Skip to content
New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

IterationEncoding: variableBased #250

Open
wants to merge 9 commits into
base: upcoming-2.0.0
Choose a base branch
from
25 changes: 25 additions & 0 deletions FORMAT_ADIOS.md
Original file line number Diff line number Diff line change
Expand Up @@ -32,3 +32,28 @@ Output from `bpls -A` for a boolean attribute `pybool` stored in the location of

There is no convention yet for a unique representation of ADIOS2 variables with boolean type.
Thus, implementations should cast the data to and from `unsigned char` instead.

## `stepBased` Encoding of Iterations
Copy link
Member Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Rename: I am not sure why, but for some reason we now call this variableBased in openPMD/openPMD-api#855

@franzpoeschel let's clarify what we pick, shall I update the standard PR to be named variableBased, too?

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Yep, we discussed this a while ago and came to the conclusion to call it variable-based since steps are an ADIOS2-specific feature, but this encoding generally relies on a backend's ability to have variable datasets.


The `iterationEncoding` mode `stepBased` must be implemented via ADIOS steps.

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Suggested change
The `iterationEncoding` mode `stepBased` must be implemented via ADIOS steps.
The `iterationEncoding` mode `variableBased` must be implemented via a backend's feature to describe *variable* datasets and attributes.
This means that such datasets and attributes are present in different versions with different contents.

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

I think I missed here that this file was ADIOS-specific.


## Datasets

An openPMD **data set** is represented by an group prefix that contains an ADIOS variable `__data__`.

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

This describes the ADIOS2 schema that we will abolish

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Generally, we will stick with a schema that is similar to the current schema (datasets and attributes are distinguished by using different backend features (variables and attributes)), but with two additions:

  • Attributes can be variable now too
  • We will use some protocol for identifying if a group is active in the current step

Since that updated schema is not yet implemented, I'd suggest we don't describe this just yet. I would not like to standardize something that in the end turns out to not work well.

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

I've started exploring that new schema here.


**attributes** are defined further below and can also appear at the dataset's **group** prefix level.

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

What does "further below" mean in this context?


## Attributes

openPMD **attributes** stored as ADIOS `Variables` at the location where they would usually be stored.
Copy link
Member Author

@ax3l ax3l Jul 29, 2021

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

@franzpoeschel raised this is currently implemented that way and I think since we don't need to change the type of attributes over time, we can keep it so:

Suggested change
openPMD **attributes** stored as ADIOS `Variables` at the location where they would usually be stored.
openPMD **attributes** stored as ADIOS `Variables` at the location where they would usually be stored.
The `__is_boolean__/...` qualifiers are still stored as ADIOS `Attribute`.

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

openPMD attributes stored as ADIOS Variables

Is this outdated? We will remove the new ADIOS2 schema where this happens


Example for a mesh record `E` with record component `x` and attributes `unitDimension` and `unitSI`:
```
double /data/meshes/E/unitDimension 10*{7}
double /data/meshes/E/x/__data__ 10*{1000}
double /data/meshes/E/x/position 10*{1}
double /data/meshes/E/x/unitSI 10*scalar
```

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Outdated. Since the old ADIOS2 schema does not yet support variable-based iteration encoding, we currently only have this kind of experimental implementation.


This example uses `stepBased` iteration encoding, but other iteration encodings would work similarly with their respective `basePath` prefix.

37 changes: 27 additions & 10 deletions STANDARD.md
Original file line number Diff line number Diff line change
Expand Up @@ -48,7 +48,9 @@ for changes in keywords).
Hierarchy of the Data File
--------------------------

The used hierarchical data file format must provide the capability to
For simplicity, we call the storage concept of a specific data format that implements the openPMD hierarchy "files", even if they are implemented in-memory or by other means.

The used hierarchical data format must provide the capability to

- create groups and sub-groups (in-file directories)
- create multi-dimensional, homogeneous array-based data structures
Expand Down Expand Up @@ -85,7 +87,10 @@ Each file's *root* group (path `/`) must at least contain the attributes:
to create a real path from it replace all occurrences
of `%T` with the integer value of the iteration, e.g.,
`/data/%T` becomes `/data/100`
- allowed value: fixed to `/data/%T/` for this version of the standard
- allowed values:
- see *Iterations and Time Series* below
- for `fileBased` and `groupBased`, this is fixed to `/data/%T/`
- for `stepBased` this is fixed to `/data/`

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Suggested change
- for `stepBased` this is fixed to `/data/`
- for `variableBased` this is fixed to `/data/`

- note: all the data that is formatted according to the present
standard (i.e. both the meshes and the particles) is to be
stored within a path of the form given by `basePath` (e.g. in
Expand Down Expand Up @@ -195,9 +200,8 @@ standard:
Iterations and Time Series
--------------------------

Iterations can be encoded in either the file name of each master-file of a
time step or in groups of the same file. (Here, an *iteration* refers
to a single simulation cycle.)
Iterations can be encoded in either the file name of each individual files, in groups of the same file, or in data sets & attributes (with supported data formats).
(Here, an *iteration* might refer to a single measurement or simulation cycle.)

The chosen style shall not vary within a related set of iterations.

Expand All @@ -212,6 +216,7 @@ Each file's *root* group (path `/`) must further define the attributes:
- allowed values:
- `fileBased` (multiple files)
- `groupBased` (one file)
- `stepBased` (one file with internal encoding for iterations, if supported by the data format)

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Suggested change
- `stepBased` (one file with internal encoding for iterations, if supported by the data format)
- `variableBased` (one file with internal encoding for iterations, if supported by the data format)


- `iterationFormat`
- type: *(string)*
Expand All @@ -222,20 +227,32 @@ Each file's *root* group (path `/`) must further define the attributes:
for `fileBased` formats the iteration must be included
in the file name;
the format depends on the selected `iterationEncoding` method
- note: it is not required that every openPMD iteration contains an update for each declared openPMD record (see below)
- examples:
- for `fileBased`:
- `filename_%T.h5` (without file system directories)
- for `groupBased`:
- for `groupBased`: (fixed value)
- `/data/%T/` (must be equal to and encoded in the `basePath`)
- for `stepBased`: (fixed value)

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Suggested change
- for `stepBased`: (fixed value)
- for `variableBased`: (fixed value)

- data-format internal convention
- *slowest varying index* of data

### `stepBased` Encoding of Iterations

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Suggested change
### `stepBased` Encoding of Iterations
### `variableBased` Encoding of Iterations


In order to correlate openPMD iterations with an index of data-format internal updates/steps or an index in the slowest varying dimension of an array, the *root* group (path `/`) must contain an additional variable once `stepBased` is chosen for `iterationEncoding`:

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Suggested change
In order to correlate openPMD iterations with an index of data-format internal updates/steps or an index in the slowest varying dimension of an array, the *root* group (path `/`) must contain an additional variable once `stepBased` is chosen for `iterationEncoding`:
In order to correlate openPMD iterations with an index of data-format internal updates/steps or an index in the slowest varying dimension of an array, the *root* group (path `/`) must contain an additional variable once `variableBased` is chosen for `iterationEncoding`:


- `snapshot`
Copy link
Member Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Question: we could allow to skip this if only one iteration (snapshot) is written.
In that case, the implied value should be 0 and there must be exactly one update/step in the data format.

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

This is what the other backends actually do when that iteration encoding is chosen, see the variableBasedSingleIteration test. The snapshot attribute is not written.

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

With the current state of openPMD/openPMD-api#949, the snapshot attribute is always written, but not required at read-time (then assumed to be 0). I should add a test somehow to ensure that reading without snapshot works as intended.

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

now tested

- type: 1-dimensional array containing N *(int)* elements, where N is the number of updates/steps in the data format

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

This is currently not implemented as an array, but as a scalar variable that changes across steps.

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

The implementation actually accepts arrays at read-time, but I should test that it works. At write time, the API currently only produces scalars.

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

now tested

- description: for each update/step in a data format, this variable needs to be updated with the corresponding openPMD iteration.
- note: in some data formats, updates/steps are absolute and not every update/step contains an update for each declared openPMD record
- advice to implementers: an openPMD iteration might be spread over multiple updates/steps, but not vice versa.
In such a scenario, an individual openPMD record's update/step must appear exactly once per iteration.

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Note: A similar situation can occur when using Append mode: An iteration is then present multiple times with redundant definitions. This will either be solved by truncation or by reading only the first/last instance of that iteration.



Required Attributes for the `basePath`
--------------------------------------

In addition to holding information about the iteration, each series of
files (`fileBased`) or series of groups (`groupBased`) should have
attributes that describe the current time and the last
time step.
In addition to holding information about the iteration, each series of files (`fileBased`), series of groups (`groupBased`) or internally encoded iterations (`stepBased`) should have attributes that describe the current time and the last time step.

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Suggested change
In addition to holding information about the iteration, each series of files (`fileBased`), series of groups (`groupBased`) or internally encoded iterations (`stepBased`) should have attributes that describe the current time and the last time step.
In addition to holding information about the iteration, each series of files (`fileBased`), series of groups (`groupBased`) or internally encoded iterations (`variableBased`) should have attributes that describe the current time and the last time step.


- `time`
- type: *(floatX)*
Expand Down