Skip to content


Switch branches/tags

Latest commit


Git stats


Failed to load latest commit information.
Latest commit message
Commit time

Earth System Data Middleware

The middleware for earth system data is a prototype to improve I/O performance for earth system simulation as used in climate and weather applications. ESDM exploits structural information exposed by workflows, applications as well as data description formats such as HDF5 and NetCDF to more efficiently organize metadata and data across a variety of storage backends.

Data model

The data model of a system organizes elements of data, standardizes how they represent data entities and how users can interact with the data. The model can be split into three layers:

  • The conceptual data model describes the entities and the semantics of the domain that are represented by the data model and the typical operations to manipulate the data. In our case, the scientific domain is NWP/climate.

  • The logical data model describes the abstraction level provided by the system, how domain entities are mapped to objects provided by the system[1], and the supported operations to access and manipulate these objects are defined. Importantly, the logical data model defines the semantics when using the operations to access and manipulate the system objects. For example, a system object of a relational model is a table – representing attributes of a set of objects – and a row of a table representing attributes of a single object.

  • The physical data model describes how the logical entities are finally mapped to objects/files/regions on available hardware. The physical data model is partly covered by the backends of ESDM, therefore, the descriptions will stop at that abstraction level.

Conceptual data model

Our conceptual data model is aligned with the needs of domain scientists from climate and weather. It reuses and extends from concepts introduced in a data model proposed for the Climate and Forecasting conventions for NetCDF data.


A variable, V, defines a set of data representing a discrete (generally scalar) quantity discretised within a “sampling” domain, d. It is accompanied by


Which will include at the minimum, a name, but may also include units, and information about additional dimensionality, directly (e.g. via a key, value pair such as that necessary to expose z = 1.5m for air temperature at 1.5m) or indirectly (e.g. via pointers to other generic coordinate variables which describe the sampled domain). There may also be a dictionary of additional metadata which may or may not conform to an external semantic convention or standard. Such metadata could include information about the tool used to observe or simulate the specific variable. Additional metadata is also required for all the other entities described below.


The sampling domain d is defined by Dimensions which define an a coordinate axis. Dimensions will also include metadata, which must again, include at a minimum a name (e.g. height, time), but may also include information about directionality, units (e.g. degrees, months, days-since-a-particular-time-using-a-particular-calendar), or details of how to construct an algorithm to find the actual sampling coordinates, perhaps using a well known algorithm such as an ISO 8601 time.


Coordinates are the set of values at which data is sampled along any given dimension. They may be explicitly defined by indexing into a coordinate variable, or implicitly defined by an algorithm. When we talk about the coordinates, it is usually clear if we mean the N-dimensional coordinate to address data in a given variable or if we just mean the (1D) coordinate along one dimension.


The data values are known at points, which may or may not represent a cell. Such cells are n-dimensional shapes where the dimensionality may or may not fully encompass the dimensionality of the domain. n-dimensional shapes can be implicitly defined in which case the Cartesian product of all dimensional coordinates forms the data "cube" of the cell, but they can also be explicitly defined, either by providing bounds on the coordinate variables (via metadata) or by introducing a new variable which explicitly defines the functional boundaries of the cell (as might happen in a finite element unstructured grid).


Variables can be aggregated into datasets. A dataset contains multiple variables that logically belong together, and should be associated with metadata describing the reason for the aggregation. Variables must have unique names within a dataset.

Our conceptual model assumes that all variables are scalars, but clearly to make use of these scalars requires more complex interpretation. In particular, we need to know the datype and operators.


defines the types of values that are valid and the operations that can be conducted. While we are mostly dealing with scalars, they may not be amenable to interpretation as simple numbers. For example, a variable may be storing an integer which points into a taxonomy of categories of land-surface-types. More complex structures could include complex data types such as vectors, compressed ensemble values, or structures within this system, provided such interpretation is handled outside of the ESDM, and documented in metadata. This allows us to limit ourselves to simple data types plus arbitrary length blocks of bits.


define the manipulations possible on the conceptual entities. The simplest operators will include creation, read, update and delete applied to an entity as a whole, or to a subset, however even these simple operators will come with constraints, for example, it should not be possible to delete a coordinate variable without deleting the parent variable as well. There will need to be a separation of concerns between operators which can be handled within the ESDM subsystem, and those which require external logic. Operators which might require external logic could include subsetting — it will be seen that the ESDM will support simple subsetting using simple coordinates — but complex subsets such as finding a region in real space from dimensions spanned using an algorithm or coordinate variable, may require knowledge of how such algorithms or variables are specified. Such knowledge is embedded in conventions such as the CF NetCDF conventions, and this knowledge could only be provided to the ESDM via appropriate operator plugins.

Logical data model

The logical data model describes how data is represented inside ESDM, the operations to interact with the data and their semantics. There are four first class entities in the ESDM logical data model: variables, fragments, containers, and metadata. ESDM entities may be linked by ESDM references, and a key property which emerges from the use of references is that no ESDM entity instance may be deleted while references to it still exist. The use of reference counting will ensure this property as well as avoid dangling pointers.

Each of these entities is now described, along with a list of supported operations:


In the logical data model, the variable corresponds directly to a variable in the conceptual data model. Each element of the variable sampled across the dimensions contains data with a prescribed DataType. Variables are associated with both Scientific Metadata and Technical Metadata. Variables are partitioned into fragments each of which can be stored on one or more “storage backend. ” A variable definition includes internal information about the domain (bounding box in some coordinate system) and dimensionality (size and shape), while the detailed information about which coordinate variables are needed to span the dimensions and how they are defined is held in the technical metadata. Similarly, where a variable is itself a coordinate variable, a link to the parent variable for which it is used is held in the technical metadata. The ESDM will not allow an attempt to delete a variable to succeed while any such references exist (see references). A key part of the variable definition is the list of fragments associated with it, and if possible, how they are organised to span the domain. Users may choose to submit code pieces that are then run within the I/O-path (not part within ESiWACE implementation), such an operation covers the typical filter, map and reduce operations of the data flow programming paradigm.

Fragments are created by the backend while appending/modifying data to a variable.


  • Variables can be created and deleted.

  • Fragments of data can be attached and deleted.

  • Fragments can be repartitioned and reshuffled.

  • Integrity can be checked.

  • Data can be read, appended or modified those operations will be translated to the responsible fragments.

  • Metadata can be atomically attached to a variable or modified.

  • A variable can be sealed to make it immutable for all subsequent modifications.

  • Process data of the variable somewhere in the I/O-path.


A fragment is a piece (subdomain) of a variable. The ESDM expects to handle fragments as atomic entities, that is, only one process can write a fragment through the ESDM, and the ESDM will write fragments as atomic entities to storage backends. The backends are free to further partition these fragments in an appropriate way, for example, by sharding using chunks as described in the mapping section. However, the ESDM is free to replicate fragments or subsets of fragments and to choose which backend is appropriate for any given fragment. This allows, for example, the ESDM to split a variable into fragments some of which are on stored on a parallel file system, while others are placed in object storage.


  • Data of fragments can be read, appended or modified.

  • Integrity of the fragment can be checked.

  • Process data of the variable somewhere in the I/O-path.


A container is a virtual entity providing views on collections of variables, allowing multiple different datasets (as defined in the conceptual model) to be realised over the variables visible to the ESDM. Each container provides a hierarchical namespace holding references to one or multiple variables together with metadata. Variables cannot be deleted while they are referenced by a container. The use of these dynamic containers provides support for much more flexible organisation of data than provided by a regular file system semantics — and efficiently support high level applications such as the Data Reference Syntax[2].

A container provides the ESDM storage abstraction which is analogous to an external file. Because many scientific metadata conventions are based on semantic structures which span variables within a file in ways that may be opaque to the ESDM without the use of a plugin, the use of a container can indicate to the ESDM that these variables are linked even though the ESDM does not understand why, and so they cannot be independently deleted. When entire files in NetCDF format are ingested into the ESDM, the respective importing tool must create a container to ensure such linking properties are not lost.


  • Creation and deletion of containers.

  • Creation and deletion of names in the hierarchical name space; the creation of links to an existing variable.

  • Attaching and modification of metadata.

  • Integrity can be checked.


Can be associated with all the other first class entities (variables, fragments, and containers). Such metadata is split into internal ESDM technical metadata, and external user-supplied semantic metadata.

Technical metadata covers, e.g., permissions, information about data location and timestamps. A backend will be able to add its own metadata to provide the information to lookup the data for the fragment from the storage backend managed by it. Metadata by itself is treaded like a normal ESDM variable but linked to the variable of choice. The implementation may embed (simple) metadata into fragments of original data (see Reference).


  • Uses can create, read, or delete arbitrary scientific metadata onto variables and containers. A future version of the ESDM may support user scientific metadata for fragments.

  • Container level metadata is generally not directly associated with variables, but may be retrieved via following references from variables to containers.

  • Queries allow to search for arbitrary metadata, e.g., for objects that have (experiment=X, model=Y, time=yesterday) returning the variables and containers in a list that match. This enables to locate scientific data in an arbitrary namespace.


A reference is a link between entities and can be used in many places, references can be embedded instead of real data of these logical objects. For example, dimensions inside a variable can be references, also a container typically uses references to variables.


  • A reference can be created from existing logical entities or removed.


ESDM does not offer a simple hierarchical namespace for the files. It provides the elementary functions to navigate data: teleportation and orientation in the following fashion: Queries about semantical data properties (e.g., experiment=myExperiment, model=myModel, date=yesterday) can be performed returning a list of matching files with their respective metadata. Iterating the list (orientation) is similar to listing a directory in a file system.

Note that this reduces the burden to define a hierarchical namespace and for data sharing services based on scientific metadata. An input/output container for an application can be assembled on the fly by using queries and name the resulting entities. As a container provides a hierachical namespace, by harnessing this capability one can search for relevant variables and map them into the local file system tree, accessing these variables as if they would be, e.g., NetCDF files. By offering a FUSE client, this feature also enables backwards compatibility for legacy POSIX applications.

Supported data backends

ESDM uses at least one storage backends but can use multiple to store the data and exactly one metadata backend.

DDN Web Object Scaler (WOS)

DDN’s Web Object Scaler or WOS is a distributed object storage system designed for extremely large data. It is available as a complete hardware/software solution or can be deployed on third-party hardware. Although, the nodes can be geographically distributed, they present a shared pool for object storage. The data can be accessed over HTTP/REST, WebDAV, Swift, and S3 interfaces. Additionally, the pools can be mounted via NFS and CIFS/SMB.

DDN Infinite Memory Engine (IME)

IME is software with a server and a client component. Rather than issuing I/O to a parallel file system client, the IME client intercepts the I/O fragments and issues these to the IME server layer which manages the NVM media and stores and protects the data. Prior to synchronizing the data to the backing file system, IME coalesces and aligns the I/O optimally for the file system. The read case works in the reverse: file data is ingested into the cache efficiently in parallel across the IME server layer and will satisfy reads from here in fragments according to the read request. IME manages at-scale flash to eliminate file system bottlenecks and the burden of creating and maintaining application-specific optimizations. It delivers:

  • New levels of I/O performance for predictable job completion in even the most demanding and complex high-performance environments.

  • Performance scaling independent of storage capacity for system designs with order of magnitude reductions in hardware.

  • Application transparency that eliminates the need to create and maintain application-specific optimizations.

Seagate Motr (MOTR)

Motr is an object storage system developed by Seagate to overcome typical limitations of traditional storage systems. In contrast to similar I/O storage system (e.g. Ceph and DAOS) Motr assesses raw block devices directly.

The design of Motr supports raw data and metadata. To achieve that it offers two types of objects: (1) A common object which is an array of fixed-size of blocks. Data can be read from and written to these objects. (2) An index for a key-value store. Key-value records can be put to and get from an index. At the moment, Motr provides a C interface.

Portable Operating System Interface (POSIX)

The Portable Operating System Interface (POSIX) is a family of standards specified by the IEEE Computer Society for maintaining compatibility between operating systems. POSIX defines the application programming interface (API), along with command line shells and utility interfaces, for software compatibility with variants of Unix and other operating systems.

Kove Direct System Architecture (KDSA)

Kove External Memory (Kunkel and Betke 2017) allows the CPU access to unlimited memory right below the CPU cache. Any data set can live in memory, close to the core, reducing processing time. The KDSA API is a low-level API that allows to access data on external memory by utilizing RDMA. Data can be transferred synchronously or asynchronously, additionally, memory can be pre-registered for use with the Infiniband HCA. Since registration of memory is time consuming, for unregistered memory regions the system may either use an internal (pre-registered) buffer and copy the user’s data to the buffer, or for larger accesses it registers the memory, performs an RDMA data transfer and then unregisters the memory again.

Amazon Simple Storage Service (S3)

Amazon Simple Storage Service (S3) is a scalable, web-based, high-speed cloud storage service with a simple-to-use API. The service allows saving and archiving of data reliably in the Amazon Web Services (AWS) of all sizes. It is suitable for use cases, like backup, archives, big-data, IoT devices, websites and much more.


Figure 1: Conversion paths.Figure 1: Conversion paths.

The documentation is organized as a set of latex and resource files. This structure simplifies integration in other documents, if only particular parts of the documentation are required. The source files are located in ./doc/latex. This is a central location, where files can be modified, if changes are required. All other locations listed in this sections are auto-generated. Therefore, after recompilation all changes will be overwritten.

Documentation centralization reduces documentation efforts. As shown in , the latex source files are exported to different formats. Change in the latex files will affect all documents.

PDF and Github-Markdown format don’t include API reference. You find API reference in the Latex and HTML formats generated by doxygen.

The listings in this section assume that current working directory is the ESDM repository. The subsections discuss the supported export possibilities in detail.


Latex documentation can be compiled by commands shown in . The result will be stored in ./doc/latex/main.pdf

cd ./doc

HTML and PDF with API reference

Doxygen depends a set of auto-generated markdown files, which should never be modified manually, because all changes will be overwritten by next compilation. Please always work with files in ./doc/latex folder and compile them running the make script as shown in .

cd ./doc

The resulting *.md files are generated in ./doc/markdown directory. Now all required source files for doxygen should be available and the final documentation can compiled by doxygen as shown in .

cd build

The resulting HTML start page is located in ./build/doc/html/index.html and main Latex document is located in ./build/doc/latex/refman.tex.

cd doc/latex
pdflatex refman.tex

The resulting PDF file is ./build/doc/latex/refman.pdf


The Github documentation ./ is generated by pandoc from the ./doc/latex/main.tex file. ./ shold be never modified manually, since changes will be overwritten by the next documentation compilation. The compilation commands are shown in .

cd ./doc

Building instructions

This guide documents installation procedures to build prerequisites as well as the prototype code base for development and testing purposes.

Figure 2: Dependencies overviewFigure 2: Dependencies overview

Dependencies (Spack)

Although, the software trees on supercomputer may provide all the standard tools and libraries required to build ESDM, it is still recommended to build, to install, and to manage dependencies via Spack[3]. The instructions below show how to setup Spack and how to install the ESDM dependencies.

  1. Download and enable Spack

        git clone --depth=2 spack
        . spack/share/spack/
  2. Set a gcc version to be used

  3. Install dependencies

        spack install gcc@$gccver +binutils
        spack compiler find
        spack install autoconf%gcc@$gccver
        spack install openmpi%gcc@$gccver gettext%gcc@$gccver
        spack install jansson%gcc@$gccver glib%gcc@$gccver
  4. Load dependencies

        spack load gcc@$gccver
        spack load -r autoconf%gcc@$gccver
        spack load -r libtool%gcc@$gccver
        spack load -r openmpi%gcc@$gccver
        spack load -r jansson%gcc@$gccver
        spack load -r glib%gcc@$gccver

ESDM prototype

Assuming all prerequisites have been installed and tested, ESDM can be configured and build as follows.

  1. Ensure environment is aware of dependencies installed using spack and dev-env

        git clone
  2. Configure and build ESDM

        cd esdm
        pushd deps
        ./configure \
          --prefix=${PREFIX} \
        cd build
        make install

ESDM-NetCDF and ESDM-NetCDF-Python

  1. Clone the NetCDF-ESDM repository

        git clone
  2. Configure and build NetCDF-ESDM. ($INSTPATH is the installation path of ESDM.)

        cd esdm-netcdf-c
        ./configure \
          --prefix=$prefix \
          --with-esdm=$INSTPATH \
          LDFLAGS="-L$INSTPATH/lib" \
          CFLAGS="-I$INSTPATH/include" \
          CC=mpicc \
        make -j
        make install
  3. If required, install the netcdf4-python module. Change to the root-directory of the esdm-netcdf repository and install the patched netcdf-python module.

        cd dev
        git clone
        cd netcdf-python
        patch -s -p1 < ../v2.patch
        python3 install --user


For the development of ESDM the directory ./dev/docker contains all requirements to quickly set up a development environment using docker. The Dockerfiles contain ESDM installation instructions for different distributions. For easy building Dockerfiles for different plattforms are provided in different flavours:

  • CentOS/Fedora/RHEL like systems

  • Ubuntu/Debian like systems


Assuming the docker service is running you can build the docker images as follows:

cd ./dev/docker
cd <choose ubuntu/fedora/..>
sudo docker build -t esdm .

After docker is done building an image, you should see the output of the ESDM test suite, which verifies that the development environment is set up correctly. The output should look similar to the following output:

Running tests...
Test project /data/esdm/build
Start 1: metadata
1/2 Test #1: metadata .........................   Passed    0.00 sec
Start 2: readwrite
2/2 Test #2: readwrite ........................   Passed    0.00 sec

100% tests passed, 0 tests failed out of 2

Total Test time (real) =   0.01 sec


Running the esdm docker container will run the test suite:

sudo docker run esdm  # should display test output

You can also explore the development environment interactively:

sudo docker run -it esdm bash

docker run dev/docker/ubuntu-whole-stack/Dockerfile


ESDM can store metadata and data separately. Furthermore, the data can be distributed over multiple storage media. How exactly the data is distributed can be specified in the configuration file. This section describes the file structure, the metadata and supported storage backends configuration and provides examples.

Search paths

The current implemention reads esdm.conf file in the current directory. In the next ESDM versions it is planned to support other common paths as well, like /etc/esdm/esdm.conf, ~/.config/esdm/esdm.conf, ~/.esdm.conf, ./esdm.conf. In addition, the configuration should also be possible via the environment variable and arguments.

File format

The configuration file format is based on JSON, i.e. all configuration files must be valid JSON files. However, ESDM places some restrictions on what keys are recognized and which types are expected. The tables summarize parameters that can be used in the configuration file. Parameter can be of four types: string, integer, float and object, which are collections of key-value pairs.

The "esdm":{} key-value pair in the root of the JSON configuration file can contain configuration for multiple data backends and one metadata backend. The backends are organized in the "backends":[] key-value pair as a list.

  "esdm":	{
    "backends": [
      <insert backend config here>,
      <insert backend config here>,
      <insert backend config here>,
    "metadata": {
      <insert metadata config here>,

	"esdm":	{
		"backends": [
				"type": "POSIX",
				"id": "p1",
				"target": "./_posix1",
				"type": "POSIX",
				"id": "p2",
				"target": "./_posix2",
		"metadata": {
			"type": "metadummy",
			"id": "md",			"target": "./_metadummy"

Data parameters

Parameter Type Default Description
id string (not set) required Unique identifier
type string (not set) required Backend name
target string (not set) required Connection specification (Bucket name for S3)
performance-model object (not set) optional Performance model definition.
max-threads-per-node integer 0 optional Maximum number of threads on a node.
write-stream-blocksize integer 0 optional Blocksize in bytes used to write fragments.
max-global-threads integer 0 optional Maximum total number of threads.
accessibility string global optional Data access permission rights.
max-fragment-size integer 10485760 optional Maximum fragment size in bytes.
fragmentation-method string contiguous optional Fragmentation methods.

Backend configuration parameters overview

Parameter: /esdm/backends/id

Unique alpha-numeric identifier.

Type string
Default (not set)
Required yes

Parameter: /esdm/backends/type

Storage backend type.

Type string
Default (not set)
Required yes

Can take a value of one of the supported backends.

Type Description
MOTR Seagate Object Storage API
DUMMY Dummy storage (Used for development)
IME DDN Infinite Memory Engine
KDSA Kove Direct System Architecture
POSIX Portable Operating System Interface
S3 Amazon Simple Storage Service
WOS DDN Web Object Scaler

Parameter: /esdm/backends/target

The target specifies connection to storage. Therefore, its value depends on the storage type. The following sections describe targets for each storage type and provide an example.

Type string
Default (not set)
Required yes
Type = MOTR

If Seagate Object Storage API is selected, then the target string is composed of four parameters separated by spaces.

[local_addr] [ha_addr] [prof] [proc_fid]


  • "local_addr" is the local address used to connect to Motr service,

  • "ha_addr" is the hardware address in Motr service,

  • "prof" is the profile FID in the Motr service, and

  • "proc_fid" is the process FID in the Motr service.

  "type": "MOTR",
  "id": "c1",
  "target": ":12345:33:103 
  <0x7000000000000001:0> <0x7200000000000001:64>"
Type = DUMMY

Used for development.

  "type": "DUMMY",
  "id": "c1",
  "target": "./_dummy",
Type = IME

The target is the IME storage mount point.

  "type": "IME",
  "id": "p1",
  "target": "./_ime"
Type = KDSA

Prefix “xpd:” followed by volume specifications. Multiple volume names can be connected by “+” sign.


A module specification consists of several parts. Caller provides a device handle, the target serial number, the target link number, the volume UUID, The convention for specifying a KDSA volume uses the following format:


where the square brackets indicate optional parts of the volume connection specification. Thus, a volumeid is nominally sufficient to specify a desired volume, and one can then optionally additionally specify the serial number of the XPD with optional link number, and/or one can optionally specify the local device to use with optional local port number. The convention for specifying multiple KDSA volumes to stripe together uses the following format:


where the square brackets indicate optional parts of the aggregated connection specification. Thus, a single volume connection specification is sufficient for a full connection specifier, and one can then optionally specify additional volume specifiers to aggregate, using the plus sign as a separator. The user may also additionally specify parameters for the aggregation, using the “at sign,” a single character as a parameter identifier, and the parameter value.

  "type": "KDSA",
  "id": "p1",
  "target": "This is the XPD connection string",
Type = POSIX

The target string is the path to a directory.

  "type": "POSIX",
  "id": "p2",
  "target": "./_posix2"
Type = S3

The target string is a bucket name with at least a 5 characters. A proper S3 configuration requires at three additional parameters: host, secret-key and access-key. Other parameters are optional.

Parameter Type Default Description
host string (not set) required A valid endpoint name for the Amazon S3 region provided by the agency.
secret-key string (not set) required Secret Access Key for the account.
access-key string (not set) required Access key ID for the account.
locationConstraint string (not set) optional Specifies the Region where the bucket will be created. If you don’t specify a Region, the bucket is created in the US East (N. Virginia) Region (us-east-1). For a list of all the Amazon S3 supported regions, see API Bucket reference (
authRegion string (not set) optional For a list of all the Amazon S3 supported regions and endpoints, see regions and endpoints in the AWS General Reference (
timeout integer (not set) optional Request timeout in milliseconds.
s3-compatible integer (not set) optional TODO (not used ?)
use-ssl integer 0 optional Use HTTPS for encryption, if enabled.
  "type": "S3",
  "id": "p1",
  "target": "./_posix1",
  "access-key" : "",
  "secret-key" : "",
  "host" : "localhost:9000"
Type = WOS

The target string is a concatenation of key=value; pairs. The host key specifies the WOS service host. The policy name is that one, that was set up on the DDN WOS console.

  "type": "WOS",
  "id": "w1",
  "target": "host=;policy=test;",

Parameter: performance-model

Parameter Type Default Domain Description
latency float 0.0  >  = 0.0 optional seconds
throughput float 0.0  > 0.0 optional MiB/s

Dynamic performance model parameters

Parameter Type Default Domain Description
latency float 0.0  >  = 0.0 optional Latency in seconds
throughput float 0.0  > 0.0 optional Throughput in MiB/s
size integer 0  > 0 optional TODO
period float 0.0  > 0.0 optional TODO
alpha float 0.0 [0.0, 1.0) optional TODO

Generic performance model parameters

Parameter: max-threads-per-node

Maximum number of threads on a node. If max-threads-per-node=0, then the ESDM scheduler estimates an optimal value.

Type integer
Default 0
Required no

Parameter: write-stream-blocksize

Block size in bytes used to read/write fragments. If write-stream-blocksize=0, then ESDM estimates optimal block size for each storage type individually.

Type integer
Default 0
Required no

Parameter: max-global-threads

Maximum total number of threads. If max-global-threads=0, then the ESDM scheduler estimates an optimal value.

Type integer
Default 0
Required no

Parameter: accessibility

Data access permission rights.

Type string
Default global
Required no
Method Description
global data is accessible from all nodes, e.g., shared file system
local data is accessible from local node only

Parameter: max-fragment-size

The amount of data that may be written into a single fragment. The amount is given in bytes.

Type integer
Default 10485760
Required no

Parameter: fragmentation-method

A string identifying the algorithm to use to split a chunk of data into fragments.

Type string
Default contiguous
Required no

Legal values are:

Method Description
contiguous This algorithm tries to form fragments that are scattered across memory as little as possible. As such, it is expected to yield the best possible write performance. However, if a transposition is performed when reading the data back, performance may be poor. Splitting a dataspace of size (50, 80, 100) into fragments of 2000 entries results in fragments of size (1, 20, 100).
equalized This algorithm tries to form fragments that have a similar size in all dimensions. As such, it is expected to perform equally well with all kinds of read requests, but it tends to write scattered data to disk which has to be sequentialized first, imposing a performance penalty on the write side. Splitting a dataspace of size (50, 80, 100) into fragments of at most 2000 entries results in fragments of sizes between (10, 11, 11) and (10, 12, 12).

Fragmentation methods

Metadata parameters

Parameter Type Default Required Description
id string (not set) (not used) Unique alpha-numeric identifier
type string (not set) yes (?)

Metadata parameters overview

Parameter: /esdm/metadata/id

Unique alpha-numeric identifier.

Type string
Default (not set)
Required yes

Parameter: /esdm/metadata/type

Metadata type. Reserved for future use.

Type string
Default (not set)
Required yes (not used)

Parameter: /esdm/metadata/target

Path to the metadata folder.

Type string
Default (not set)
Required yes


This section illustrated how ESDM can be used with different programming languages and tools. The applications with NetCDF support don’t need to be changed or recompiled as long as they are linked against the shared netcdf library. They can simply use the ESDM-NetCDF version and benefit of ESDM. To force them to use the ESDM-NetCDF version, the library can be pre-loaded by LD_PRELOAD environment variable.


NetCDF (Network Common Data Form) (“Network Common Data Form (NetCDF),” n.d.) is a set of software libraries and machine-independent data formats that support the creation, access, and sharing of array-oriented scientific data. In particular, NetCDF is widely used in climate research.

The following examples write and read data to/from a NetCDF file. (For simplicity, we omitted the error handling.)

The write example (./write) creates a two dimensional variable in the esdm:// file.

#include <netcdf.h>

#define NDIMS 2
#define NX 6
#define NY 12

int main() {
  int ncid, x_dimid, y_dimid, varid;
  int dimids[NDIMS];
  int data_out[NX][NY];
  int x, y, retval;

  for (x = 0; x < NX; x++)
    for (y = 0; y < NY; y++)
      data_out[x][y] = x * NY + y;

  nc_create("esdm://", NC_CLOBBER, &ncid);
  nc_def_dim(ncid, "x", NX, &x_dimid);
  nc_def_dim(ncid, "y", NY, &y_dimid);

  dimids[0] = x_dimid;
  dimids[1] = y_dimid;

  nc_def_var(ncid, "var1", NC_INT, NDIMS, dimids, &varid);
  nc_put_var_int(ncid, varid, &data_out[0][0]);
  return 0;

The read example (./read) reads this variable.

#include <stdlib.h>
#include <stdio.h>
#include <netcdf.h>

int main() {
  int ncid, varid;
  int data_in[6][12];
  int x, y, retval;

  nc_open("esdm://", NC_NOWRITE, &ncid);
  nc_inq_varid(ncid, "var1", &varid);
  nc_get_var_int(ncid, varid, &data_in[0][0]);

  // do something with data ...

  return 0;

For a successful run ESDM requires three things. Firstly, the esdm.conf file must be located in the search path. In this case, it is the current directory. Secondly, ESDM requires an initialized ESDM directory structure. If not existing, it can be created by the mkfs.esdm tool. Thirdly, files need to be prefixed with esdm://. This prefix is an indicator to the NetCDF library to redirect data to ESDM. Otherwise, NetCDF will work as usual and create an *.nc file.

The following script initialize the ESDM directory structure and runs the two examples.


mkfs.esdm -g -l --create  --remove --ignore-errors

Currently, ESDM does not provide a full compatibility to NetCDF. The NetCDF compatibility is documented in the report released in


The Climate Data Operator (CDO) software is a collection of many operators for standard processing of climate and forecast model data. The operators include simple statistical and arithmetic functions, data selection and subsampling tools, and spatial interpolation. CDO was developed to have the same set of processing functions for GRIB and NetCDF datasets in one package. The Climate Data Interface [CDI] is used for the fast and file format independent access to GRIB and NetCDF datasets. To enable ESDM, the files need the esdm:// prefix. In the example data is converted from GRIP to NetCDF format and written by ESDM according to the configuration in the esdm.conf file.


mkfs.esdm -g -l --create  --remove --ignore-errors
cdo -f nc -copy esdm:// 


XIOS (“XIOS a library dedicated to I/O management in climate codes,” n.d.), standing for XML-IO-Server, is a library dedicated to I/O management, primarily designed to manage NetCDF outputs of climate models. Its main features are management of output diagnostics and history files, temporal post-processing operations, (e.g., average, min, max) and spatial post-processing operations (e.g,. arithmetic operations, re-griding). XIOS simplifies I/O management, by reducing the number of I/O operations and arguments in the source code. The output definitions are outsourced into XML files, which can be changed without re-compilation of the source code. XIOS achieves high performance by asynchronous and parallel I/O.

  1. Prerequisites

    • Install ESDM

    • Install ESDM-NetCDF

    • Install NetCDF-Fortran interface

  2. Checkout XIOS code

        svn checkout
  3. Add to bld.cfg

        bld::tool::cflags the ESDM include directory of the installation
        bld::tool::ldflags the ESDM lib directory and add it also as rpath -Wl,--rpath=<ESDM lib directory>
  4. Modify arch.env. Change the NetCDF directory to the ESDM dir

  5. On some Linux distribution gmake may be not available. Fortunately, make provides the necessary functionality.

        sudo ln -s /usr/bin/make /usr/bin/gmake
        ./make_xios --dev --arch GCC_LINUX
  6. Check that the proper NetCDF library was linked. This should output the NetCDF library created from ESDM.

        ldd bin/test_client.exe |grep netcdf
  7. Fix iodef.xml file:

        cp inputs/iodef.xml .
        Change  <file id="output" name="esdm://output" enabled=".TRUE.">
  8. Create ESDM configuration file in the local directory

        cp <ESDM install directory>/src/test/esdm.conf .
  9. Create ESDM data directories

        export PATH=<ESDM install directory>/bin/:$PATH
        mkfs.esdm -g -l --create  --remove --ignore-errors
  10. Run XIOS MPMD with OpenMPI:

        mpiexec -np 2 bin/xios_server.exe : -np 3 bin/test_client.exe


The patched NetCDF4-Python module is able to redirect I/O to ESDM if the file name contains the esdm:// prefix, esdm.conf is located in the search path, and the ESDM directory structure is initialized with the mkfs.esdm tool. The following example illustrates write access to a NetCDF file.

#!/usr/bin/env python3

from netCDF4 import Dataset
import numpy as np

root_grp = Dataset('esdm://', 'w', format='NETCDF4')
#root_grp = Dataset('', 'w')
root_grp.description = 'Example simulation data'

ndim = 128 # Size of the matrix ndim*ndim
xdimension = 0.75
ydimension = 0.75

# dimensions
root_grp.createDimension('time', None)
root_grp.createDimension('x', ndim)
root_grp.createDimension('y', ndim)

time = root_grp.createVariable('time', 'f8', ('time',))
x = root_grp.createVariable('x', 'f4', ('x',))
y = root_grp.createVariable('y', 'f4', ('y',))
field = root_grp.createVariable('field', 'f8', ('time', 'x', 'y',))

x_range =  np.linspace(0, xdimension, ndim)
y_range =  np.linspace(0, ydimension, ndim)
x[:] = x_range
y[:] = y_range
for i in range(5):
   time[i] = i*50.0
   field[i,:,:] = np.random.uniform(size=(len(x_range), len(y_range)))



Dask is able to utilize ESDM by redirection I/O to the patched NetCDF4-Python module. By passing the engine=netcdf4 argument to the to_netcdf() function we can make sure, that the correct engine is used in the background. We just have to take care, that esdm.conf is located in the search path and that the ESDM directory structure is initialized by the mkfs.esdm tool.

#!/usr/bin/env python
import xarray as xr
import numpy as np
import pandas as pd
ds = xr.Dataset(
   {"var1": [1,2,3,4,5,6]},
ds.to_netcdf(path="esdm://", format='NETCDF4', engine="netcdf4")
#!/usr/bin/env python3
import xarray as xr
ds = xr.open_dataset('esdm://', engine='netcdf4')


NetCDF Performance Benchmark Tool (NetCDF-Bench) was developed to measure NetCDF I/O performance on devices ranging from notebooks to large HPC systems. It mimics the typical I/O behavior of scientific climate applications and captures the performance on each node/process. In the end, it aggregates the data to the human-readable summary.

Initialize ESDM directory structure by means of mkfs.esdm and make sure esdm.conf is located in the search path. The benchmark can be run by the following script.

mkfs.esdm -g -l --create  --remove --ignore-errors
benchtool -r -w -f=esdm://


ParaView is an open-source, multi-platform data analysis and visualization application. It can load a variety of different data formats, visualize the data and analyse it by qualitative and quantitative techniques. In particular, Paraview supports the NetCDF data format and different grid types, and is often used in climate and earth science. The data import capabilities in Paraview are implemented as plugins. The NetCDF Plugin is able to load NetCDF files with CF conventions and the CDI Reader Plugin is able to load ICON data (Crueger et al. 2018). We could successfully test both them.

The ESDM-NetCDF library need to be linked to Paraview. On Linux this can be done by LD_PRELOAD feature, like shown in the following script. The $INSTPATH variable show to the ESDM-NetCDF library location.


A NetCDF file was converted to ESDM representation and loaded in Paraview via the NetCDF plugin. Loading data works slightly different compared to ordinary NetCDF files. If the pseudo ESDM file (an empty file with the :esdm suffix) exists, then it can be selected in the open dialog as usual. Otherwise, the user have to know the file name and must enter it manually. Another difference is that on the left side in the GUI (see ) the data set is labeled by the :esdm suffix. Once data is loaded, other steps can be done as usual.

Figure 3: Visualization of ice data set in Paraview. Data source is an ESDM file.Figure 3: Visualization of ice data set in Paraview. Data source is an ESDM file.

Developers guide

This document aims to give a quick conceptual overview of the API that is available to programs which link directly to ESDM. It is not meant to be exhaustive, but rather to give enough background information so that reading the documentation of the ESDM API becomes easy. The intended audience of this guide are developers of scientific applications who wish to benefit of the full potential of the ESDM API.

Readers who are interested in using an ESDM file system with a NetCDF based application should read the users guide instead. Unfortunately, there is no guide for ESDM backend developers and ESDM core contributors yet.

Internal data model

The first thing to understand is the data model used by ESDM. This data model is very similar to that employed by NetCDF, but it adds some abstractions. This section describes the model.


A data space describes the layout of data in memory. It is a mapping of logical data space to sequential bytes in memory. All data handled by ESDM is associated with a data space. Mostly the data spaces are user-provided. Several copies of the same data may use distinct datas paces. In this case, ESDM allows users to transform data from one layout to another by providing a source and a destination data space (esdm_dataspace_copy_data()). User code needs to provide a data space when it defines a variable (here, the data layout part is irrelevant), when it stores data, and when it reads data.

The logical data space is always a multidimensional, axis-aligned box in ESDM. As such, a data space consists of the following information:

  • the dimension count

  • start and end coordinates for each axis

  • a data type describing a single value within the multidimensional array

  • (data serialization information)

The data serialization information is usually implicit. By default, ESDM assumes a C multidimensional array layout. Fortran programs will need to set explicitly the data serialization information to match Fortran’s inverse dimension order. The data serialization information (stride) can also achieve unorthodox effects like arrays with holes or replicating a single 2D slice along a third axis.

Creating and destroying dataspaces

ESDM provides two distinct mechanisms to create a data space: A generic API that allocates the data space on the heap and an API to create a throw-away data space on the stack quickly.

The generic API

The functions used to create dataspaces are:

  • esdm_dataspace_create() Construct a data space with a given dimension count, size, and datatype. It assumes that the hypercube starts at (0, 0, ..., 0) and C data layout.

  • esdm_dataspace_create_full() Like esdm_dataspace_create(), but allows the user to specify start coordinates.

  • esdm_dataspace_copy() Copy constructor.

  • esdm_dataspace_subspace() Create a data space that contains a subset of the logical data space of its parent data space. It copies the datatype and assumes the C data layout for the subspace. If this is not desireable, follow up with a call to esdm_dataspace_copyDatalayout().

  • esdm_dataspace_makeContiguous() Create a data space with the same logical space and datatype, but which uses the C data layout.

All these functions return a pointer to a new data space which must be destroyed with a call to esdm_dataspace_destroy().

Data layout can only be set explicitly after a data space has been created. This is done by a call to esdm_dataspace_set_stride() or to esdm_dataspace_copyDatalayout(). The first allows the user to specify any standard data layout, including but not limited to Fortran’s data layout. The latter assumes that the data space will access the same data buffer as the data space from which the data layout is copied. As such, esdm_dataspace_copyDatalayout() is very convenient for data subseting operations.

The simple API

For convenience and performance reasons, ESDM provides preprocessor macros that allocate a data space on the stack. Macros are provided for 1D, 2D, and 3D data spaces and come in variants that either assume start coordinates at the origin of the coordinate system, or allow the offset to be specified explicitly. The macros without explicit start coordinates are:


The macros that take start coordinates simply add an "o" to the end of the name:


The result of these macros is a value of type esdm_simple_dspace_t, which is a struct containing a single member ptr which includes the pointer value that can subsequently be used in all ESDM calls that require a data space, i.e., a typical usage might be:

esdm_simple_dspace_t region = esdm_dataspace_2do(x, width, y, height, type);
esdm_read(dataset, buffer, region.ptr);

As the esdm_simple_dspace_t lives on the stack, it must not be destroyed with esdm_dataspace_destroy(). It ceases to exist when the surrounding block exits.


A dataset in ESDM is what a variable is in NetCDF or HDF5. Each data set is associated with a data space that describes the logical extends of the data, and it acts as a container for data written into this logical space.

There is no requirement to fill the entire logical space with data. Usually, reading nonexistent data results in an error. However, a dataset can also be associated with a fill value to handle data with holes seamlessly.

There is also no requirement to write non-overlapping data. When write accesses overlap, ESDM will assume that both accesses place the same data in the overlapping area. If this condition does not hold, there is no guarantee which data will be returned on a read.

In addition to the data and the logical space, datasets can also contain several attributes. Like attributes in NetCDF, these are meant to associate metadata with a dataset.

User code can either create a dataset with esdm_dataset_create() or look up an existing dataset from a container with esdm_dataset_open(). In either case, the reference to the dataset must later be dropped by a call to esdm_dataset_close(). A dataset can also be deleted with a call to esdm_dataset_delete() which will remove its data from the file system, as well as its name and link from its container.


Containers provide a means to organize and retrieve datasets. When a dataset is created, it is added to a container and associated with a name for later retrieval.

Like datasets, containers are created with esdm_container_create() or looked up from the file system with esdm_container_open() and need to be closed with esdm_container_close() when their reference is not required anymore. Closing a container requires closing all datasets it contains first. esdm_container_delete() removes a container from the file system.


The grid abstraction exists for performance reasons only: While it is possible to think of a data set as a set of possibly overlapping chunks of data, it is surprisingly hard to determine minimal sets of chunks to satisfy a read request. On the other hand, user code generally does not use overlapping chunks of data. Instead, user code can be assumed to work on (semi-)regular non-overlapping chunks of data. Passing this chunking information to ESDM allows the library to make the right decisions faster.

Grids also allow user code to inquire how the data is available on disk, allowing consumer code to iterate over complete and non-overlapping sets of data chunks in the most efficient fashion. To work with grids, the header esdm-grid.h must be included.

Like a data space, a grid covers an axis-aligned hyper box in the logical space. This space is called the grid’s domain, and it is defined when a grid is created with esdm_grid_create(), esdm_grid_createSimple() allows omitting the start coordinates to use the origin as one corner of the hyper box.

Once a grid has been created, its axes can be subdivided individually by calls to esdm_grid_subdivide(). It allows the user to specify all the bounds for an axis explicitly. In many contexts, however, it will be simpler to use esdm_grid_subdivideFixed() or esdm_grid_subdivideFlexible() which instruct ESDM to generate bounds in a regular way. The fixed subdivision will produce intervals of a given size. Flexible subdivision instructs ESDM to generate a specific number of similar size intervals.

After the grid axis has been defined, individual grid cells may be turned into subgrids via esdm_grid_createSubgrid(). After this, the parent grid axes are fixed, and the subdivision calls cannot be used anymore. On the other hand, the subrid is a newly created grid with the parent grid’s cell bounds as its domain. Usually, user code will follow up with calls to esdm_grid_subdivide*() on the subgrid to define its axes. Subgrids may be constructed recursively to any depth.

This subgrid feature is useful to define grids with semi-regular decompositions: For instance, an image may be decomposed into stripes, which are themselves decomposed into rectangles, but the rectangle bounds of one stripe do not match those of another stripe. Such semi-regular decompositions are a typical result of load balancing of earth system simulations.

Once a grids structure has been defined fully, it can be used to read/write data via esdm_read_grid() and esdm_write_grid(). Parallel applications will want to distribute the grid to all involved processes first by calling esdm_mpi_grid_bcast() (include esdm-mpi.h for this). Both use a grid for input/output, and communicating it over MPI will fix the grid’s structure, prohibiting future calls to subdivide axes or create subgrids.

It is possible to iterate over all (sub-)cells of a grid. This is done using an iterator, the three relevant functions are esdm_gridIterator_create(), esdm_gridIterator_next() and esdm_gridIterator_destroy(). This is meant to be used by readers who inquire about an available grid from a dataset using esdm_dataset_grids(). This method of reading avoids any cropping or stitching together of data chunks within ESDM, delivering the best possible performance.

Grids always remain in possession of their data space. Consequently, it is not necessary to dispose of them explicitly. However, closing a data space invalidates all associated (sub-)grids.

Usage examples

Learning usage of an API is easiest by seeing it in action in concrete examples. As such, this section provides four relatively basic examples of how the ESDM API is supposed to be used, which nevertheless cover all the necessary core functionality.

Basic writing

The simplest way to write a greyscale image to ESDM is as follows:

//assume image data stored in either
uint16_t imageBuffer[height][width];
uint16_t (*imageBuffer)[width] = malloc(height*sizeof*imageBuffer);

//initialize ESDM
esdm_status result = esdm_init();
assert(result == ESDM_SUCCESS);

//create the container, dataspace, and dataset
esdm_container_t* container;
//pass true to allow overwriting of an existing container
result = esdm_container_create("path/to/container", false, &container); 
assert(result == ESDM_SUCCESS);
//the data is 2D and consists of uint64_t values
esdm_simple_dspace_t space = 
  esdm_dataspace_2d(height, width, SMD_TYPE_UINT16);  esdm_dataset_t* dataset;
result = esdm_dataset_create(container, "myImage", space.ptr, &dataset);
assert(result == ESDM_SUCCESS);

//write the data
result = esdm_write(dataset, imageBuffer, space.ptr);
assert(result == ESDM_SUCCESS);

result = esdm_dataset_close(dataset);
assert(result == ESDM_SUCCESS);
result = esdm_container_close(container);
assert(result == ESDM_SUCCESS);

//bring down ESDM
result = esdm_finalize();
assert(result == ESDM_SUCCESS);

In this example, the same dataspace is used to create the dataset and write the data, writing all the data in one large piece. Although it is not necessary, the data spaces that are passed to esdm_write() may be smaller than the data set, calling esdm_write() as many times as required to write the entire data.

Grid-based writing

When using grid-based writing, creating of the container and the dataset is the same. The creation of the grid, however, is added explicitly. In this case, we are going to slice a 3D data space into ten slices of similar size along the z-axis and into rows of 256 lines along the y axis:

// define the grid
esdm_grid_t* grid;
result = esdm_grid_createSimple(dataset, 3, (int64_t[3]){depth, height, width}, &grid);
assert(result == ESDM_SUCCESS);
// the last parameter allows the last interval to be smaller than 256 lines
result = esdm_grid_subdivideFixed(grid, 1, 256, true);  
assert(result == ESDM_SUCCESS);
result = esdm_grid_subdivideFlexible(grid, 0, 10);
assert(result == ESDM_SUCCESS);

for(int64_t z0 = 0, slice = 0, curDepth; z0 < depth; z0 += curDepth, slice++) {
    // inquire the depth of the current slice
    // use of esdm_grid_subdivideFlexible() generally 
    // requires use of a grid bounds inquiry function
    int64_t z1;
    result = esdm_grid_getBound(grid, 0, slice + 1, &z1);
    assert(result == ESDM_SUCCESS);
    curDepth = z1 - z0;

    for(int64_t row = 0; row*256 < height; row++) {
        // compute the height of the current row
        // we can calculate this ourselves as we have used esdm_grid_subdivideFixed()
        int64_t height = (row < height/256 ? 256 : height - row*256);

        // set contents of dataBuffer

        // use the grid to write one chunk of data, the grid knows 
        // to which dataset it belongs
        result = esdm_write_grid(
            esdm_dataspace_3do(z0, curDepth, row*256, height, 0, width, SMD_TYPE_DOUBLE).ptr, 
        assert(result == ESDM_SUCCESS);

// no need to cleanup the grid, it belongs to the dataset and 
// will be disposed off when the dataset is closed

Simple reading

Reading data is very similar to writing it. Nevertheless, a simple example is given to read an entire dataset in one piece:

// open the container and dataset, and inquire the dataspace
esdm_container_t* container;
result = esdm_container_open("path/to/container", ESDM_MODE_FLAG_READ, &container);
assert(result == ESDM_SUCCESS);
esdm_dataset_t* dataset;
result = esdm_dataset_open(container, "myDataset", ESDM_MODE_FLAG_READ, &dataset);
assert(result == ESDM_SUCCESS);
esdm_dataspace_t* dataspace;
// this returns a reference to the internal dataspace, do not destroy or modify it
esdm_dataset_get_dataspace(dataset, &dataspace);  

// allocate a buffer large enough to hold the data and generate a dataspace for it
// the buffer will be in contiguous C data layout
result = esdm_dataspace_makeContiguous(dataspace, &dataspace);  
assert(result == ESDM_SUCCESS);
int64_t bufferSize = esdm_dataspace_total_bytes(dataspace);
void* buffer = malloc(bufferSize);

// read the data
result = esdm_read(dataset, buffer, dataspace);
assert(result == ESDM_SUCCESS);

// do stuff with buffer and dataspace

// cleanup
// esdm_dataspace_makeContiguous() creates a new dataspace
result = esdm_dataspace_destroy(dataspace); 
assert(result == ESDM_SUCCESS);
result = esdm_dataset_close(dataset);
assert(result == ESDM_SUCCESS);
result = esdm_container_close(container);
assert(result == ESDM_SUCCESS);

Grid-based reading

Reading an entire dataset as a single chunk is generally a terrible idea. Data sets, especially those generated by earth system models, can be massive, many times larger than the available main memory. Reading a dataset in the form of reader-defined chunks is possible with esdm_read(), but not necessarily efficient. The chunks on disk may not match those which are used to read, requiring esdm_read() to

  • read multiple chunks from disk and stitch them together,

  • and to read more data from disk than is required.

If the dataset has been written using a grid, this grid can be recovered to inform the reading process of the actual data layout on disk:

    // open the container and dataset, and inquire the available grids
    esdm_container_t* container;
    result = esdm_container_open("path/to/container", ESDM_MODE_FLAG_READ, &container);
    assert(result == ESDM_SUCCESS);
    esdm_dataset_t* dataset;
    result = esdm_dataset_open(container, "myDataset", ESDM_MODE_FLAG_READ, &dataset);
    assert(result == ESDM_SUCCESS);
    int64_t gridCount;
    esdm_grid_t** grids;
    result = esdm_dataset_grids(dataset, &gridCount, &grids);
    assert(result == ESDM_SUCCESS);

    // select a grid, here we just use the first one
    assert(gridCount >= 1);
    esdm_grid_t* grid = grids[0];
    free(grids);  //we are responsible to free this array

    // iterate over the data, reading the data one stored chunk at a time
    esdm_gridIterator_t* iterator;
    result = esdm_gridIterator_create(grid, &iterator);
    assert(result == ESDM_SUCCESS);
    while(true) {
        esdm_dataspace_t* cellSpace;
        result = esdm_gridIterator_next(&iterator, 1, &cellSpace);
        assert(result == ESDM_SUCCESS);

        if(!iterator) break;

        // allocate a buffer large enough to hold the data
        int64_t bufferSize = esdm_dataspace_total_bytes(cellSpace);
        void* buffer = malloc(bufferSize);

        // read the data
        result = esdm_read_grid(grid, cellSpace, buffer);
        assert(result == ESDM_SUCCESS);

        // do stuff with buffer and cellSpace

        // cleanup
        result = esdm_dataspace_destroy(cellSpace);
        assert(result == ESDM_SUCCESS);

    // cleanup
    // no cleanup necessary for the iterator, it has already been destroyed 
    // by esdm_gridIterator_next()
    result = esdm_dataset_close(dataset);
    assert(result == ESDM_SUCCESS);
    result = esdm_container_close(container);
    assert(result == ESDM_SUCCESS);

Use Cases

This part of the documentation presents several use-cases for middleware to handle earth system data. The description is organized as follows:

  • typical workloads in climate and weather forecasts

  • involved stakeholders/actors and systems

  • and the actual use cases

The use cases can extend each other and are generally constructed so that they are not limited to the ESDM and apply to similar middleware. The use of backends is kept abstract where possible so that, in principle, implementations can be swapped with only minor semantic changes to the sequence of events.

Workloads in climate and weather

The climate and weather forecast communities have their typical workflows and objectives and share various methods and tools (e.g., the ICON model is used and developed together by climate and weather scientists). This section briefly collects and groups the data-related high-level use-cases by the community and their motivation.

Numerical weather prediction focuses on producing a short-time forecast based on initial sensor (satellite) data and generates derived data products for specific end-users (e.g., the weather forecast for the general public or military). As the compute capabilities and requirements for users increase, new services are added, or existing services are adapted. Climate predictions run for long-time periods and may involve complex workflows to compute derived information such as monthly mean or identify specific patterns in the forecasted data such as tsunamis.

A list of characteristic high-level workloads and use-cases that are typically performed per community is given in the following, These use-cases resemble what a user/scientist usually has in mind when dealing with NWP and climate simulation; there are several supportive use-cases from the data center’s perspective that will be discussed later.


  • Data ingestion: Store incoming stream of observations from satellites, radar, weather stations, and ships.

  • Pre-Processing: Cleans adjusts observation data and then transforms it to the data format used as an initial condition for the prediction. For example, insufficient sampling makes pre-processing necessary so models can be initialized correctly.

  • Now Casting (0-6h): Precise prediction of the weather in the near future. Uses satellite data and data from weather stations, extrapolates radar echos.

  • Numeric Model Forecasts (0-10+ Days): Run a numerical model to predict the weather for the next few days. Typically, multiple models (ensembles) are run with some perturbed input data. The model usually proceeds as follows: 1) Read-Phase to initialize simulation; 2) create periodic snapshots (write) for the model time, e.g., every hour.

  • Post-Processing: create data products that may be used for multiple purposes.

    • for Now Casting: multi-sensor integration, classification, ensembles, impact models

    • for Numeric Model Forecasts: statistical interpretation of ensembles, model-combination

    • generation of data products like GRIB files as service for customers of weather forecast data

  • Visualizations: Create fancy presentations of the future weather; this is usually part of the post-processing.


Many use cases in climate are very similar:

  • Pre-Processing: Similar to the NWP use case.

  • Forecasting with Climate Models: Similar to the NWP use case, with the following differences:

    • The periodic snapshots (write) uses variable depending frequencies, e.g., some variables are written out with higher frequencies than others
  • Post-Processing: create data products that are useful, e.g., run CDOs (Climate Data Operations) to generate averages for certain regions. The performed post-processing depends on the task the scientist has in mind. While some products are straightforward and may be used to diagnose the simulation run itself at the model’s runtime, later scientists might be interested in running additional post-processing to look for new phenomena.

  • Dynamic visualization: use interactive tools to search for interesting patterns. Tools such as VTK and Paraview are used.

  • Archive data: The model results are stored on a long-term archive. They may be used for later analysis – often at other sites and by another user, e.g., to search for some exciting pattern or to serve as input data for localized higher-resolution models. Also, it supports the reproducibility of research.

Stakeholders and Systems

List of Use-Cases

The use cases are organized as one document per use case. The available use cases are:

  • Independent Write

  • Independent Read

  • Simulation

  • Pre/Post Processing


Styleguide for ESDM development

This document describes the style guide to use in the code.

General rules

  • Do not break the line after a fixed number of characters as this is the duty of the editor to use some softwrap.

    • You may use a line wrap, if that increases readability (see example below).
  • Use two characters for indentation per level

  • Documentation with Doxygen needs to be added on the header files

  • Ensure that the code does not produce WARNINGS

  • Export only the functions to the user that is needed by the user

  • The private (module-internal) interface is defined in internal.h

Naming conventions

  • use lower case for the public interface

  • functions for users provided by ESDM start with esdm_

  • auxiliary functions that are used internally start with ea_ (ESDM auxiliary) and shall be defined inside esdm-internal.h

Example Code

//First add standard libraries
#include <stdio.h>
#include <stdlib.h>

// Add an empty line before adding any ESDM include file
#include <esdm-internal.h>

struct x_t{
  int a;
  int b;
  int *p;

// needs always to be split separately, 
// to allow it to coexist in a public header file
typedef struct x_t x_t; 

int testfunc(int a){
    // Additional basic block
  if (a == 5){



  // allocate variables as late as possible, 
  // such that we can see when it is needed and what it does.
  int ret;



  return 0;

Callgraph for accessing metadata

Guiding question:

  • How and when do we fetch metadata in the read path?

  • When do we serialize metadata to JSON in the write path?

General responsibility

  • Container holds references to datasets (under different name possibly)

  • Dataset holds information about the dataset itself:

    • User-defined attributes (scientific metadata) => it is not know a-priori what that is

    • Technical attributes (some may be optional) => well known how they look like

    • Fragment information is directly inlined as part of datasets

  • Fragments

    • Information about their shape, backend plugin ID, backend-specific options (unknown to us)


Applies for datasets, containers, fragments:

  • All data is kept in appropriate structures in main memory

  • Serialize to JSON just before calling the metadata backend to store the metadata (and free JSON afterwards)

  • Backend communicates via JSON to ESDM layer

Callgraph from user perspective:

  • c = container_create("name")

  • d = dataset_create(c, "dset")

  • write(d) creates fragments and attaches the metadata to the dataset

  • dataset_commit(d) => make the dataset persistent, also write the dataset + fragment metadata

  • container_commit(c) => makes the container persistent TODO: check that the right version of data is linked to it.

  • dataset_destroy(d)

  • container_destroy(c)


Applies for datasets, containers, fragments:

  • All data is kept in appropriate structures in main memory AND fetched when the data is queried initially

  • De-serialized from JSON at the earliest convenience in the ESDM layer

  • Backend communicates via JSON to ESDM layer

Callgraph from user perspective:

  • c = container_open("name") <= here we read the metadata for "name" and generate appropriate structures

  • d = dataset_open("dset") <= here we read the metadata for "dset" and generate appropriate structures

  • read(d)

  • dataset_destroy(d)

  • container_destroy(c)

Alternative workflow with unkown dsets:

  • c = container_open("name") <= here we read the metadata for "name" and generate appropriate structures

  • it = dataset_iterator()

  • for x in it:

    • dataset_iterator_dataset(x)
  • read(d)

  • dataset_destroy(d)

  • container_destroy(c)

Grid deduplication


  • The user must be allowed to create throw-away grids for reading

  • It must be avoided that there are several fragments with the exact same shape

  • It must be avoided that there are several grids with the exact same axis and subgrid structure

  • deduplication must happen on two distinct levels: the fragment level and the grid level

  • It is not possible to reference count grids or fragments, as these objects contain a reference to their owning dataset, and thus must not survive its destruction

    • deduplication must rely on proxy objects that reference the already existing objects
  • User code can hold pointers to subgrids

    • grids cannot be replaced with proxy objects, they must be turned into proxy objects

    • grids that are turned into proxy objects must retain their subgrid structure to ensure proper destruction

  • A fragment cannot be matched with a subgrid, and a subgrid cannot be matched with another subgrid that differs in one of the axes or subgrids

    • deduplication cannot happen while the grid is still in fixed axes state

    • esdm_read_grid() and esdm_write_grid() should be defined to put a grid into fixed structure state, allowing deduplication on their first call


  • Fragments are managed in a centralized container (hashtable keyed by their extends), and referenced by the grids.

    • A count of referencing grids can be added if necessary. This is not a general ref count, the dataset still owns the fragments and destructs them in its destructor. The grid count would allow the dataset to delete a fragment when its containing grids get deleted.

    • The fragment references in the grids remain plain pointers, but the grids loose their ownership over the fragments.

    • The MPI code that handles grids must be expanded to also communicate fragment information explicitly.

    • As a side effect, this also allows the dataset to unload any fragments to prevent run-away memory consumption.

  • Grids contain a delegate pointer.

    • All grid methods must first resolve any delegate chain.

    • The delegate pointer is set when the grid is first touched with an I/O or MPI call and detected to be structurally identical to an existing grid.

    • When the delegate pointer is set, the delegate pointers of all subgrids are also set.

    • There are two possible implementations for managing grid proxy objects:

      • The proxy grid’s structure data remains valid (axes and cell matrix), the proxy’s destructor recursively destructs its subgrid proxies.

      • All existing grids are managed via a flat list of grids within the dataset, and grids that become proxies get their axis and matrix data deleted immediately.

Roadmap for Implementation

  1. Create a hash function that works on hypercubes and offset/size arrays.

  2. Centralize the storage of fragments in a hash table. This deduplicates fragments, and takes fragment ownership away from grids. This will break the MPI code.

  3. Fix the MPI interface by communicating fragment metadata separately from grid metadata.

  4. Centralize the storage of grids in the dataset. The dataset should have separate lists for complete top-level grids, incomplete top-level grids, and subgrids.

  5. Implement delegates for grids.

  6. Implement the grid matching machinery to create the delegates.

Crueger, T., M. A. Giorgetta, R. Brokopf, M. Esch, S. Fiedler, C. Hohenegger, L. Kornblueh, et al. 2018. “ICON-a, the Atmosphere Component of the ICON Earth System Model: II. Model Evaluation.” Journal of Advances in Modeling Earth Systems 10 (7): 1638–62.

Kunkel, Julian, and Eugen Betke. 2017. “An MPI-IO in-Memory Driver for Non-Volatile Pooled Memory of the Kove XPD.” In High Performance Computing, edited by Julian M. Kunkel, Rio Yokota, Michela Taufer, and John Shalf, 679–90. Cham: Springer International Publishing.

“Network Common Data Form (NetCDF).” n.d.

XIOS a library dedicated to I/O management in climate codes.” n.d.

[1] A entity of the domain model such as a car could be mapped to one or several objects.

[2] Taylor et al (2012): CMIP5 Data Reference Syntax (DRS) and Controlled Vocabularies.