Release notes v2.3#

This release introduces several new features and improvements of the engine calculators. Moreover, our packaging strategy has been revolutionized and now we provide packages for most operating systems - including previously abandoned systems such as Ubuntu 12.04 - by using the exact same versions of the libraries on every supported platform (Linux, Windows, Mac).

More than 20 pull requests were closed in oq-hazardlib and more than 70 pull requests were closed in oq-engine. For the complete list of changes, please see the changelogs: https://github.com/gem/oq-hazardlib/blob/engine-2.3/debian/changelog and https://github.com/gem/oq-engine/blob/engine-2.3/debian/changelog.

A summary follows.

New features in the engine#

From this release it is possible to take into account the topography of the region of interest. It is enough to write the elevation of the hazard sites in the sites file or in the site model file, as a negative depth, and the engine will compute the ground shaking correctly. Before there was no way to specify the elevation of the hazard sites and the ground shaking was computed only at the sea level.

From this release it is also possible to use magnitude-dependent integration distances. This is expected to have a substantial impact on the calculation time, since small magnitude ruptures will affects a lot less sites than before. In order to use the feature the user has to specify in the job.ini file a list of pairs (magnitude, distance) for each tectonic region type. In the past the user would write

maximum_distance = {'Active Shallow Crust': 200}

and the engine would use an integration distance of 200 km for all ruptures of the given tectonic region type, disregarding the magnitude. Now, the user can write

maximum_distance = {'Active Shallow Crust': [(8, 200), (7, 100), (5, 20)]}

and the engine will use an integration distance of 200 km for ruptures of magnitude 8, of 100 km for ruptures of magnitude 7, and of 20 km for ruptures of magnitude 5. Intermediate magnitudes will use a linearly interpolated distance. It is up to the user to provide a sound magnitude-distance function. For the future, we plan to release tools to help the users in the task of inferring magnitude-distance functions. If in doubt, you can always use the old syntax, equivalent to a constant maximum distance for all magnitudes, and the engine will keep working as in the past.

A lot of work went into the UCERF calculators, but they are still officially marked as experimental and are left undocumented on purpose. There is a new time-dependent classical calculator, while the old calculators were substantially improved.

There is an experimental command run_tiles to split a sites.csv file in tiles and run multiple calculations at once. This is meant for power users only.

There is a new command oq reset that will remove all calculations of the current user from the database and the filesystem. It is meant for cleaning up a testing machine were there are no important calculations. Use it with care.

New features in hazardlib#

New GMPEs - site adjusted Pezeshk et al. 2011, plus a few modified GMPEs for Armenia - were added.

The new GriddedSurface object introduced in engine 2.2 can now be serialized in the NRML 0.5 format properly.

The calculator calc_hazard_curves now supports mutually exclusive sources and ruptures too.

Performance improvements#

The calculation of ground motion fields is a lot faster than before in situations with a large amount of hazard sites. The reason is that now the distances are computed in the workers, whereas before they were computed in the controller node. This has a significant impact on users with a cluster.

Changes#

There are two new limits in the event based calculators:

  1. it is impossible to run calculations with more than 65536 tasks

  2. it is impossible to have tasks generating more than 65536 events each

These limits have been added to prevent users from running impossibly large calculations: an error message will be displayed early rather than causing machines to crash.

The event loss table exporter is now faster than in release 2.2. It does not export the tectonic region type string anymore, however you can infer it from the event_tag field which contains the source group ID. The association between the source group ID and the tectonic region type is contained in the new export sourcegroups.csv. The source_id has been stripped from the event tag since it was taking a lot of memory in the case of extra-large calculations (UCERF).

The calculation of quantiles hazard curves used different algorithms in the case of logic tree full enumeration and logic tree sampling. For the sake of consistency, we now use always the same algorithm, the one of full enumeration. This change may produce small differences in the quantile curves if you were doing computations with sampling of the logic tree.

The separator for the exported file ruptures.csv has changed from comma to tab, since this file now contains MULTIPOLYGON geometries containing commas.

We do not compute the mean anymore if there is a single realization, even if the user has set mean_hazard_curves=true.

We changed the DbServer port from port 1999 to port 1908. The reason is that some anti-viruses were blocking port 1999.

Fixes to the Web UI#

The “Remove” button now actually removes the calculation. Before it was only hiding it from the Web UI. Be careful with it, since it is final. If you remove a calculation, the only way to restore it is to repeat the computation.

We fixed the names of the files downloaded from the Web UI. The names of files downloaded from the Web UI previously contained slash (‘/’) characters that were then replaced with dashes or underscores by the browser. We no longer include slash characters in file names.

The named of the zipped outputs contained a spurious “.”. This has been fixed.

Other#

We removed a warning about missing the rtree library that was too noisy.

A bug in the logic tree storage was fixed. Now it possible to run a computation on a machine, copy the datastore on a different machine and to export the results from there.

The was a bug in scenario_risk calculations in the case of exposures featuring multiple assets of the same taxonomy on the same point without insured losses. This is now fixed.

There was a bug when combining hazard curves happening in rare situations, when a source group was generating probabilities of exceedence all zeros. This has been fixed.

Configuration files with empty conditional_loss_poes and non-empty loss_ratios were giving an ugly error message at export time. Now the non-existing loss maps are simply not exported.

The parallelization library is now more robust against unexpected exceptions.

Calculations in which a single task is generated are now executed in core, i.e. no subprocess is spawned.

As usual the internal representation of several data structures has changed in the datastore. In particular, we removed several usages of variable-length strings, since they were buggy in the case of large arrays. As always, users should not rely on the stability of the internal datastore formats.

UCERF calculations (and more generally any calculations requiring a shared directory) now work even on clusters without a shared directory: they will only use the controller node, but at least they will not fail.

The header of the exported losses for scenario calculations with occupants was the ugly string occupants:float32: now it is just occupants.

The command oq engine --make-html-report now works with Python 3 too.

The commands oq info and oq info --report have been fixed and improved; in particular, now the report contains the correct number of effective ruptures (i.e. ruptures within the integration distance) even for classical calculations.

The openquake.cfg file is included in the installation, so users that installed the engine via pip install openquake.engine will not get any complain about the missing configuration file.

The repository with the latex manual has been merged into the oq-engine repository, to help reducing disalignement between the code and the documentation.

We fixed an issue with rabbitmq on Ubuntu 16.04 by updating our configuration file openquake.cfg to not use the guest account.