practical_guide.md 63.3 KB
Newer Older
# Practical guide for processing large data sets with startR
This guide includes explanations and practical examples for you to learn how to use startR to efficiently process large data sets in parallel on the BSC's HPCs (Nord3-v2, CTE-Power 9, Marenostrum 4, ...). See the main page of the [**startR**](README.md) project for a general overview of the features of startR, without actual guidance on how to use it.
If you would like to start using startR rightaway on the BSC infrastructure, you can directly go through the "Configuring startR" section, copy/paste the basic startR script example shown at the end of the "Introduction" section onto the text editor of your preference, adjust the paths and user names specified in the `Compute()` call, and run the code in an R session after loading the R and ecFlow modules.
Nicolau Manubens's avatar
Nicolau Manubens committed
1. [**Motivation**](inst/doc/practical_guide.md#motivation)
2. [**Introduction**](inst/doc/practical_guide.md#introduction)
3. [**Configuring startR**](inst/doc/practical_guide.md#configuring-startr)
4. [**Using startR**](inst/doc/practical_guide.md#using-startr)
   1. [**Start()**](inst/doc/practical_guide.md#start)
   2. [**Step() and AddStep()**](inst/doc/practical_guide.md#step-and-addstep)
   3. [**Compute()**](inst/doc/practical_guide.md#compute)
      1. [**Compute() locally**](inst/doc/practical_guide.md#compute-locally)
      2. [**Compute() on CTE-Power 9**](inst/doc/practical_guide.md#compute-on-cte-power-9)
      3. [**Compute() on the fat nodes and other HPCs**](inst/doc/practical_guide.md#compute-on-the-fat-nodes-and-other-hpcs)
   4. [**Collect() and the EC-Flow GUI**](inst/doc/practical_guide.md#collect-and-the-ec-flow-gui)
5. [**Additional information**](inst/doc/practical_guide.md#additional-information)
6. [**Other examples**](inst/doc/practical_guide.md#other-examples)
7. [**Compute() cluster templates**](inst/doc/practical_guide.md#compute-cluster-templates)
## Motivation
Nicolau Manubens's avatar
Nicolau Manubens committed
What would you do if you had to apply a custom statistical analysis procedure to a 10TB climate data set? Probably, you would need to use a scripting language to write a procedure which is able to retrieve a subset of data from the file system (it would rarely be possible to handle all of it at once on a single node), code the custom analysis procedure in that language, and apply it carefully and efficiently to the data. Afterwards, you would need to think of and develop a mechanism to dispatch the job mutiple times in parallel to an HPC of your choice, each of the jobs processing a different subset of the data set. You could do this by hand but, ideally, you would rather use EC-Flow or a similar general purpose workflow manager which would orchestrate the work for you. Also, it would allow you to visually monitor and control the progress, as well as keep an easy-to-understand record of what you did, in case you need to re-use it in the future. The mentioned solution, although it is the recommended way to go, is a demanding one and you could easily spend a few days until you get it running smoothly. Additionally, when developing the job script, you would be exposed to the difficulties of efficiently managing the data and applying the coded procedure to it.
With the constant increase of resolution (in all possible dimensions) of weather and climate model output, and with the growing need for using computationally demanding analytical methodologies (e.g. bootstraping with thousands of repetitions), this kind of divide-and-conquer approach becomes indispensable. While tools exist to simplify and automate this complex procedure, they usually require adapting your data to specific formats, migrating to specific database systems, or an advanced knowledge of computer sciences or of specific programming languages or frameworks.

startR is yet another tool which allows the R user to apply user-defined functions or procedures to collections of NetCDF files (see Note 1) as large as desired, transparently using computational resources in HPCs (see Note 2) to minimize the time to solution. Although it has been designed to require as few mandatory technical parameters as possible from the user, an experienced user can configure a number of additional parameters to adjust the execution. startR operates on, and provides as outputs, multidimensional arrays with named dimensions, a basic and widely used data structure in R, and this makes the framework more familiar to the general R user.

Although startR can be difficult to use if learnt from the documentation of its functions, it can also be used with little effort if re-using and adapting already existing startR scripts, such as the ones provided later in this guide. startR scripts are written in R, and are short (usually under 50 lines of code), concise, and easy to read.

Other things you can expect from startR:
- Combining data from multiple model executions or observational data sources.
- Extracting values for a certain time period, geographical location, etc., from a collection of NetCDF files.
- Obtaining data arrays with results of analytical procedures that are to be plotted or stored as RData or NetCDF for later use in the analysis workflow.
Things that are not supposed to be done with startR:
- Curating/homogenizing model output files or generating files to be stored under /esarchive following the department/community conventions. Although metadata is understood and used by startR, its handling is not 100% consistent yet.
Nicolau Manubens's avatar
Nicolau Manubens committed
_**Note 1**_: The data files do not need to be migrated to a database system, nor have to comply with any specific convention for their file names, name, number or order of the dimensions and variables, or distribution of files in the file system. Although only files in the NetCDF format are supported for now, plug-in functions (of approximately 200 lines of code) can be programmed to add support for additional file formats.

_**Note 2**_: The HPCs startR is designed to run on are understood as multi-core multi-node clusters. startR relies on a shared file system across all HPC nodes, and does not implement any kind of distributed storage system for now.

## Introduction

In order to use startR you will need to follow the configuration steps listed in the "Configuring startR" section of this guide to make sure startR works on your workstation with the HPC of your choice.
Afterwards, you will need to understand and use five functions, all of them included in the startR package:
 - **Start()**, for declaing the data sets to be processed
 - **Step()** and **AddStep()**, for specifying the operation to be applied to the data
 - **Compute()**, for specifying the HPC to be employed, the execution parameters (e.g. number of chunks and cores), and to trigger the computation
 - **Collect()** and the **EC-Flow graphical user interface**, for monitoring of the progress and collection of results
Next, you can see an example startR script performing the ensemble mean of a small data set on an HPC cluster such as Nord3-v2 or CTE-Power9, for you to get a broad picture of how the startR functions interact and the information that is represented in a startR script. Note that the `queue_host`, `temp_dir` and `ecflow_suite_dir` parameters in the `Compute()` call are user-specific.

```r
library(startR)

repos <- '/esarchive/exp/ecmwf/system5_m1/6hourly/$var$/$var$_$sdate$.nc'
data <- Start(dat = repos,
              var = 'tas',
              sdate = '20180101',
              ensemble = 'all',
              time = 'all',
              latitude = indices(1:40),
              longitude = indices(1:40))

fun <- function(x) {
  # Expected inputs:
  #   x: array with dimensions ('ensemble', 'time')
  apply(x + 1, 2, mean)
}

step <- Step(fun, 
             target_dims = c('ensemble', 'time'), 
             output_dims = c('time'))

wf <- AddStep(data, step)

res <- Compute(wf,
               chunks = list(latitude = 2,
                             longitude = 2),
               threads_load = 2,
               threads_compute = 4,
               cluster = list(queue_host = 'p9login1.bsc.es',
                              queue_type = 'slurm',
                              temp_dir = '/gpfs/scratch/bsc32/bsc32473/startR_hpc/',
                              r_module = 'R/3.5.0-foss-2018b',
                              job_wallclock = '00:10:00',
                              cores_per_job = 4,
                              max_jobs = 4,
                              bidirectional = FALSE,
                              polling_period = 10
                             ),
               ecflow_suite_dir = '/home/Earth/nmanuben/startR_local/')
```

## Configuring startR

At BSC, the only configuration step you need to follow is to set up passwordless connection with the HPC. You do not need to follow the complete list of deployment steps under [**Deployment**](inst/doc/deployment.md) since all dependencies are already installed for you to use.
Specifically, you need to set up passwordless, userless access from your machine to the HPC login node, and from the HPC login node to your machine if at all possible. In order to establish the connection in one of the directions, you need to do the following:
1. generate an ssh pair of keys on the origin host if you do not have one, using `ssh-keygen -t rsa`
2. ssh to the destionation node and create a directory where to store the public key, using `ssh username@hostname_or_ip mkdir -p .ssh`. 'hostname_or_ip' refers to the host name or IP address of the login node of the selected HPC, and 'username' to your account name on the HPC, which may not coincide with the one in your workstation.
3. dump your public key on a new file under that folder, using `cat .ssh/id_rsa.pub | ssh username@hostname_or_ip 'cat >> .ssh/authorized_keys'`
4. adjust the permissions of the repository of keys, using `ssh username@hostname_or_ip "chmod 700 .ssh; chmod 640 .ssh/authorized_keys"`
5. if your username is different on your workstation and on the login node of the HPC, add an entry in the file .ssh/config in your workstation as follows:
```
  Host short_name_of_the_host
    HostName hostname_or_ip
    User username
    IdentityFile ~/.ssh/id_rsa
```

After following these steps for the connections in both directions (although from the HPC to the workstation might not be possible), you are good to go.
Do not forget adding the following lines in your .bashrc on the HPC machine.

If you are planning to run it on CTE-Power:
```
if [[ $BSC_MACHINE == "power" ]] ; then
  module unuse /apps/modules/modulefiles/applications
  module use /gpfs/projects/bsc32/software/rhel/7.4/ppc64le/POWER9/modules/all/
fi
```
If you are on Nord3-v2, then you'll have to add:
```
if [ $BSC_MACHINE == "nord3v2" ]; then
  module purge
  module use /gpfs/projects/bsc32/software/suselinux/11/modules/all
  module unuse /apps/modules/modulefiles/applications /apps/modules/modulefiles/compilers /apps/modules/modulefiles/tools /apps/modules/modulefiles/libraries /apps/modules/modulefiles/environment
fi
```
You can add the following lines in your .bashrc file on your workstation for convenience:
alias ctp='ssh -XY username@hostname_or_ip'
alias start='module load R CDO ecFlow'
```

## Using startR

If you have successfully gone through the configuration steps, you will just need to run the following commands in a terminal session and a fresh R session will pop up with the startR environment ready to use.

The library can be loaded as follows:
library(startR)
```

### Start()

In order for startR to recognize the data sets you want to process, you first need to declare them. The first step in the declaration of a data set is to build a special path string that encodes where all the involved NetCDF files to be processed are stored. It contains some wildcards in those parts of the path that vary across files. This special path string is also called "path pattern".
Before defining an example path pattern, let's introduce some target NetCDF files. In the "esarchive" at BSC, we can find the following files:

```
/esarchive/exp/ecmwf/system5_m1/6hourly/
  |--tas/
  |   |--tas_19930101.nc
  |   |--tas_19930201.nc
  |   |        ...
  |   |--tas_20171201.nc
  |--tos/
      |--tos_19930101.nc
      |--tos_19930201.nc
      |        ...
      |--tos_20171201.nc
```

A path pattern that can be used to encode the location of these files in a compact way is the following:

```r
repos <- '/esarchive/exp/ecmwf/system5_m1/6hourly/$var$/$var$_$sdate$.nc'
```

The wildcards used (the pieces wrapped between '$' symbols) can be given any names you like. They do not necessarily need to be 'var' or 'sdate' or match any specific key word (although in this case, as explained later, the 'var' name will trigger a special feature of `Start()`).
Once the path pattern is specified, a `Start()` call can be built, in which you need to provide, as parameters, the specific values of interest for each of the dimensions defined by the wildcards (also called outer dimensions, or file dimensions), as well as for each of the dimensions inside the NetCDF files (inner dimensions).
You can check in advance which dimensions are inside the NetCDF files by using e.g. easyNCDF on one of the files:
```r
easyNCDF::NcReadDims('/esarchive/exp/ecmwf/system5_m1/6hourly/tas/tas_19930101.nc',
                     var_names = 'tas')

This will show the names and lengths of the dimensions of the selected variable:

```r
      var longitude  latitude  ensemble      time 
        1      1296       640        25       860
_**Note**_: If you check the dimensions of that file with `ncdump -h`, you will realize the 'var' dimension is actually not defined inside. `NcReadDims()` and `Start()`, though, perceive the different variables inside a file as if stored along a virtual dimension called 'var'. You can ignore this for now and assume 'var' is simply a file dimension (since it appears as a wildcard in the path pattern). Read more on this in Note 1 at the end of this section.
Once we know the dimension names, we have all the information we need to put the `Start()` call together:

```r
data <- Start(dat = repos,
              # outer dimensions
              var = 'tas',
              sdate = '19930101',
              # inner dimensions
              ensemble = 'all',
              time = 'all',
              latitude = 'all',
              longitude = 'all')
```

For each of the dimensions, the values or indices of interest (a.k.a. selectors) can be specified in three possible ways:
- Using one or more numeric indices, for example `time = indices(c(1, 3, 5))`, or `sdate = indices(3:5)`. In the latter case, the third, fourht and fifth start dates appearing in the file system in alphabetical order would be selected ('19930301', '19930401' and '19930501').
- Using one or more actual values, for example `sdate = values('19930101')`, or `ensemble = values(c('r1i1p1', 'r2i1p1'))`, or `latitude = values(c(10, 10.5, 11))`. The `values()` helper function can be omitted (as shown in the example). See Note 2 for details on how value selectors are handled when specified for inner dimensions.
- Using a list of two numeric values, for example `sdate = indices(list(5, 10))`. This will take all indices from the 5th to the 10th.
- Using a list of two actual values, for example `sdate = values(list('r1i1p1', 'r5i1p1'))` or `latitude = values(list(-45, 75))`. This will take all values, in order, placed between the two values specified (both ends included).
- Using the special keywords 'all', 'first' or 'last'.

The dimensions specified in the `Start()` call do not need to follow any specific order, not even the actual order in the path pattern or inside the file. The order, though, can have an impact on the performance of `Start()` as explained in Note 3.

Running the `Start()` call shown above will display some progress and information messages:

```r
* Exploring files... This will take a variable amount of time depending
*   on the issued request and the performance of the file server...
* Detected dimension sizes:
*         dat:    1
*         var:    1
*       sdate:    1
*    ensemble:   25
*        time:  860
*    latitude:  640
*   longitude: 1296
* Total size of involved data:
*   1 x 1 x 1 x 25 x 860 x 640 x 1296 x 8 bytes = 132.9 Gb
* Successfully discovered data dimensions.
Warning messages:
1: ! Warning: Parameter 'pattern_dims' not specified. Taking the first dimension,
!   'dat' as 'pattern_dims'. 
2: ! Warning: Could not find any pattern dim with explicit data set descriptions (in
!   the form of list of lists). Taking the first pattern dim, 'dat', as
!   dimension with pattern specifications.
```

The warnings shown are normal, and could be avoided with a more wordy specification of the parameters to the `Start()` function.
The dimensions of the selected data set and the total size are shown. As you have probably noticed, this `Start()` call is very fast, even though several GB of data are involved (see Note 4 on the size of the data in R). This is because `Start()` is simply discovering the location and dimension of the involved data.
You can give a quick look to the collected metadata as follows:
```r
Class 'startR_header' length 9 Start(dat = "/esarchive/exp/ecmwf/system5_m1/6hourly/$var$/$var$_$sdate$.nc",      var = "tas", sdate = "19930101", ensemble = "all", time = "all", latitude = "all",  ...
  ..- attr(*, "Dimensions")= Named num [1:7] 1 1 1 25 860 ...
  .. ..- attr(*, "names")= chr [1:7] "dat" "var" "sdate" "ensemble" ...
  ..- attr(*, "Variables")=List of 2
  .. ..$ common: NULL
  .. ..$ dat1  : NULL
  ..- attr(*, "ExpectedFiles")= chr [1, 1, 1] "/esarchive/exp/ecmwf/system5_m1/6hourly/tas/tas_19930101.nc"
  ..- attr(*, "FileSelectors")=List of 1
  .. ..$ dat1:List of 3
  .. .. ..$ dat  :List of 1
  .. .. .. ..$ : chr "dat1"
  .. .. ..$ var  :List of 1
  .. .. .. ..$ : chr "tas"
  .. .. ..$ sdate:List of 1
  .. .. .. ..$ : chr "19930101"
  ..- attr(*, "PatternDim")= chr "dat"
```

The retrieved information can be accessed with the `attr()` function. For example:

```r
attr(data, 'FileSelectors')$dat1
```
```r
$dat
$dat[[1]]
[1] "dat1"


$var
$var[[1]]
[1] "tas"


$sdate
$sdate[[1]]
[1] "19930101"
```
If you are interested in actually loading the entire data set in your machine you can do so in two ways (_**be careful**_, loading the data involved in the `Start()` call in this example will most likely stall your machine. Try it with a smaller region or a subset of forecast time steps):
- adding the parameter `retrieve = TRUE` in your `Start()` call.
- evaluating the object returned by `Start()`: `data_load <- eval(data)`
Nicolau Manubens's avatar
Nicolau Manubens committed
See the section on "How to choose the number of chunks, jobs and cores" for indications on working out the maximum amount of data that can be retrieved with a `Start()` call on a specific machine. 
aho's avatar
aho committed
You may realize that this functionality is similar to the `Load()` function in the s2dv package. In fact, `Start()` is more advanced and flexible, although `Load()` is more mature and consistent for loading typical seasonal to decadal forecasting data. `Load()` will be adapted in the future to use `Start()` internally.
There are no constrains for the number or names of the outer or inner dimensions used in a `Start()` call. In other words, `Start()` will handle NetCDF files with any number of dimensions with any name, as well as files distributed across folders in complex ways, since you can use customized wildcards in the path pattern.

There are a number of advanced parameters and features in `Start()` to handle heterogeneities across files involved in a `Start()` call, such as the `synonims` parameter, or to handle dimensions extending across multiple NetCDF files, such as the `*_across` parameter. See the documentation on the [**Start()**](inst/doc/start.md) function or `?Start` for more information.
_**Note 1**_ _on the 'var' dimension_: as mentioned above in this section, `NcVarReader()` is showing as if a virtual dimension 'var' appeared inside the file. The existence of this dimension is justified by the fact that, many times, NetCDF files contain more than one variable. The 'var' dimension should hence be considered a "inner" dimension. But, in our example, the dimension 'var' is also defined as a file dimension in the path pattern. So, following the logic of `Start()`, there would be two 'var' dimensions, one of them outer and the other inner, and we should consequently specify selectors for each of them. However, as exception, they are automatically understood to be the same dimension, and the target variable name specified as selector for the outer 'var' dimension is also re-used to select the variable inside the file. This is a feature triggered only by the 'var' dimension name and, if other dimension names appeared more than once as inner or outer dimensions, `Start()` would crash throw an error. The feature described here is useful for the very common case where file paths contain the variable name and that variable is the only climate variable inside the file. If this feature was not available, one could still define the data set as shown in the code snippet below, where there would be some redundancy in the `Start()` call and in the dimensions of the resulting array.
_**Note 2**_ _on providing values as selectors for inner dimensions_: when values are requested for a inner dimension, the corresponding numeric indices are automatically calculated by comparing the provided values with a variable inside the file with the same name as the dimension for which the values have been requested. In the last example, where specific values are requested for the latitudes, the variable 'latitude' is automatically retrieved from the files. If the name of the variable does not coincide with the name of the dimension, the parameter `*_var` can be specified in the `Start()` call, as detailed in `?Start`.
_**Note 3**_ _on the order of dimensions_: neither the file dimensions nor the inner dimensions need to be specified in the same order as they appear in the path pattern or inside the NetCDF files, respectively. The resulting arrays returned by `Start()` will have the dimensions in the same order as requested in `Start()`, so changing the order in the call can potentially trigger automatic reordering of data, which is time consuming. But, depending on the use case, it may be a good idea to ask for a specific dimension order so that the data is properly arranged for making posterior calculations more efficient. Remember that the order of the dimensions in R is "big endian"; the values are consecutive along the first (left-most) dimension. In contrast, the order of dimensions in NetCDF files is "little endian". This means that if you want to respect the order of the data values in memory as stored in the NetCDF files, you should request the dimensions in your `Start()` call in the opposite order.
_**Note 4**_ _on the size of the data in R_: if you check the size of the involved file in the example `Start()` call used above ('/esarchive/exp/ecmwf/system5_m1/6hourly/tas/tas_19930101.nc'), you will realize it only weighs 34GB. Why is the data reported to occupy 134GB then? This is due to two facts: by one side, NetCDF files are usually compressed, and their uncompressed size can be substantially greater. In this case, the uncompressed data would occupy about 72GB. Besides, the variable we are targetting in the example ('tas') is defined as a float variable inside the NetCDF file. This means each value is a 4-byte real number. However, R cannot represent 4-byte real numbers; it always takes 8 bytes to represent a real number. This is why, when float numbers are represented in R, the effective size of the data is doubled. In this case, the 72GB of uncompressed float numbers need to be represented using 132GB of RAM in R.

### Step() and AddStep()

Once the data sources are declared, you can define the operation to be applied to them. The operation needs to be encapsulated in the form of an R function receiving one or more multidimensional arrays with named dimensions (plus additional helper parameters) and returning one or more multidimensional arrays, which should also have named dimensions. For example:

```r
fun <- function(x) {
  r <- sqrt(sum(x ^ 2) / length(x))
  for (i in 1:100) {
    r <- sqrt(sum(x ^ 2) / length(x))
  }
  dim(r) <- c(time = 1)
  r
}
```

As you see, the function only needs to operate on the essential dimensions, not on the whole set of dimensions of the data set. This example function receives only one multidimensional array (with dimensions 'ensemble' and 'time', although not expressed in the code), and returns one multidimensional array (with a single dimension 'time' of length 1). startR will automatically feed the function with subsets of data with only the essential dinmensions, but first, a startR "step" for the function has to be built with with the `Step()` function.
The `Step()` function requires, as parameters, the names of the dimensions of the input arrays expected by the function (in this example, a single array with the dimensions 'ensemble' and 'time'), as well as the names of the dimensions of the arrays the function returns (in this example, a single array with the dimension 'time'):

```r
step <- Step(fun = fun, 
             target_dims = c('ensemble', 'time'), 
             output_dims = c('time'))
```

The step function should ideally expect arrays with the dimensions in the same order as requested in the `Start()` call, and consequently the dimension names specified in the `Step()` function should appear in the same order. If a different order was specified, startR would reorder the subsets for the step function to receive them in the expected dimension order.

Functions that receive or return multiple multidimensional arrays are also supported. In such cases, lists of vectors of dimension names should be provided as `target_dims` or `output_dims`.

Since functions wrapped with the `Step()` function will potentially be called thousands of times, it is recommended to keep them as light as possible by, for example, avoiding calls to the `library()` function to load other packages or interacting with files on disk. See the documentation on the parameter `use_libraries` of the `Step()` function, or consider adding additional parameters to the step function with extra information. 

Once the step is built, a workflow of steps can be assembled as follows:

```r
wf <- AddStep(data, step)
```

If the step involved more than one data source, a list of data sources could be provided as first parameter. You can find examples using more than one data source further in this guide.
It is not possible for now to define workflows with more than one step, but this is not a crucial gap since a step function can contain more than one statistical analysis procedure. Furthermore, it is usually enough to perform only the first or two first steps of the analysis workflow on the HPCs because, after these steps, the volume of data involved is reduced substantially and the analysis can go on with conventional methods.
### Compute()
Once the data sources are declared and the workflow is defined, you can proceed to specify the execution parameters (including which platform to run on) and trigger the execution with the `Compute()` function.

Next, a few examples are shown with `Compute()` calls to trigger the processing of a dataset locally (only on the machine where the R session is running) and different HPCs (Nord3-v2, CTE-Power9 and other HPCs). However, let's first define a `Start()` call that involves a smaller subset of data in order not to make the examples too heavy.

```r
library(startR)

repos <- '/esarchive/exp/ecmwf/system5_m1/6hourly/$var$/$var$_$sdate$.nc'
data <- Start(dat = repos,
              var = 'tas',
              sdate = '19930101',
              ensemble = 'all',
              time = indices(1:2),
              latitude = 'all',
              longitude = 'all')

fun <- function(x) {
  # Expected inputs:
  #   x: array with dimensions ('ensemble', 'time')
  apply(x + 1, 2, mean)
}

step <- Step(fun,
             target_dims = c('ensemble', 'time'),
             output_dims = c('time'))
wf <- AddStep(data, step)
```

The output of the `Start()` call follows, where the size of the selected data is reported.
```r
* Exploring files... This will take a variable amount of time depending
*   on the issued request and the performance of the file server...
* Detected dimension sizes:
*         dat:    1
*         var:    1
*       sdate:    1
*    ensemble:   25
*        time:    2
*    latitude:  640
*   longitude: 1296
* Total size of involved data:
*   1 x 1 x 1 x 25 x 2 x 640 x 1296 x 8 bytes = 316.4 Mb
* Successfully discovered data dimensions.
Warning messages:
1: ! Warning: Parameter 'pattern_dims' not specified. Taking the first dimension,
!   'dat' as 'pattern_dims'. 
2: ! Warning: Could not find any pattern dim with explicit data set descriptions (in
!   the form of list of lists). Taking the first pattern dim, 'dat', as
!   dimension with pattern specifications.
```
#### Compute() locally
When only your own workstation is available, startR can still be useful to process a very large dataset by chunks, thus avoiding a RAM memory overload and consequent crash of the workstation. startR will simply load and process the dataset by chunks, one after the other. The operations defined in the workflow will be applied to each chunk, and the results will be stored on a temporary file. `Compute()` will finally gather and merge the results of each chunk and return a single data object, including one or multiple multidimensional data arrays, and additional metadata.

A list of the dimensions which to split the data along, and the number of slices (or "chunks") to make for each, is the only piece of information required for `Compute()` to run locally. It will only be possible to request chunks for those dimensions not required by any of the steps in the workflow built by `Step()` and `AddStep()`.

Following the worklfow of steps defined in the example, where the step uses 'time' and 'ensemble' as target dimensions, the dimensions remaining for chunking would be 'dat', 'var', 'sdate', 'latitude' and 'longitude'. Note that defining a step which has many target dimensions should be avoided as it will reduce the chunking options.

As an example, we could request for computation performing two chunks along the 'latitude' dimension, and two chunks along the 'longitude' dimension. This would result in the data being processed in 4 chunks of about 80MB (the size of the involved data, 316MB, divided by 4).

Calculate the size of the chunks before executing the computation, and make sure they fit in the RAM of your machine. You should have as much free RAM as 2x or 3x the expected size of one chunk. Read more on adjusting the number of chunks in the section "How to choose the number of chunks, jobs and cores". 

```r
res <- Compute(wf,
               chunks = list(latitude = 2,
                             longitude = 2))
```

Once the `Compute()` call is executed, the R session will wait for it to return the result. Progress messages will be shown, with a remaining time estimate after the first chunk has been processed.
```r
* Processing chunks... remaining time estimate soon...
* Loading chunk 1 out of 4 ...
* Processing...
* Remaining time estimate (at 2019-01-27 23:31:07) (neglecting merge
*   time): 1.662909 mins
* Loading chunk 2 out of 4 ...
* Processing...
* Loading chunk 3 out of 4 ...
* Processing...
* Loading chunk 4 out of 4 ...
* Processing...
```

After all chunks have been processed, a summary of the execution and its performance is reported.
```r
* Computation ended successfully.
*   Number of chunks: 4
*   Max. number of concurrent chunks (jobs): 1
*   Requested cores per job: NA
*   Load threads per chunk: 1
*   Compute threads per chunk: 1
*   Total time (s): 123.636801242828
*     Chunking setup: 0.000516414642333984
*     Data upload to cluster: 0
*     All chunks: 123.416877985001
*     Transfer results from cluster: 0
*     Merge: 0.219406843185425
*     Each chunk: 
*       queue: 
*         mean: 0
*         min: 0
*         max: 0
*       job setup: 
*         mean: 0
*         min: 0
*         max: 0
*       load: 
*         mean: 3.07621091604233
*         min: 2.05077886581421
*         max: 5.05771064758301
*       compute: 
*         mean: 27.7766432762146
*         min: 27.4815557003021
*         max: 28.199684381485
```

Also, some warning messages will be displayed corresponding to the execution of the `Start()` function to retrieve the data for each chunk.

`Compute()` will return a list of data arrays in your R session, one data array for each output returned by the last step in the specified workflow. In the example here, ony one array is returned.
```r
str(res)
```
```r
List of 1
 $ output1: num [1:2, 1, 1, 1, 1:640, 1:1296] 250 250 251 250 251 ...
 - attr(*, "startR_compute_profiling")=List of 14
  ..$ nchunks          : num 4
  ..$ concurrent_chunks: num 1
  ..$ cores_per_job    : logi NA
  ..$ threads_load     : num 1
  ..$ threads_compute  : num 1
  ..$ bychunks_setup   : num 0.000516
  ..$ transfer         : num 0
  ..$ queue            : num 0
  ..$ job_setup        : num 0
  ..$ load             : num [1:4] 5.06 3.14 2.06 2.05
  ..$ compute          : num [1:4] 28.2 27.8 27.5 27.7
  ..$ transfer_back    : num 0
  ..$ merge            : num 0.219
  ..$ total            : num 124
```

The configuration details and profiling information are attached as attributes to the returned list of arrays.

If you check the dimensions of one of the ouput arrays, you will see that it preserves named dimensions, and that the output dimensions of the workflow steps appear on the first prositions (left-most).

```r
dim(res$output1)
```
```r
     time       dat       var     sdate  latitude longitude 
        2         1         1         1       640      1296
```

In addition to performing the computation in chunks, you can adjust the number of execution threads to use for the data retrieval stage (with `threads_load`) and for the computation (with `threads_compute`). Using more than 2 threads for the retrieval will usually be perjudicial, since two will already be able to make full use of the bandwidth between the workstation and the data repository. The optimal number of threads for the computation will depend on the number of processors in your machine, the number of cores they have, and the number of threads supported by each of them.

```r
res <- Compute(wf,
               chunks = list(latitude = 2,
                             longitude = 2),
               threads_load = 2,
               threads_compute = 4)
```
```r
* Computation ended successfully.
*   Number of chunks: 4
*   Max. number of concurrent chunks (jobs): 1
*   Requested cores per job: NA
*   Load threads per chunk: 2
*   Compute threads per chunk: 4
*   Total time (s): 44.6467976570129
*     Chunking setup: 0.000483036041259766
*     Data upload to cluster: 0
*     All chunks: 44.4387269020081
*     Transfer results from cluster: 0
*     Merge: 0.207587718963623
*     Each chunk: 
*       queue: 
*         mean: 0
*         min: 0
*         max: 0
*       job setup: 
*         mean: 0
*         min: 0
*         max: 0
*       load: 
*         mean: 3.08622789382935
*         min: 2.77512788772583
*         max: 3.93441939353943
*       compute: 
*         mean: 8.02220791578293
*         min: 8.01178908348083
*         max: 8.03660178184509
In order to run the computation on a HPC, you will need to make sure the passwordless connection with the login node of that HPC is configured, as shown at the beginning of this guide. If possible, in both directions. Also, you will need to know whether there is a shared file system between your workstation and that HPC, and will need information on the number of nodes, cores per node, threads per core, RAM memory per node, and type of workload used by that HPC (Slurm, PBS and LSF supported).
You will need to add two parameters to your `Compute()` call: `cluster` and `ecflow_suite_dir`.
The parameter `ecflow_suite_dir` expects a path to a folder in the workstation where to store temporary files generated for the automatic management of the workflow. As you will see later, the EC-Flow workflow manager is used transparently for this purpose.
The parameter `cluster` expects a list with a number of components that will have to be provided a bit differently depending on the HPC you want to run on. You can see next an example cluster configuration that will execute the previously defined workflow on Nord3-v2.
  # user-defined
  temp_dir <- '/gpfs/scratch/bsc32/bsc32734/startR_hpc/'
  ecflow_suite_dir <- '/home/Earth/aho/startR_local/'

  res <- Compute(wf,
                 chunks = list(latitude = 2,
                               longitude = 2),
                 threads_load = 2,
                 threads_compute = 4,
                 cluster = list(queue_host = 'nord4.bsc.es',
                   queue_type = 'slurm',
                   temp_dir = temp_dir,
                   cores_per_job = 4,
                   job_wallclock = '00:10:00',
                   max_jobs = 4,
                   extra_queue_params = list('#SBATCH --mem-per-cpu=3000'),
                   bidirectional = FALSE,
                   polling_period = 10
                 ),
                 ecflow_suite_dir = ecflow_suite_dir,
                 wait = TRUE
                 ) 
The cluster components and options are explained next:
- `queue_host`: must match the 'short_name_of_the_host' you associated to the login node of the selected HPC in your .ssh/config file.
- `queue_type`: one of 'slurm', 'pbs' or 'lsf'.
- `temp_dir`: directory on the HPC where to store temporary files. Must be accessible from the HPC login node and all HPC nodes.
- `lib_dir`: directory on the HPC where the startR R package and other required R packages are installed, accessible from all HPC nodes. These installed packages must be compatible with the R module specified in `r_module`. This parameter is optional; only required when the libraries are not installed in the R module.
- `r_module`: name of the UNIX environment module to be used for R. If not specified, 'module load R' will be used.
- `cores_per_job`: number of computing cores to be requested when submitting the job for each chunk to the HPC queue. Each node may be capable of supporting more than one computing thread.
- `job_wallclock`: amount of time to reserve the resources when submitting the job for each chunk. Must follow the specific format required by the specified `queue_type`.
- `max_jobs`: maximum number of jobs (chunks) to be queued simultaneously onto the HPC queue. Submitting too many jobs could overload the bandwidth between the HPC nodes and the storage system, or could overload the queue system.
- `extra_queue_params`: list of character strings with additional queue headers for the jobs to be submitted to the HPC. Mainly used to specify the amount of memory to book for each job (e.g. '#SBATCH --mem-per-cpu=30000'), to request special queuing (e.g. '#SBATCH --qos=bsc_es'), or to request use of specific software (e.g. '#SBATCH --reservation=test-rhel-7.5').
- `bidirectional`: whether the connection between the R workstation and the HPC login node is bidirectional (TRUE) or unidirectional from the workstation to the login node (FALSE).
- `polling_period`: when the connection is unidirectional, the workstation will ask the HPC login node for results each `polling_period` seconds. An excessively small value can overload the login node or result in temporary banning.
After the `Compute()` call is executed, an EC-Flow server is automatically started on your workstation, which will orchestrate the work and dispatch jobs onto the HPC. Thanks to the use of EC-Flow, you will also be able to monitor visually the progress of the execution. See the "Collect and the EC-Flow GUI" section.

The following messages will be displayed upon execution:
* ATTENTION: Dispatching chunks on a remote cluster. Make sure
*   passwordless access is properly set in both directions.
ping server(bscearth329:5678) succeeded in 00:00:00.001342  ~1 milliseconds
server is already started
* Processing chunks...
* Remaining time estimate soon...
```
At this point, you may want to check the jobs are being dispatched and executed properly onto the HPC. For that, you can either use the EC-Flow GUI (covered in the next section), or you can `ssh` to the login node of the HPC and check the status of the queue with `squeue` or `qstat`, as shown below.
```
             JOBID PARTITION     NAME     USER ST       TIME  NODES NODELIST(REASON)
            757026      main /STARTR_ bsc32734  R       0:46      1 s02r2b24
            757027      main /STARTR_ bsc32734  R       0:46      1 s04r1b61
            757028      main /STARTR_ bsc32734  R       0:46      1 s04r1b63
            757029      main /STARTR_ bsc32734  R       0:46      1 s04r1b64
Here the output of the execution after waiting for about a minute:
```r
* Remaining time estimate (neglecting queue and merge time) (at
*   2019-01-28 01:16:59): 0 mins (46.22883 secs per chunk)
* Computation ended successfully.
*   Number of chunks: 4
*   Max. number of concurrent chunks (jobs): 4
*   Requested cores per job: 4
*   Load threads per chunk: 2
*   Compute threads per chunk: 4
*   Total time (s): 58.3834805488586
*     Chunking setup: 1.13428020477295
*     Data upload to cluster: 0
*     All chunks: 52.9015641212463
*     Transfer results from cluster: 4.05326294898987
*     Merge: 0.294373273849487
*     Each chunk: 
*       queue: 
*         mean: 10.5
*         min: 7
*         max: 14
*       job setup: 
*         mean: 3.3917236328125
*         min: 1.16518902778625
*         max: 5.61825823783875
*       load: 
*         mean: 5.77260076999664
*         min: 5.27595114707947
*         max: 6.26925039291382
*       compute: 
*         mean: 23.7037765979767
*         min: 23.4937765598297
*         max: 23.9137766361237
* Computation ended successfully.
Warning messages:
1: ! Warning: Parameter 'ecflow_server' has not been specified but execution on
!cluster has been requested. An ecFlow server instance will
!be created on localhost:5678. 
2: ! Warning: ATTENTION: The source chunks will be removed from the
!system. Store the result after Collect() ends if needed.
```

As you have probably realized, this execution has been slower than the local execution, even if 4 simultaneous jobs have been executed on CTE-Power. This is due to the small size of the data being processed here. The overhead of queuing and starting jobs at CTE-Power is large compared to the required computation time for this amount of data. The benefit would be obvious in use cases with larger inputs.
Usually, in use cases with larger data inputs, it will be preferrable to add the parameter `wait = FALSE` to your `Compute()` call. With this parameter, `Compute()` will return an object with all the information on your startR execution which you will be able to store in your disk. After doing that, you will be able to close your R session and collect the results later on with the `Collect()` function. This is discussed in the next section.

As mentioned above in the definition of the `cluster` parameters, it is strongly recommended to check the section on "How to choose the number of chunks, jobs and cores".

You can find the `cluster` configuration for other HPCs at the end of this guide [Compute() cluster templates](#compute-cluster-templates)
### Collect() and the EC-Flow GUI

Usually, in use cases where large data inputs are involved, it is convenient to add the parameter `wait = FALSE` to your `Compute()` call. With this parameter, `Compute()` will immediately return an object with information about your startR execution. You will be able to store this object onto disk. After doing that, you will not need to worry in case your workstation turns off in the middle of the computation. You will be able to close your R session, and collect the results later on with the `Collect()` function.

```r
  res <- Compute(wf,
                 chunks = list(latitude = 2,
                               longitude = 2),
                 threads_load = 2,
                 threads_compute = 4,
                 cluster = list(queue_host = 'nord4.bsc.es',
                   queue_type = 'slurm',
                   temp_dir = '/gpfs/scratch/bsc32/bsc32734/startR_hpc/',
                   cores_per_job = 4,
                   job_wallclock = '00:10:00',
                   max_jobs = 4,
                   extra_queue_params = list('#SBATCH --mem-per-cpu=3000'),
                   bidirectional = FALSE,
                   polling_period = 10
                 ),
                 ecflow_suite_dir = '/home/Earth/aho/startR_local/',
                 wait = FALSE
                 ) 
  saveRDS(res, file = 'test_collect.Rds')
```

At this point, after storing the descriptor of the execution and before calling `Collect()`, you may want to visually check the status of the execution. You can do that with the EC-Flow graphical user interface. You need to open a new terminal, load the EC-Flow module if needed, and start the GUI:

```
module load ecFlow
ecflow_ui &
```

After doing that, a window will pop up. You will be able to monitor the status of your EC-Flow suites there. However, if it is the first time you are using the EC-Flow GUI with startR, you will need to register the EC-Flow server that has been started automatically by `Compute()`. You can open the top menu "Manage servers" > "New server" > set host to 'localhost' > set port to '5678' > save.

Note that the host and port can be adjusted with the parameter `ecflow_server` in `Compute()`, which must be provided in the form `c(host = 'hostname', port = port_number)`.

After registering the EC-Flow server, an expandable entry will appear, where you can see listed the jobs to be executed, one for each chunk, with their status represented by a colour. Gray means pending, blue means queuing, green means in progress, and yellow means completed.

You will see that, if you are running on an HPC where the connection with its login node is unidirectional, the jobs remain blue (queuing). This is because the jobs, upon start or completion, cannot send the signals back. In order to retrieve this information, the `Collect()` function must be called from an R terminal.
library(startR)

res <- readRDS('test_collect.Rds')

result <- Collect(res, wait = TRUE)

In this example, `Collect()` has been run with the parameter `wait = TRUE`. This will be a blocking call, in which `Collect()` will retrieve information from the HPC, including signals and outputs, each `polling_period` seconds (as described above). `Collect()` will not return until the results of all chunks have been received. Meanwhile, the status of the EC-Flow workflow on the EC-Flow GUI will be updated periodically and you will be able to monitor the status, as shown in the image below (image taken from another use case).

<img src="inst/doc/ecflow_monitoring.png" width="600" />

Upon completion, `Collect()` returns the merged data array, as seen in the "Compute locally" section.

```r
str(result)
```
```r
List of 1
 $ output1: num [1:2, 1, 1, 1, 1:640, 1:1296] 251 252 249 246 245 ...
 - attr(*, "startR_compute_profiling")=List of 14
  ..$ nchunks          : num 4
  ..$ concurrent_chunks: num 4
  ..$ cores_per_job    : num 4
  ..$ threads_load     : num 2
  ..$ threads_compute  : num 4
  ..$ bychunks_setup   : num 1.16
  ..$ transfer         : num 0
  ..$ queue            : Named num [1:4] 18 18 18 18
  .. ..- attr(*, "names")= chr [1:4] "queue" "queue" "queue" "queue"
  ..$ job_setup        : Named num [1:4] 2.25 2.25 2.25 2.25
  .. ..- attr(*, "names")= chr [1:4] "job_setup" "job_setup" "job_setup" "job_setup"
  ..$ load             : Named num [1:4] 5.64 6.19 5.64 5.62
  .. ..- attr(*, "names")= chr [1:4] "load" "load" "load" "load"
  ..$ compute          : Named num [1:4] 24.5 24.4 24.3 24.4
  .. ..- attr(*, "names")= chr [1:4] "compute" "compute" "compute" "compute"
  ..$ transfer_back    : num 1.72
  ..$ merge            : num 0.447
  ..$ total            : logi NA
```

Note that, when the results are collected with `Collect()` instead of calling `Compute()` with the parameter `wait = TRUE`, it will not be possible to know the total time taken by the entire data processing workflow, but we will still be able to know the timings of most of the stages.

You can also run `Collect()` with `wait = FALSE`. This will crash with an error if the results are not yet available, or will return the merged array otherwise.

`Collect()` also has a parameter called `remove`, which by default is set to `TRUE` and triggers removal of all data results received from the HPC (and stored under `ecflow_suite_dir`). If you would like to preserve the data, you can set `remove = FALSE` and `Collect()` it as many times as desired. Alternatively, you can `Collect()` with `remove = TRUE` and store the merged array right after with `saveRDS()`.
## Additional information
### How to choose the number of chunks, jobs and cores
#### Number of chunks and memory per job

Adjusting the number of chunks, simultaneous jobs, cores and memory per job is crucial for a successful and efficient execution of your startR workflow on the HPC.

The most important choice is that of the number of chunks and memory per job. Your first priority should be to make the data chunks fit in the HPC nodes' RAM memory.

After running the `Start()` call (or calls) for a use case, you will get a message on the terminal with the total size of the involved data. You should divide the data in as many chunks as required to make one of them fit in the blocks of RAM memory on the HPC. If you check the specifications of your target HPC, you will see that each node has a fixed amount of RAM (e.g., for CTE-Power 9, each node has 512GB of RAM). But that is not what you are looking for. You want to know the size of the RAM memory modules (e.g., for CTE-Power 9, the memory modules are of 32 GB). startR can not use memory regions allocated in multiple memory modules.

Knowing the size of the involved data and the size of the memory modules, you can work out the ideal number of chunks with a simple division. If the data weighs 100GB and the memory modules are of 32GB, 4 chunks will be required ideally.

However, we must take into account that the handling of these data chunks is not ideal, and that neither startR nor the functions to be applied to the data will manage the data in a 100% efficient manner. This is why the number of chunks should be finally doubled or multiplied by 3.

The amount of memory requested for each job, in this case where the data to be processed is a few times larger than the memory modules, must be equal to the size of the memory modules.

If the data to be processed was smaller than the memory modules, it could still be interesting to split it in chunks in order to parallelize the computation and get a faster result. In that case, the memory requested for each job must be equal to 2 or 3 times the size of one chunk.

#### Maximum number of simultaneous jobs

The maximum number of simultaneous jobs should be adjusted according to the capacity of the file system to deal with multiple simultanous file accesses, according to the bandwidth of the connection between the HPC nodes and the file system, and according to the number of jobs per user supported/allowed by the HPC queue. A recommended practice is to make a test with a subset of the data and try first with a small number of simultaneous jobs (e.g. 2), and keep repeating the test while increasing the number of simultaneous jobs, until the performance of the process is not improved due to any of the reasons mentioned before.

#### Number of cores and computing threads

The number of cores per job should be as large as possible, with two limitations:
- If the HPC nodes have each "M" total amount of memory with "m" amount of memory per memory module, and have each "N" amount of cores, the requested amount of cores "n" should be such that "n" / "N" does not exceed "m" / "M".
- Depending on the shape of the chunks, startR has a limited capability to exploit multiple cores. It is recommended to make small tests increasing the number of cores to work out a reasonable number of cores to be requested. 
### How to clean a failed execution
- Work out the startR execution ID, either by inspecting the execution description by `Compute()` when called with the parameter `wait = FALSE`, or by checking the `ecflow_suite_dir` with `ls -ltr`.
- ssh to the HPC login node and cancel all jobs of your startR execution.
- Close the R session from where your `Compute()` call was made.
- Remove the folder named with your startR execution ID under the `ecflow_suite_dir`.
- ssh to the HPC login node and remove the folder named with your startR execution ID under the `temp_dir`.
- Optionally remove the data under `data_dir` on the HPC login node if the file system is not shared between the workstation and the HPC and you do not want to keep the data in the `data_dir`, used as caché for future startR executions.
- Open the EC-Flow GUI and remove the workflow entry (a.k.a. suite) named with your startR execution ID with right click > "Remove".
### Visualizing the profiling of the execution
As seen in previous sections, profiling measurements of the execution are provided together with the data output. These measurements can be visualized with the `PlotProfiling()` function made available in the source code of the startR package.

This function has not been included as part of the official set of functions of the package because it requires a number of extense plotting libraries which take time to load and, since the startR package is loaded in each of the worker jobs on the HPC, this could imply a substantial amount of time spent in repeatedly loading unused visualization libraries during the computing stage.

The function takes as inputs one or a list of 'startR_compute_profiling' attribute objects from one or more `Compute()` executions.
```r
source('https://earth.bsc.es/gitlab/es/startR/raw/master/inst/PlotProfiling.R')
PlotProfiling(attr(res, 'startR_compute_profiling'))
```

A chart displays the timings for the different stages of the computation, as shown in the image below. Note that these results have been taken from a use case different to the ones used in this guide. 

<img src="inst/doc/compute_profiling.png" width="800" />

You can click on the image to expand it.
### What to do if your function has too many target dimensions
### Pending features
- Adding feature for `Copute()` to run on multiple HPCs or workstations.
- Adding plug-in to read CSV files.
- Supporting multiple steps in a workflow defined by `AddStep()`.
- Adding feature in `Start()` to read sparse grid points.
- Allow for chunking along "essential" (a.k.a. "target") dimensions.
## Other examples
### Using experimental and (date-corresponding) observational data

```r
repos <- paste0('/esnas/exp/ecmwf/system4_m1/6hourly/',
                '$var$/$var$_$sdate$.nc')

system4 <- Start(dat = repos,
                 var = 'sfcWind',
                 #sdate = paste0(1981:2015, '1101'),
                 sdate = paste0(1981:1984, '1101'),
                 #time = indices((30*4+1):(120*4)),
                 time = indices((30*4+1):(30*4+4)),
                 ensemble = 'all',
                 #ensemble = indices(1:6),
                 #latitude = 'all',
                 latitude = indices(1:10),
                 #longitude = 'all',
                 longitude = indices(1:10),
                 return_vars = list(latitude = NULL,
                                    longitude = NULL,
                                    time = c('sdate')))

repos <- paste0('/esnas/recon/ecmwf/erainterim/6hourly/',
                '$var$/$var$_$file_date$.nc')

dates <- attr(system4, 'Variables')$common$time
dates_file <- sort(unique(gsub('-', '', sapply(as.character(dates),
substr, 1, 7))))

erai <-    Start(dat = repos,
                 var = 'sfcWind',
                 file_date = dates_file,
                 time = values(dates),
                 #latitude = 'all',
                 latitude = indices(1:10),
                 #longitude = 'all',
                 longitude = indices(1:10),
                 time_var = 'time',
                 time_tolerance = as.difftime(1, units = 'hours'),
                 time_across = 'file_date',
                 return_vars = list(latitude = NULL,
                                    longitude = NULL,
                                    time = 'file_date'),
                 merge_across_dims = TRUE,
                 split_multiselected_dims = TRUE)

step <- Step(eqmcv_atomic,
             list(a = c('ensemble', 'sdate'),
                  b = c('sdate')),
             list(c = c('ensemble', 'sdate')))

res <- Compute(step, list(system4, erai),
               chunks = list(latitude = 5,
                             longitude = 5,
                             time = 2),
               cluster = list(queue_host = 'bsceslogin01.bsc.es',
                              max_jobs = 4,
                              cores_per_job = 2),
               shared_dir = '/esnas/scratch/nmanuben/test_bychunk',
               wait = FALSE) 
```

### Example of computation of weekly means
### Example with data on an irregular grid with selection of a region
### Example on MareNostrum 4
repos <- paste0('/esarchive/exp/ecmwf/system5_m1/6hourly/',
                '$var$-longitudeS1latitudeS1all/$var$_$sdate$.nc')
# Slower alternative, using files with a less efficient NetCDF 
# compression configuration
#repos <- '/esarchive/exp/ecmwf/system5_m1/6hourly/$var$/$var$_$sdate$.nc'
data <- Start(dat = repos,
              var = 'tas',
              sdate = indices(1),
              ensemble = 'all',
              time = 'all',
              latitude = indices(1:40),
              longitude = indices(1:40),
              retrieve = FALSE)
lons <- attr(data, 'Variables')$common$longitude
lats <- attr(data, 'Variables')$common$latitude

fun <- function(x) apply(x + 1, 2, mean)
step <- Step(fun, c('ensemble', 'time'), c('time'))
wf <- AddStep(data, step)

res <- Compute(wf,
               chunks = list(latitude = 2,
                             longitude = 2),
               threads_load = 1,
               threads_compute = 2,
               cluster = list(queue_host = 'mn2.bsc.es',
                              queue_type = 'slurm',
                              data_dir = '/gpfs/projects/bsc32/share/startR_data_repos/',
                              temp_dir = '/gpfs/scratch/pr1efe00/pr1efe03/startR_tests/',
                              r_module = 'R/3.4.0',
                              cores_per_job = 2,
                              job_wallclock = '00:10:00',
                              max_jobs = 4,
                              extra_queue_params = list('#SBATCH --qos=prace'),
                              bidirectional = FALSE,
                              polling_period = 10,
                              special_setup = 'marenostrum4'
                             ),
               ecflow_suite_dir = '/home/Earth/nmanuben/test_remove/',
               ecflow_server = NULL,
               silent = FALSE,
               debug = FALSE,
               wait = TRUE)
```

### Example on CTE-Power using GPUs
### Seasonal forecast verification example on cca

```r
crps <- function(x, y) {
  mean(SpecsVerification::EnsCrps(x, y, R.new = Inf))
}

library(startR)

repos <- '/perm/ms/spesiccf/c3ah/qa4seas/data/seasonal/g1x1/ecmf-system4/msmm/atmos/seas/tprate/12/ecmf-system4_msmm_atmos_seas_sfc_$date$_tprate_g1x1_init12.nc'

data <- Start(dat = repos,
              var = 'tprate',
              date = 'all',
              time = 'all',
              number = 'all',
              latitude = 'all',
              longitude = 'all',
              return_vars = list(time = 'date'))

dates <- attr(data, 'Variables')$common$time

repos <- '/perm/ms/spesiccf/c3ah/qa4seas/data/ecmf-ei_msmm_atmos_seas_sfc_19910101-20161201_t2m_g1x1_init02.nc'

obs <- Start(dat = repos,
             var = 't2m',
             time = values(dates),
             latitude = 'all',
             longitude = 'all',
             split_multiselected_dims = TRUE)

s <- Step(crps, target_dims = list(c('date', 'number'), c('date')),
                output_dims = NULL)
wf <- AddStep(list(data, obs), s)

r <- Compute(wf,
             chunks = list(latitude = 10,
                           longitude = 3),
             cluster = list(queue_host = 'cca',
                            queue_type = 'pbs',