-
Notifications
You must be signed in to change notification settings - Fork 41
/
TODO-1.2
50 lines (34 loc) · 2.12 KB
/
TODO-1.2
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
- Make a utility for HPSS that will make an index only file
- fsck utility to fix files that were not fully completely created correctly.
- Is the Fortran module interface complete?
- Do we need the ADIOS_WRITE macros anymore? If so, we need to do this:
#define ADIOS_WRITE(a,b) call adios_write(a,'b'//char(0),b)
to
#define ADIOS_WRITE(a,b) call adios_write(a,#b//char(0),b)
- Auto generate the index file intended for the auto-split output (POSIX).
- Formalize the resilience features of BP and add any features necessary.
- fix MPI_STRIPE method to not require the conditional compilation to work
properly.
- Figure out why and fix append mode slowdown.
- Add write="no" back in. This will force any array dimensions that use this
to be stored as absolute numbers rather than references to vars.
- Figure out how to use the same XML file for the app and staging. We need
some flag to recognize and parse differently. The idea is that we will specify
the transport for a group is the staging method, but add a parameter, for
example, that states which method to actually use in the staging area to write
the data to disk. There might be more changes necessary to make this work.
- Make the group_size optional. This would require making the buffer a
maximum size and delaying the offset calculation until the close operation.
There is also the issue of what to do if the buffer is exceeded. It probably
only makes sense to abort. Otherwise, we'd have to do a big synch point that
some processes might not be expecting. I don't think this is really a reasonable
thing to do. Abort is about the only thing that can really work consistently.
- Consider refiniements to the free-memory-percentage to better address
very dynamic memory usage.
- Add explicit support for "no values present" like NetCDF has.
- Add a checksum data characteristic to the writing so that we can validate
the data on reading.
- fix the phdf5 method so that it uses the progressive assembly approach
of the nc4 method.
- fix the tests/genarray so that the 2d and 3d XML files don't collide on
the group name internally (gpp.py will overwrite one with the other).