Documentation Source Text

Check-in [ac27ab7f6f]
Login

Many hyperlinks are disabled.
Use anonymous login to enable hyperlinks.

Overview
Comment:Updates to the 35% faster document.
Timelines: family | ancestors | descendants | both | trunk
Files: files | file ages | folders
SHA3-256:ac27ab7f6f0995ed596eecbcdd57b40c97bb0ef93b7d55169616a2148d53b01d
User & Date: drh 2017-05-05 16:55:30
Context
2017-05-10
16:34
Add the SQLITE_READ authorizer change to the change log for 3.19.0. check-in: 84e21ebf22 user: drh tags: trunk
2017-05-05
16:55
Updates to the 35% faster document. check-in: ac27ab7f6f user: drh tags: trunk
2017-05-04
20:23
More tweaks to the 35% faster document. check-in: 3563b95114 user: drh tags: trunk
Changes
Hide Diffs Unified Diffs Ignore Whitespace Patch

Changes to pages/fasterthanfs.in.

11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
...
114
115
116
117
118
119
120
121
122
123















124
125
126
127
128
129
130
131
132
133
...
139
140
141
142
143
144
145







<p>Furthermore, a single SQLite database holding many thousands
of thumbnail-sized blobs uses about 20% less disk space than
storing the blobs in individual files.

<p>The performance difference arises (we believe) because when
reading from an SQLite database, the open() and close() system calls
only need to be invoked once, whereas
open() and close() must be invoked once for each blob
when reading the blobs from individual files.  In other words, the
overhead of calling open() and close() is greater than the overhead
of using the database.  The size reduction arises from the fact that
individual files are padded out to the next multiple of the filesystem
block size, whereas the blobs are packed more tightly into an SQLite
database.

<h1>How Performance Was Measured</h1>

<p>The performance comparison is accomplished using the
[https://www.sqlite.org/src/file/test/kvtest.c|kvtest.c] program
found in the SQLite source tree.
To compile the test program, first gather the kvtest.c source file
into a directory with the [amalgamation|SQLite amalgamation] source
files "sqlite3.c" and "sqlite3.h".  Then on unix, run a command like
................................................................................
option, a separate SQL statement is run to read each blob and
the performance of reading from the database is approximately
the same as the performance from reading directly from files.
This is still a significant finding, since few people would
expect a [full-featured SQL] database to run as fast as direct file reads,
and yet SQLite does.

<p>If the --random option is added to the "run" command, that causes the
blobs to be read in a random order.  This causes a noticable decrease in
the performance of database reads, since it forces more movement of B-Tree















cursors.  When --random is used and --blob-api is omitted, reading directly
from files on disk is generally a little faster, but reads from the database
are still competitive.

<h1>Other Considerations</h1>

<p>Some other SQL database engines advise developers to store blobs in separate
files and then store the filename in the database.  In that case, where
the database must first be consulted to find the filename before opening
and reading the file, simply storing the entire blob in the database is
................................................................................
we expect that write performance into SQLite will be slower than writing
directly to individual files.  However, if ACID transactions are disabled
via [PRAGMA journal_mode|PRAGMA journal_mode=OFF]
(thus putting SQLite on equal footing with the filesystem) and the
[sqlite3_blob_write()] interface is used, SQLite might well be competitive
or even faster than writes to separate files on disk.  That is an
experiment we have not yet run.













|
|
|






|







 







|
|
|
>
>
>
>
>
>
>
>
>
>
>
>
>
>
>
|
|
|







 







>
>
>
>
>
>
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
...
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
...
154
155
156
157
158
159
160
161
162
163
164
165
166

<p>Furthermore, a single SQLite database holding many thousands
of thumbnail-sized blobs uses about 20% less disk space than
storing the blobs in individual files.

<p>The performance difference arises (we believe) because when
reading from an SQLite database, the open() and close() system calls
are invoked only once, whereas
open() and close() are invoked once for each blob
when reading the blobs from individual files.  It appears that the
overhead of calling open() and close() is greater than the overhead
of using the database.  The size reduction arises from the fact that
individual files are padded out to the next multiple of the filesystem
block size, whereas the blobs are packed more tightly into an SQLite
database.

<h1>How These Measurements Are Made</h1>

<p>The performance comparison is accomplished using the
[https://www.sqlite.org/src/file/test/kvtest.c|kvtest.c] program
found in the SQLite source tree.
To compile the test program, first gather the kvtest.c source file
into a directory with the [amalgamation|SQLite amalgamation] source
files "sqlite3.c" and "sqlite3.h".  Then on unix, run a command like
................................................................................
option, a separate SQL statement is run to read each blob and
the performance of reading from the database is approximately
the same as the performance from reading directly from files.
This is still a significant finding, since few people would
expect a [full-featured SQL] database to run as fast as direct file reads,
and yet SQLite does.

<p>The --random option on the "run" command causes the
blobs to be read in a random order.  This causes the performance of database
reads to decrease.  The reason is that the blobs are tightly packed in
the database, rather than being padded out to the next block size as when
they are stored in the filesystem.  Some pages contain parts of
adjacent blobs.  When the blobs are read sequentially, those pages are
only read into memory once and cached and then used to reconstruct
adjacent blobs, but when blobs are read in a random order, those pages
that share parts of two or more blobs tend to be read multiple times,
leading to decreased performance.

<p>The "--mmap SIZE" option on the "run" command causes the database file
to be accessed using mmap() instead of via read().  The SIZE argument is
the size of the memory mapped region, and should be the size of the database
file for maximum performance.  Using "--mmap 1G" causes the database reads
to be almost twice as fast as disk reads even when the --random
option is used.

<p>When --random is used and both --blob-api and --mmap are omitted,
reading directly from files on disk is generally a little faster, but
reads from the database are still competitive.

<h1>Other Considerations</h1>

<p>Some other SQL database engines advise developers to store blobs in separate
files and then store the filename in the database.  In that case, where
the database must first be consulted to find the filename before opening
and reading the file, simply storing the entire blob in the database is
................................................................................
we expect that write performance into SQLite will be slower than writing
directly to individual files.  However, if ACID transactions are disabled
via [PRAGMA journal_mode|PRAGMA journal_mode=OFF]
(thus putting SQLite on equal footing with the filesystem) and the
[sqlite3_blob_write()] interface is used, SQLite might well be competitive
or even faster than writes to separate files on disk.  That is an
experiment we have not yet run.

<p>Remember that the relative performance of database reads and reads from
the filesystem will depend on both the hardware and the operating system.
Please try the tests above on your own system.  If you encounter cases
there database reads do not perform favorably in comparison to filesystem
reads, please report your findings in the [mailing lists|SQLite mailing list].