On Wed, Jul 23, 2025 at 9:59 PM Peter Geoghegan <pg@bowt.ie> wrote:
> Tomas' index-prefetch-simple-master branch:
> │ I/O Timings: shared read=1490.918
> │ Execution Time: 2015.731 ms
> Complex patch (same prewarming/eviction are omitted this time):
> │ I/O Timings: shared read=138.856
> │ Execution Time: 768.454 ms
> I'm using direct IO in both cases. This can easily be repeated, and is stable.
Forget to add context about the master branch: Master can do this in
2386.850 ms, with "I/O Timings: shared read=1825.161". That's with
buffered I/O (not direct I/O), and with the same pg_prewarm +
pg_buffercache_evict_relation function calls as before. I'm running
"echo 3 > /proc/sys/vm/drop_caches" to drop the filesystem cache here,
too (unlike when testing the patches, where my use of direct i/o makes
that step unnecessary).
In summary, the simple patch + direct I/O clearly beats the master
branch + buffered I/O here -- though not by much. While the complex
patch gets a far greater benefit.
--
Peter Geoghegan
