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
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
|
VFD SWMR Punch List
4 November 2019
1. Design optimizations in index writes. Perhaps write deltas only,
with a full index every n ticks?
2. Design work for NFS and object store versions.
3. Design work for journaling variation.
4. Add support for opening multiple files in either VFD SWMR writer or
VFD SWMR writer mode. See addition of the EOT queue in section 3.2.2
of the RFC, and related changes in sections 3.3 and 3.3.2.
5. Add the pb\_expansion\_threshold field to the
H5F\_vfd\_swmr\_config\_t structure, and update
H5Pset\_vfd\_swmr\_config()and
H5Pget\_vfd\_swmr\_config()accordingly.
6. Implement end tick now API call. See section 3.1.2 of the RFC for
specifications.
7. Implement enable / disable EOT call. See section 3.1.3 of the RFC
for specifications.
8. Implement option to flush raw data as part of EOT -- write
performance test to characterize the performance hit.
9. Move VFD SWMR specific H5F code to its own file -- say
H5Fvfd\_swmr.c.
10. Update the new page buffer to support the pb\_expansion\_threshold,
and trigger an early end of tick if the threshold is exceeded.
11. Modify metadata file write call to allow the location of the index
to float and thus be of arbitrary size.
12. Add code to remove entries from the index after they have been
written to the HDF5 file, and have not been modified for at least
max\_lag ticks.
13. Tidy short cuts in the initial implementation. These include:
- Odd behavior in the superblock refresh routine (see comments in
code). Figure out what is going on, and then either bypass the
issue or fix it as seems appropriate.
- Comment H5F\_vfd\_swmr\_config\_t in H5Fpublic.h properly.
- Cleanup EOA hack in H5FD\_read().
- Address file open failures in SWMR tests. These appear to be
cases in which the writer finishes before the reader opens --
causing the reader to fail as the metadata file no longer
exists. If so, handle this more gracefully.
14. Add support for specifying the VFD that sits under the VFD SWMR
reader VFD. Probably do this as part of the pluggable VFD design
that Jake is working on.
15. Modify the metadata cache so that we don't allocate space for the
page hash table unless the file is opened in VFD SWMR reader mode.
16. Implement the logging facility (section 3.14 RFC)
17. Test code to expose existing page buffer bugs and fix same. Note
that they seem to appear primarily on Jelly.
- DLL pre remove sanity check assertion failure
- vfd\_swmr\_addrem\_writer: H5PB.c:2981:
H5PB\_\_mark\_entry\_dirty: Assertion
'(entry\_ptr)-\>delay\_write\_until \> (pb\_ptr)-\>cur\_tick'
failed.
18. Flesh out designs for unit, integration and performance tests suites
as outlined in the RFC. Implement same.
19. Fix memory leak in sparse-reader test.
20. Test John's patch that repairs the superblock flags mismatch that
crashes the reader.
21. Investigate a potential time-of-check, time-of-use race condition
involving EOA/EOF and the skip\_read variable in some of the H5PB
routines.
22. Understand use of H5F\_t on branch feature/vfd\_swmr instead of
H5C\_t as on develop branch.
23. *Temporarily* reserve a new superblock flag for VFD SWMR for
development purposes.
24. Prior to feature merge, *permanently* reserve a new superblock flag.
25. Revisit the global heap. The global heap receives data types (a form
of metadata) and variable-length (VL) data such as strings. It is
translated to raw data on its way down the HDF5 software stack. If
the global heap is modified/replaced as currently planned, then VFD
SWMR does not have to deal with it. However, if the global heap
overhaul does not take place, then we have more work to do.
|
