panthema / 2012 / 1119-eSAIS-Inducing-Suffix-and-LCP-Arrays-in-External-Memory / eSAIS-DC3-LCP-0.5.4 / stxxl / TODO (Download File)
* asynchronous pipelining
  (currently being developed in branch parallel_pipelining_integration)

* if the stxxl disk files have been enlarged because more external memory
  was requested by the program, resize them afterwards to
  max(size_at_program_start, configured_size)
  https://sourceforge.net/forum/message.php?msg_id=4925158

* integrate unordered_map branch

* allocation strategies: provide a method get_num_disks()
  and don't use stxxl::config::get_instance()->disks_number() inappropriately

* implement recursion in stable_ksort and do not assume random key
  distribution, do sampling instead
  as a start, at least abort early if the expected size of a bucket is larger
  than the memory available to sort it

* debug stable_ksort in depth, there are still some crashing cases left

* continue using the new approach for STXXL_VERBOSE:
  $(CXX) -DSTXXL_VERBOSE_FOO=STXXL_VERBOSEx

* check+fix all sorted_runs() calls to not cause write I/Os

* on disk destruction, check whether all blocks had been deallocated before,
  i.e. free_bytes == disk_size

* implement an allocator which looks at the available free space on disks
  when distributing blocks to disks, or does load-balancing depending
  on the given speed of the disks

* implement new disk queuing strategy that supports NCQ,
  which is now widely available in HDDs/SSDs;
  probably most interesting for rather small block sizes
  (currently begin developed in branch kernelaio)

* abstract away block manager so every container can attach to a file.

* retry incomplete I/Os for all file types (currently only syscall)

* iostats: add support for splitting I/Os and truncating reads at EOF

* do not specify marginal properties such as allocation strategy as part of the template type.
  instead, make such properties dynamically configurable using run-time polymorphism,
  which would incur only a negligible running time overhead (one virtual function call per block).

* If we get rid of the sentinels (at least for sorting), we can drop the
  min_value()/max_value() requirement on the comparator and therefore
  unmodified comparators could be used for stxxl.

* Traditionally stxxl only supports PODs in external containers, e.g. nothing
  that has non-trivial constructors, destructors or copy/assignemnt operators.
  That is neccessary to allow fast operations by accessing the underlying
  blocks directly without caring about copying/moving/otherwise manipulating
  the content or uninitialized elements.

  For some applications it could be helpful to have containers that take
  non-POD elements, but you would probably lose the features you gain by having
  direct access to the blocks.

  Some discussion regarding using some shared_ptr<> in a stxxl::vector can be
  found here: https://sourceforge.net/projects/stxxl/forums/forum/446474/topic/4099030

* The following code is not correct:

  file * f = create_file(..., RDONLY);
  vector<> v(f);
  v[42]; // non-const access, sets dirty flag

  but the error message it produces is very unclear and may come very
  asynchronous:

  terminate called after throwing an instance of 'stxxl::io_error' what():
  Error in function virtual void stxxl::mmap_file::serve(const
  stxxl::request*): Info: Mapping failed. Page size: 4096 offset modulo page
  size 0 Permission denied

  Caused by stxxl::vector<> when swapping out a dirty page. Reproducible with
  containers/test_vector_sizes.stxxl.bin

  Possible solution: vector::write_page() should check if
  vector-is-bound-to-a-file and file-is-opened-read-only throw "please use only
  const access to a read only vector"

* I've noticed that the destructor for stxxl::vector flushes any dirty data to
  disk. Possibly that makes sense for a file-backed vector (I haven't looked at
  how those work), but for a vector that uses the disks configured in .stxxl it
  is wasted I/O as the data is immediately freed. For the moment I am working
  around this by calling thevector.clear() immediately before destroying it,
  which seems to prevent the writeback (based on querying the stats).

* There should be a function like

   bool supports_filetype(const char *)

   that allows to check at runtime whether a file type (given as string) is
   available. Useful for tests that take a file type as parameter - they
   currently throw an exception "unsupported filetype".

* Change the constructor
  stxxl::vector<>::vector(stxxl::file*)
  to
  stxxl::vector<>::vector(shared_ptr<stxxl::file>)
  so that ownership of the file can be transferred to the vector and cleanup
  (closing etc.) of the file can happen automatically

* materialize() that takes an empty vector as argument and grows it, including
  allocating more blocks of fills the existing part of a vectorand once we came
  to it's end automatically expands the vector (including proper resizes). This
  approach should be more efficient (due to overlapping) than repeated
  push_back()

* streamop passthrough() useful if one wants to replace some streamop by a
  no-op by just redefining one type needs peformance measurements, hopefully
  has no impact on speed.

* stream::discard() should not need to call in.operator*() - are there side
  effects to be expected?
  -> add a streamop "inspect" or so for this purpose
  -> drop *in call from discard

* add discard(StreamOp, unsigned_irgendwas n): discard at most n elements
RSS 2.0 Weblog Feed Atom 1.0 Weblog Feed Valid XHTML 1.1 Valid CSS (2.1)
Copyright 2005-2014 Timo Bingmann - Impressum