
If you were logged in you would be able to see more operations.
|
|
|
QA Status: |
Done successfully
|
Test Details: |
NOTE.
Actual execution is skipped in case when FB works in SC/CS mode (see stub in the code of test).
Do not forget to remove this stub after improvement will be implemented for SC/CS.
29.10.2018: deferred change min_version to 2.5.9 , have Q expected result. Sent letter to dimitr, 29.10.2018 22:23
NOTE.
Actual execution is skipped in case when FB works in SC/CS mode (see stub in the code of test).
Do not forget to remove this stub after improvement will be implemented for SC/CS.
29.10.2018: deferred change min_version to 2.5.9 , have Q expected result. Sent letter to dimitr, 29.10.2018 22:23
|
Currently engine writes contents of Header and TIP pages to disk immediately after any change (for ex. transaction start increments value of "Next transaction" counter, commit changes transaction state etc).
I.e. Header and TIP pages are written to disk not waiting for commit (flush).
This was done by at least two major reasons:
a) safety
b) in CS (and SC) page locks for Header and TIP pages is released immediately after use to avoid AST overhead
In the case of many light read-only transactions these writes makes almost 100% of whole transaction time.
Since read-only transactions does not change data it is safe to update on disk contents of Header and TIP pages
somewhat later, for example when write transaction started (or committed) or when whole page cache is flushed
at database shutdown.
Note, reason (b) doesn't affect SuperServer.
So, improvement is that read-only transaction will not force write of Header\TIP page content to disk immediately after change
(in SS mode, of course). In the case of many light read-only transactions it gives significant performance gain.
PS to address this issue in CS mode there is more complex solution, probably it will be implemented in v4 a bit later
|
Description
|
Currently engine writes contents of Header and TIP pages to disk immediately after any change (for ex. transaction start increments value of "Next transaction" counter, commit changes transaction state etc).
I.e. Header and TIP pages are written to disk not waiting for commit (flush).
This was done by at least two major reasons:
a) safety
b) in CS (and SC) page locks for Header and TIP pages is released immediately after use to avoid AST overhead
In the case of many light read-only transactions these writes makes almost 100% of whole transaction time.
Since read-only transactions does not change data it is safe to update on disk contents of Header and TIP pages
somewhat later, for example when write transaction started (or committed) or when whole page cache is flushed
at database shutdown.
Note, reason (b) doesn't affect SuperServer.
So, improvement is that read-only transaction will not force write of Header\TIP page content to disk immediately after change
(in SS mode, of course). In the case of many light read-only transactions it gives significant performance gain.
PS to address this issue in CS mode there is more complex solution, probably it will be implemented in v4 a bit later
|
Show » |
There are no comments yet on this issue.
|
|