Hi,
I am working on using the read stream in autoprewarm. I observed ~10%
performance gain with this change. The patch is attached.
The downside of the read stream approach is that a new read stream
object needs to be created for each database, relation and fork. I was
wondering if this would cause a regression but it did not (at least
depending on results of my testing). Another downside could be the
code getting complicated.
For the testing,
- I created 50 databases with each of them having 50 tables and the
size of the tables are 520KB.
- patched: 51157 ms
- master: 56769 ms
- I created 5 databases with each of them having 1 table and the size
of the tables are 3GB.
- patched: 32679 ms
- master: 36706 ms
I put debugging message with timing information in
autoprewarm_database_main() function, then run autoprewarm 100 times
(by restarting the server) and cleared the OS cache before each
restart. Also, I ensured that the block number of the buffer returning
from the read stream API is correct. I am not sure if that much
testing is enough for this kind of change.
Any feedback would be appreciated.
--
Regards,
Nazir Bilal Yavuz
Microsoft
