Debian Codegraph @

dune/uggrid/parallel/ddd/TODO

raw
1 <!--
2 SPDX-FileCopyrightInfo: Copyright © DUNE Project contributors, see file LICENSE.md in module root
3 SPDX-License-Identifier: LGPL-2.1-or-later
4 -->
5
6
7 ---------------------------------------------------------------------
8
9 DDD TODO-LIST
10
11 971008 kb created
12
13 Apart from many TODOs spread out in the source code there are
14 some more features which are still missing.
15
16 ---------------------------------------------------------------------
17
18
19
20 ---------------------------------------------------------------------
21
22 1. Dynamic ddd_CplTable.
23 ddd_CplTable, ddd_NCplTable and ddd_ObjTable are allocated
24 dynamically since 1.8.10. However, the strategy is double-on-overflow,
25 which temporarily needs more memory than with double-sized tables.
26 There should be an option (at compile-time) to switch the
27 strategy to kind-of segmented, i.e., segments of fixed size are
28 allocated and freed on demand. Then, accessing the tables will
29 be more time-expensive, but we could handle more delicate
30 low-memory situations.
31
32 ---------------------------------------------------------------------
33
34 2. DDD_NCopiesPrio().
35 Implement a function returning the number of object copies with
36 a given priority.
37
38 ---------------------------------------------------------------------
39
40 3. XferCopyObjX and additional bytes.
41 XferCopyObjX sends objects which are bigger than their
42 declared DDD_TYPE. There should be a way to specify whether the
43 additional data bytes are LDATA or GDATA or a mixture of both.
44
45 ---------------------------------------------------------------------
46
47 4. IF-communication and DDD_ATTR.
48 In all IF-communication functions there could be an additional
49 function with masked DDD_ATTR, i.e. a value and a mask. With this
50 feature one could communicate over a set of DDD_ATTR-subinterfaces.
51
52 ---------------------------------------------------------------------
53
54 5. Integration check via DDD_HEADER checksums in debug-mode.
55 There should be a debug-mode which allows to detect memory errors
56 (bad pointers). In this debug-mode, each DDD_HEADER contains a
57 checksum variable, which is a parity-like encoding of the data in the
58 DDD_HEADER. Before each time the HEADER data is accessed (i.e.,
59 read _and_ write access), the checksum is computed and compared
60 with the stored one. Errors according to bad pointers and memory
61 faults could be found more easily.
62
63 ---------------------------------------------------------------------
64
65 6. Communicators.
66 MPI-like communicators into DDD. This should be supported already
67 by PPIF. (i.e. groups of processors execute certain DDD functions,
68 like Interface-comm, Xfer, Identify).
69
70 ---------------------------------------------------------------------
71
72 7. Make Notify-module faster.
73 Perhaps it is possible to schedule the small messages in Notify not
74 only through the PPIF-tree, but also to immediate processor neighbours,
75 where this is possible. This possibly can eliminate bottlenecks on
76 the master processor.
77
78 ---------------------------------------------------------------------
79
80 8. F_FRONTEND: ARCH_FLINK.
81 Introduce ARCH_FLINK definition in all arch subdirectories. This
82 could be necessary for the F_FRONTEND production version.
83
84 ---------------------------------------------------------------------
85
86 9. AddData without previous XferCopyObj.
87 It should be possible to start the Xfer-communication directly with
88 calls to DDD_XferAddData (with previous specification of a
89 destination processor). This would enable simple Xfers based on
90 objects w/o DDD_HEADER.
91
92 ---------------------------------------------------------------------
93