Previously the basket size configuration could only be coarsely applied to all
branches but in most real world files the size of the data in each branch varies
significantly. Setting a small value causes large branches to be packed into
individual baskets and prevents effective compression. Setting a large value
improves the compression ratio of large branches but causes excessive memory
usage.

Instead we change the parameters to only specifiy the maximum and minimum
branch size and provide a target for the number of events to pack into each
basket. We then use the first event to estimate the size of the branch.

This makes the branching very sensitive to the first event size however this
seems to work okay in practice. If needed the basket sizes could be refined
after several events to give a more precise estimate.

Thomas Cluzel authored 4 years ago and Marco Clemencic committed 4 years ago

see https://its.cern.ch/jira/browse/LBCOMP-23



Co-authored-by: Marco Clemencic <marco.clemencic@cern.ch>

See https://sft.its.cern.ch/jira/browse/ROOT-10514

GaudiKernel/GaudiKernel/StreamBuffer.h:569
GaudiKernel/src/Lib/EventSelectorDataStream.cpp:81
RootCnv/src/RootNTupleCnv.cpp:540

by adding the [[fallthrough]] attribute (available since C++17)

Gitlab CI authored 6 years ago and Marco Clemencic committed 6 years ago

patch generated by standalone job

- prefer boost::string_ref instead of const std::string& in interfaces
- add transparant comparison when using std::map, so that a string_ref
  can be used to search in a map which has keys of type std::string
- move standalone functions only visible in a single compilation unit
  into an anonymous namespace so that these functions are not visible
  to the linker

- require public constructor for components
- removed XyzFactory classes
- enabled warning for XyzFactory headers
- some clean up and modernization

This reverts merge request !524

The StatusCode now consists of a value and category/domain. By default
StatusCodes are created within a "default category". But users can
define new categories and StatusCodes using typed enums. The design is
mainly borrowed from std::error_code.

Main changes:
- StatusCode::setCode has been removed
- No implicit conversions from/to bool or int anymore. This greatly
  increases type safety.
- Converted all unscoped enums in Gaudi to their own StatusCode
  category (Clients will have to adjust their code to the scoped enum
  names (e.g. ERR -> Status::ERR)
- Added operator bool() that is equivalent to isSuccess()
- Added ternary (bitwise) AND(&) and OR(|) operators to combine
  StatusCodes according to three-valued logic
- Added extensive unit tests for all StatusCode features

Impact on clients:
- Code that relies on the implicit StatusCode<->bool/int conversion has
  to be changed to explicitly construct/convert StatusCodes via either
  the constructor, operator bool() or getCode().
- Clients using Gaudi StatusCodes need to adjust to using the scoped
  enums. E.g. clients of the ConversionSvc need to change BAD_STORAGE_TYPE to
  Status::BAD_STORAGE_TYPE when checking for specific StatusCode return
  types.

- and some minor format tweaking

- remove destructors qualified with 'override = default'
- prefer inheriting constructors

... plus some formatting tweaks,

In preparation for a larger re-design of StatusCode clean up a few
places where the code unnecessarily relies on an implicit long
type-conversion. In detail:

- ObjectVector::remove: Return the index of the deleted object as
  documented in the base class. Before it always returned 1
- RegistryEntry: Change return type of some methods from long to
  StautsCode
- IODataManager: Get rid of the S_OK/S_ERROR enum alias as this relies
  on the actual type of the StatusCode enums
- RootDataConnection: Change return type of setCompression from
  long to StatusCode

See https://github.com/hhatto/autopep8/issues/346