CA-408552: Improve bootstrom performance by save db ops #6541
Conversation
|
|
liulinC
force-pushed
the
private/linl/perfpr
branch
from
7d4f2a8 to
1d43b77
Compare
There was a problem hiding this comment.
This is difficult to review on github due to all the moved code.
But the following settings in .config/git/config locally makes reviewing easier, this will showed code that was moved but otherwise unchanged in grey (and ignores indentation changes):
[diff]
algorithm = histogram
colorMoved = dimmed_zebra
colorMovedWS = allow-indentation-change
There was a problem hiding this comment.
It might be easier to review if this was split into multiple commits, e.g. one commit that just introduces the 'fun op' and reindents code, then another that introduces names for things we want to cache ahead of fun op (but doesn't move the code yet) (e.g. 'vdi_is_ha_state_or_redolog'), and then a final commit that just moves the cached values that don't depend on op outside of fun op.
I thought it was a simple update, as I only moved codes to the beginning with low risk. |
events_from_xenopsd thread is critical as it sync up VM status in case of bootstorm, this thread is flood as lots of events comes from xenopsd waiting for process. During processing of the events, VM/VDI/VBD update_allowed_operations will be called to refresh the allowed operations. However, for each ops (start/suspend,etc) for the same object(VM), the object info is always the same no matter what the ops is. Thus, it is not necessary to query the object information over and over again. Disclosure is used to resovle the issue. Query once and the disclosure will just remember the query result. The performance test for starting 500 VM on 4 hosts improve around 10% performance for both XS8 and XS9 This commit just introduce disclosure and caller call the disclosure instead of the original function Signed-off-by: Lin Liu <[email protected]>
Define db ops into variables and keep them inside returned function for code review Signed-off-by: Lin Liu <[email protected]>
Move ops unrelated db operation outside of returned function Signed-off-by: Lin Liu <[email protected]>
liulinC
force-pushed
the
private/linl/perfpr
branch
from
1d43b77 to
e1a57fa
Compare
Just follow the suggestion to split the commits, @edwintorok Please take a look, thanks. |
| Ok () | ||
| in | ||
| match op with | ||
| | `forget -> |
There was a problem hiding this comment.
This could be encoded using when condition ->; it's a matter or taste. I like when because it avoids nesting.
There was a problem hiding this comment.
I have the same taste.
But would just leave it as it was, to de-risk regression.
events_from_xenopsd thread is critical as it sync up VM status in case of bootstorm, this thread is flood as lots of events comes from xenopsd waiting for process.
During processing of the events, VM/VDI/VBD update_allowed_operations will be called to refresh the allowed operations.
However, for each ops (start/suspend,etc) for the same object(VM), the object info is always the same no matter what the ops is. Thus, it is not necessary to query the object information over and over again.
Disclosure is used to resovle the issue. Query once and the disclosure will just remember the query result.
The performance test for starting 500 VM on 4 hosts improve around 10% performance for both XS8 and XS9