| 1 |
|
| 2 |
Since the beginning, Firebird had no rules of how to allocate disk space |
| 3 |
for database file(s). It just writes new allocated pages in not determined |
| 4 |
order (because of dependencies between pages to serve "careful write" strategy). |
| 5 |
|
| 6 |
This approach is very simple but has some drawbacks : |
| 7 |
|
| 8 |
- because of not determined order of writes, there may be such situation when |
| 9 |
page cache contains many dirty pages at time when new allocated page must be |
| 10 |
written but can't because out of disk space. In such cases often all other |
| 11 |
dirty pages are lost because administrators prefer to shutdown database |
| 12 |
before making some space on disk available. This leads to serious corruptions. |
| 13 |
|
| 14 |
- allocating disk space by relatively small chunks may lead to significant file |
| 15 |
fragmentation at file system level and reduce performance of large scans (for |
| 16 |
example during backup). |
| 17 |
|
| 18 |
|
| 19 |
Using new ODS 11.1, Firebird changes its disk space allocation algorithm to |
| 20 |
avoid corruptions in out of disk space conditions and to give the file system a |
| 21 |
chance to avoid fragmentation. These changes are described below. |
| 22 |
|
| 23 |
a) Every newly allocated page is written on disk immediately before returning to |
| 24 |
the engine. If page can't be written then allocation doesn't happen, PIP bit |
| 25 |
is not cleared and appropriate IO error is raised. This error can't lead to |
| 26 |
corruptions as we have a guarantee that all dirty pages in cache have disk |
| 27 |
space allocated and can be written safely. |
| 28 |
|
| 29 |
This change makes one additional write of every newly allocated page compared |
| 30 |
with old behavior. So performance penalty is expected during the database file |
| 31 |
growth. To reduce this penalty Firebird groups writes of newly allocated pages |
| 32 |
up to 128KB at a time and tracks number of "initialized" pages at PIP header. |
| 33 |
|
| 34 |
Note : newly allocated page will be written to disk twice only if this page |
| 35 |
is allocated first time. I.e. if page was allocated, freed and allocated again |
| 36 |
it will not be written twice on second allocation. |
| 37 |
|
| 38 |
b) To avoid file fragmentation, Firebird used appropriate file system's API to |
| 39 |
preallocate disk space by relatively large chunks. Currently such API exists |
| 40 |
only in Windows but it was recently added into Linux API and may be implemented |
| 41 |
in such popular file system's as ext2, etc in the future. So this feature is |
| 42 |
currently implemented only in Windows builds of Firebird and may be implemented |
| 43 |
in Linux builds in the future. |
| 44 |
|
| 45 |
For better control of disk space preallocation, new setting in Firebird.conf |
| 46 |
was introduced : DatabaseGrowthIncrement. This is upper bound of preallocation |
| 47 |
chunk size in bytes. Default value is 128MB. When engine needs more disk space |
| 48 |
it allocates 1/16th of already allocated space but no less than 128KB and no more |
| 49 |
than DatabaseGrowthIncrement value. If DatabaseGrowthIncrement is set to zero then |
| 50 |
preallocation is disabled. Space for database shadow files is not preallocated. |
| 51 |
Also preallocation is disabled if "No reserve" option is set for database. |
| 52 |
|
| 53 |
Note : preallocation also allows to avoid corruptions in out of disk space |
| 54 |
condition - in such case there is a big chance that database has enough space |
| 55 |
preallocated to operate until administrator makes some disk space available. |
| 56 |
|
| 57 |
|
| 58 |
Author: Vlad Khorsun, <hvlad at users sourceforge net> |
| 59 |
|