StorageElement

DIRAC provides an abstraction of a SE interface that allows to access different kind of them with a single interface. The access to each kind of SE ( SRMv2, DIRAC SE, …) is achieved by using specific plugin modules that provide a common interface. The information necessary to define the proper plugin module and to properly configure this plugin to access a certain SE has to be introduced in the DIRAC Configuration. An example of such configuration is:

CERN-USER
{
  OccupancyPlugin = WLCGAccountingJson
  OccupancyLFN = /lhcb/spaceReport.json
  SpaceReservation = LHCb_USER
  ReadAccess = Active
  WriteAccess = Active
  AccessProtocol.1
  {
    # The name of the DIRAC Plugin module to be used for implementation
    # of the access protocol
    PluginName = SRM2
    # Flag specifying the access type (local/remote)
    Access = remote
    # Protocol name
    Protocol = srm
    # Host endpoint
    Host = srm-lhcb.cern.ch
    Port = 8443
    # WSUrl part of the SRM-type PFNs
    WSUrl = /srm/managerv2?SFN=
    # Path to navigate to the VO namespace on the storage
    Path = /castor/cern.ch/grid
    # SRM space token
    SpaceToken = LHCb_USER
    # VO specific path definitions
    VOPath
    {
      biomed = /castor/cern.ch/biomed/grid
    }
  }
}

Configuration options are:

  • BackendType: just used for information. No internal use at the moment

  • SEType: Can be T0D1 or T1D0 or T1D1. it is used to asses whether the SE is a tape SE or not. If the digit after T is 1, then it is a tape.

  • UseCatalogURL: default False. If True, use the url stored in the catalog instead of regenerating it

  • ChecksumType: default ADLER32. NOT ACTIVE !

  • Alias: when set to the name of another storage element, it instanciates the other SE instead.

  • ReadAccess: default True. Allowed for Read if no RSS enabled (Activate RSS)

  • WriteAccess: default True. Allowed for Write if no RSS enabled

  • CheckAccess: default True. Allowed for Check if no RSS enabled

  • RemoveAccess: default True. Allowed for Remove if no RSS enabled

  • OccupancyLFN: default (/<vo>/occupancy.json). LFN where the json file containing the space reporting is to be found

  • OccupancyPlugin: default (empty). Plugin to find the occupancy of a given storage.

  • SpaceReservation: just a name of a zone of the physical storage which can have some space reserved. Extends the SRM SpaceToken concept.

  • ArchiveTimeout: for tape SE only. If set to a value in seconds, enables the FTS Archive Monitoring feature

  • BringOnlineTimeout: for tape SE only. If set to a value in seconds, specify the BringOnline parameter for FTS transfers. Otherwise, the default is whatever is in the FTS3Job class.

VO specific paths

Storage Elements supporting multiple VO’s can have definitions slightly differing with respect to the Path used to navigate to the VO specific namespace in the physical storage. If a generic Path can not be suitable for all the allowed VO’s a VOPath section can be added to the Plugin definition section as shown in the example above. In this section a specific Path can be defined for each VO which needs it.

StorageElementBases

Installations tend to have several StorageElements, with very similar configurations (e.g., the same Host and Port). It could be useful to factorize the SEs configuration to avoid repeating it. In order to factorize the configuration, it is possible to use BaseSE, which acts just like inheritance in object programming. You define a SE just like any other but in the StorageElementBases section. This SE can then be refered to by another SE. This new SE will inherit all the configuration from its parents, and can override it. For example:

StorageElementBases
{
  CERN-EOS
  {
    BackendType = Eos
    SEType = T0D1
    AccessProtocol.1
    {
      Host = srm-eoslhcb.cern.ch
      Port = 8443
      PluginName = GFAL2_SRM2
      Protocol = srm
      Path = /eos/lhcb/grid/prod
      Access = remote
      SpaceToken = LHCb-EOS
      WSUrl = /srm/v2/server?SFN=
    }
  }
}
StorageElements
{
  CERN-DST-EOS
  {
    BaseSE = CERN-EOS
  }
  CERN-USER
  {
    BaseSE = CERN-EOS
    PledgedSpace = 205
    AccessProtocol.1
    {
      PluginName = GFAL2_SRM2
      Path = /eos/lhcb/grid/user
      SpaceToken = LHCb_USER
    }
  }
  GFAL2_XROOT
  {
    Host = eoslhcb.cern.ch
    Port = 8443
    Protocol = root
    Path = /eos/lhcb/grid/user
    Access = remote
    SpaceToken = LHCb-EOS
    WSUrl = /srm/v2/server?SFN=
  }
}

This definition would be strictly equivalent to:

StorageElementBases
{
  CERN-EOS
  {
    BackendType = Eos
    SEType = T0D1
    AccessProtocol.1
    {
      Host = srm-eoslhcb.cern.ch
      Port = 8443
      PluginName = GFAL2_SRM2
      Protocol = srm
      Path = /eos/lhcb/grid/prod
      Access = remote
      SpaceToken = LHCb-EOS
      WSUrl = /srm/v2/server?SFN=
    }
  }
}
StorageElements
{
  CERN-DST-EOS
  {
    BackendType = Eos
    SEType = T0D1
    AccessProtocol.1
    {
      Host = srm-eoslhcb.cern.ch
      Port = 8443
      PluginName = GFAL2_SRM2
      Protocol = srm
      Path = /eos/lhcb/grid/prod
      Access = remote
      SpaceToken = LHCb-EOS
      WSUrl = /srm/v2/server?SFN=
    }
  }
  CERN-USER
  {
    BackendType = Eos
    SEType = T0D1
    PledgedSpace = 205
    AccessProtocol.1
    {
      Host = srm-eoslhcb.cern.ch
      Port = 8443
      PluginName = GFAL2_SRM2
      Protocol = srm
      Path = /eos/lhcb/grid/user
      Access = remote
      SpaceToken = LHCb_USER
      WSUrl = /srm/v2/server?SFN=
    }
  }
  GFAL2_XROOT
  {
    Host = eoslhcb.cern.ch
    Port = 8443
    PluginName =  GFAL2_XROOT
    Protocol = root
    Path = /eos/lhcb/grid/user
    Access = remote
    SpaceToken = LHCb-EOS
    WSUrl = /srm/v2/server?SFN=
  }
}

Note that base SE must be separated from the inherited SE in two different sections. You can also notice that the name of the protocol section can be a plugin name. In this way, you do not need to specify a plugin name inside.

Available protocol plugins

DIRAC comes with a bunch of plugins that you can use to interact with StorageElements. These are the plugins that you should define in the PluginName option of your StorageElement definition.

  • DIP: used for dips, the DIRAC custom protocol (useful for example for DIRAC SEs).

  • File: offers an abstraction of the local access as an SE.

  • RFIO (deprecated): for the rfio protocol.

  • Proxy: to be used with the StorageElementProxy.

  • S3: for S3 (e.g. AWS, CEPH) support (see S3 support in DIRAC)

There are also a set of plugins based on the gfal2 libraries.

  • GFAL2_SRM2: for srm, replaces SRM2

  • GFAL2_XROOT: for xroot, replaces XROOT

  • GFAL2_HTTPS: for https

  • GFAL2_GSIFTP: for gsiftp

Default plugin options:

  • Access: Remote or Local. If Local, then this protocol can be used only if we are running at the site to which the SE is associated. Typically, if a site mounts the storage as NFS, the file protocol can be used.

  • InputProtocols/OutputProtocols: a given plugin normally contain a hard coded list of protocol it is able to generate or accept as input. There are however seldom cases (like SRM) where the site configuration may change these lists. These options are here to accomodate for that case.

GRIDFTP Optimisation

For efficiency reasons the environment variable DIRAC_GFAL_GRIDFTP_SESSION_REUSE should be exported in the server bashrc files:

export DIRAC_GFAL_GRIDFTP_SESSION_REUSE=True

This enables the session reuse for the GRIDFTP plugin. This cannot be enabled generally because it can lead to denial of service like attacks when thousands of jobs keep their connections to an SE alive for too long.

Another variable that can tune GRIDFTP plugin is the DIRAC_GFAL_GRIDFTP_ENABLE_IPV6. Because of a globus bug, pure ipv4 nodes can’t talk to dual stack nodes. A fix is under way (aug. 2019 https://its.cern.ch/jira/browse/LCGDM-2817), but in the meantime, this environement variable allows for some flexibility. Ideally, you should leave the default (i.e. True), but you can disable it for problematic sites.

Space occupancy

Several methods allow to know how much space is left on a storage, depending on the protocol:

  • dips: a simple system call returns the space left on the partition

  • srm: the srm is able to return space occupancy based on the space token

  • any other: a generic implementation has been made in order to retrieve a JSON file containing the necessary information.

A WLCG working group is trying to standardize the space reporting. So a standard will probably emerge soon (before 2053). For the time being, we shall consider that the JSON file will contain a dictionary with keys Total and Free in Bytes. For example:

{
  "Total": 20,
  "Free": 10
}

The LFN of this file is by default /<vo>/occupancy.json, but can be overwritten with the OccupancyLFN option of the SE.

The SpaceReservation option allows to specify a physical zone of the storage which would have space reservation (for example LHCb_USER, LHCb_PROD, etc). It extends the concept of SpaceToken that SRM has. This option is only used if the StoragePlugin does not return itself a SpaceReservation value.

The OccupancyPlugin allows to change the way space occupancy is measured. Several plugins are available (please refer to the module documentation):

Locally mounted filesystems

Some sites mount their storage directly on the worker nodes. In order to access these files, you can rely on the GFAL2_SRM2 plugin or on the File plugin.

With File

If the local path follows the DIRAC convention (i.e. finishes with the LFN), then you can use the File plugin. This is simply defined like that:

File
{
  Protocol = file
  Path = /mnt/lustre_3/storm_3/lhcbdisk/
  Host = localhost
  Access = local
}

With SRM2

In case there is some mangling done somewhere, and the file path does not follow the DIRAC convention, you may need to ask the local path to SRM. You need to define a protocol section with SRM, specifying that a file URL can be generated and that it is valid only in local:

GFAL2_SRM2_LOCAL
{
  PluginName = GFAL2_SRM2
  Host = storm-fe-lhcb.cr.cnaf.infn.it
  Port = 8444
  Protocol = srm
  Path = /disk
  # This is different from the ``standard`` definition
  Access = local
  SpaceToken = LHCb-Disk
  WSUrl = /srm/managerv2?SFN=
  # This is different from the ``standard`` definition
  OutputProtocols = file, https, gsiftp, root, srm
}

Multi Protocol

There are several aspects of multi protocol:

  • One SE supports several protocols

  • SEs with different protocols need to interact

  • We want to use different protocols for different operations

DIRAC supports all of them. The bottom line is that before executing an action on an SE, we check among all the plugins defined for it, which plugins are the most suitable. There are 5 Operation options under the DataManagement section used for that:

  • RegistrationProtocols: used to generate a URL that will be stored in the FileCatalog

  • AccessProtocols: used to perform the read operations

  • WriteProtocols: used to perform the write and remove operations

  • ThirdPartyProtocols: used in case of replications

  • StageProtocols: used when issuing a stage request (see below)

When performing an action on an SE, the StorageElement class will evaluate, based on these lists, and following this preference order, which StoragePlugins to use.

The behavior is straightforward for simple read or write actions. It is however a bit more tricky when it comes to third party copies.

Each StoragePlugins has a list of protocols that it is able to accept as input and a list that it is able to generate. In most of the cases, for protocol X, the plugin is able to generate URL for the protocol X, and to take as input URL for the protocol X and local files. There are plugins that can do more, like GFAL2_SRM2 plugins that can handle many more (xroot, gsiftp, etc). It may happen that the SE can be writable only by one of the protocol. Suppose the following situation: you want to replicate from storage A to storage B. Both of them have as plugins GFAL2_XROOT and GFAL2_SRM2; AccessProtocols is “root,srm”, WriteProtocols is “srm” and ThirdPartyProtocols is “root,srm”.

The negociation between the storages to find common protocol for third party copy will lead to “root,srm”. Since we follow the order, the sourceURL will be a root url, and it will be generated by GFAL2_XROOT because root is its native protocol (so we avoid asking the srm server for a root turl). The destination will only consider using GFAL2_SRM2 plugins because only srm is allowed as a write plugin, but since this plugins can take root URL as input, the copy will work.

The WriteProtocols, AccessProtocols and StageProtocols lists can be locally overwritten in the SE definition.

Multi Protocol with FTS

External services like FTS requires pair of URLs to perform third party copy. This is implemented using the same logic as described above. There is however an extra step: once the common protocols between 2 SEs have been filtered, an extra loop filter is done to make sure that the selected protocol can be used as read from the source and as write to the destination. Finally, the URLs which are returned are not necessarily the url of the common protocol, but are the native urls of the plugin that can accept/generate the common protocol. For example, if the common protocol is gsiftp but one of the SE has only an SRM plugin, then you will get an srm URL (which is compatible with gsiftp).

Added in version v7r1p37: The FTS3Agent can now use plugins to influence the list of TPC protocols used. See FTS3 support in DIRAC

MultiHop support

See FTS3 support in DIRAC

Protocol matrix

In order to make it easier to debug, the script dirac-dms-protocol-matrix will generate a CSV files that allows you to see what would happen if you were to try transfers between SEs

Staging

Up to recently, any protocol that was defined as AccessProtocols was also used for staging (plot twister: only srm could do it). This is not true for CTA . Because CTA can stage with xroot only, but we may need to use another protocol to transfer to a another site, we need to distinguish between staging and accessing. To the best of my knowledge, only CTA is like this, and thus, it is the only place where you may need to define StageProtocols. In case of FTS transfer from CTA where the stage and transfer protocols are different, we rely on the multihop mechanism of FTS to do the protocol translations. More technical details are available in DIRAC.DataManagementSystem.Client.FTS3Job

StorageElementGroups

StorageElements can be grouped together in a StorageElementGroup. This allows the systems or the users to refer to any storage within this group.

Mapping Storages to Sites and Countries

Both Sites and Countries can have StorageElement (discouraged) or StorageElementGroup associated. This shows particularly useful if we want to restrict the job output upload to specific locations, due to network constraints for example. This is done using the AssociatedSEs parameter of the Site or Country. The resolution order and logic is explained in getDestinationSEList() and well illustrated with examples in the associated tests