with the same guarantees of portability across engine versions.

There is now an abort WebAPI and an abort button in the WebUI and QGIS plugin, using which it is possible to cleanly kill a running calculation.

We fixed a small issue: now when the WebUI is stopped all calculations started by it are automatically killed.

The integration between the plugin and the web tools (in particular the Input Preparation Toolkit) has been improved as well; for the new features of the plugin please read its release notes.

Bug fixes/additional checks

We fixed a minor bug with the .npz exports: the command oq engine --run job.ini --exports npz now works as expected.

We raise a clear error when the parameter risk_investigation_time is missing in the job_risk.ini file.

There is a better error message when the user wants to compute loss maps and the parameter asset_loss_table is missing.

There is a clear warning when the engine produces no output, usually because mean_hazard_curves is set to false. Also, in that case, no output rows are incorrectly added to the database.

There is a check on the path names used in the source model logic tree file: they must be relative paths, not absolute paths, to make it possible to port the input files across different machines and operating systems.

There is a check on calculations starting from precomputed hazard results: if job.ini file contains a site model or logic tree files, the user is warned that they will not used for the subsequent calculation and should be removed.

We improved the error message in the case of source models in format nrml/0.4 uncorrectly marked as nrml/0.5.

The CSV rupture exporter now also exports the parameter ses_per_logic_tree_path in the header of the CSV file.

A lot of work went into the .rst reports which are now more reliable than ever.

oq commands

There is a new command oq shell that starts an IPython shell (or Python shell, if IPython is not available) in the engine virtual environment. This is very useful to explore interactively the results of a computation.

There is a new command oq abort <job_id> that allows to kill properly running calculations.

The command oq engine --delete-calculation <job_id> has been extended to abort running calculations before removing them from the database and the file system.

The command oq zip has been extended and now works for all kind of calculations.

The command oq extract has now a signature consistent with oq export and has grown the ability to connect to a remote engine server.

The oq dump command has been extended so that it is possible to extract a single calculation, if so desired.

IT

We have specialized packages for clusters even for Ubuntu. The packages meant for the controller node includes rabbitmq and celery, which are missing in the packages meant for the worker nodes. Previously, we were installing rabbitmq and celery even when not needed.

Docker containers can be used to deploy a multi-node installation of the Engine on a Docker cluster with dynamic scaling.

It is possible to specify a custom_tmp variable in the openquake.cfg file to configure a custom path for the temporary directory. This is useful in situations when the system /tmp directory is in a partition which is too small (this happened to users with a machine in the Azure cloud).

The version of numpy officially supported by the engine is 1.11 however we made some work to ensure that the tests run also with the most recent versions. In the near future we will probably migrate to numpy 1.14.

Since we added a field with the process ID in the job table in order to implement the abort functionality, upgrading to engine 2.9 requires a database migration if you have an older version of the engine and you want to keep your old calculations. You will just need to say yes the first time you run a calculation, when the engine will ask you to upgrade the database. The migration is immediate and safe.

Deprecations/removals

Since the GMF-reading feature has been incorporated into the core event_based_risk calculator, the experimental gmf_ebrisk calculator is gone.

Reading the GMFs from a NRML file has been deprecated. The suggested approach is to use the CSV format, as documented in the manual.

The obsolete hazard GeoJSON exporters, deprecated months ago, have been removed. The right way to convert hazard outputs into geospatial data is to use the QGIS plugin.

We deprecated the function openquake.hazardlib.nrml.parse, now superseded by the function openquake.hazardlib.nrml.to_python. This is of interest only to people using hazardlib from Python.

The ability to run two .ini files together (with the command oq engine --run job_h.ini,job_r.ini) has been removed, since it could produce issues in some cases and was never documented anyway.

Python 3 migration

All the installers provided by GEM (for Linux, Windows, macOS), all the virtual machines, all the Linux packages (CentOS, Ubuntu) and the docker containers are now based on Python 3.5. All use of Python 2.7 has been discontinued internally at GEM. You can still install the engine manually with Python 2.7 and it will work, but this is the last version to guarantee compatibility with Python 2. Starting with the next version (3.0) the engine will require at least Python 3.5 to work. The change has no effect on people using the engine as a tool since the right version of Python is already included in our installers/packages, but power users developing their own tools/libraries based on the engine will have to migrate their scripts to Python 3. If your scripts do not import engine libraries directly but they only perform calls to the WebAPI they will keep working of course. This is why the QGIS plugin is still working with engine, even if the plugin itself is using Python 2.7.