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
|
<!DOCTYPE HTML PUBLIC "-//IETF//DTD HTML//EN">
<html>
<head>
<title>Data Caching</title>
</head>
<body>
<h1>Meta Data Caching</h1>
<p>The HDF5 library caches two types of data: meta data and raw
data. The meta data cache holds file objects like the file
header, symbol table nodes, global heap collections, object
headers and their messages, etc. in a partially decoded
state. The cache has a fixed number of entries which is set with
the file access property list (defaults to 10k) and each entry
can hold a single meta data object. Collisions between objects
are handled by preempting the older object in favor of the new
one.
<h1>Raw Data Chunk Caching</h1>
<p>Raw data chunks are cached because I/O requests at the
application level typically don't map well to chunks at the
storage level. The chunk cache has a maximum size in bytes
set with the file access property list (defaults to 1MB) and
when the limit is reached chunks are preempted based on the
following set of heuristics.
<ul>
<li>Chunks which have not been accessed for a long time
relative to other chunks are penalized.
<li>Chunks which have been accessed frequently in the recent
past are favored.
<li>Chunks which are completely read and not written, completely
written but not read, or completely read and completely
written are penalized according to <em>w0</em>, an
application-defined weight between 0 and 1 inclusive. A weight
of zero does not penalize such chunks while a weight of 1
penalizes those chunks more than all other chunks. The
default is 0.75.
<li>Chunks which are larger than the maximum cache size do not
participate in the cache.
</ul>
<p>One should choose large values for <em>w0</em> if I/O requests
typically do not overlap but smaller values for <em>w0</em> if
the requests do overlap. For instance, reading an entire 2d
array by reading from non-overlapping "windows" in a row-major
order would benefit from a high <em>w0</em> value while reading
a diagonal accross the dataset where each request overlaps the
previous request would benefit from a small <em>w0</em>.
<h1>The API</h1>
<p>The cache parameters for both caches are part of a file access
property list and are set and queried with this pair of
functions:
<dl>
<dt><code>herr_t H5Pset_cache(hid_t <em>plist</em>, unsigned int
<em>mdc_nelmts</em>, size_t <em>rdcc_nbytes</em>, double
<em>w0</em>)</code>
<dt><code>herr_t H5Pget_cache(hid_t <em>plist</em>, unsigned int
*<em>mdc_nelmts</em>, size_t *<em>rdcc_nbytes</em>, double
<em>w0</em>)</code>
<dd>Sets or queries the meta data cache and raw data chunk cache
parameters. The <em>plist</em> is a file access property
list. The number of elements (objects) in the meta data cache
is <em>mdc_nelmts</em>. The total size of the raw data chunk
cache and the preemption policy is <em>rdcc_nbytes</em> and
<em>w0</em>. For <code>H5Pget_cache()</code> any (or all) of
the pointer arguments may be null pointers.
</dl>
<hr>
<address><a href="mailto:matzke@llnl.gov">Robb Matzke</a></address>
<!-- Created: Tue May 26 15:20:14 EDT 1998 -->
<!-- hhmts start -->
Last modified: Tue May 26 15:38:27 EDT 1998
<!-- hhmts end -->
</body>
</html>
|
