Musing about using profile-guided-optimization for PostgreSQL JIT, to improve future JITed programs would be worthwhile. Difficulty would be when and for how long to to train, and how to deal with different workloads.
-
-
Replying to @AndresFreundTec
pg_ctl --volkswagen to detect benchmark runs?
1 reply 0 retweets 1 like -
Replying to @d_gustafsson
:P If we had caching (which we won't for v11), I'd say we count how often a certain expression is used and then start to instrument it after a threshold and collect profiling data *just* for that expression. And then optimize just that expression with PGO.
2 replies 0 retweets 1 like -
Replying to @AndresFreundTec
Sounds reasonable. Is there an estimated cost involved such that we can profile expressions that we thought would take X but ended up taking kX, and profile expressions were k > something-clever?
1 reply 0 retweets 0 likes -
Replying to @d_gustafsson
I doubt it. For one I've no idea how to estimate expected time, for another measuring the timing would be expensive.
1 reply 0 retweets 0 likes -
Replying to @AndresFreundTec
Figured as much. Would be neat to profile outliers and anomalies for long tail optimization on top of hot spots.
1 reply 0 retweets 0 likes
profile != profile guided optimization. I don't think monitoring performance for outliers should be tied to PGO.
Loading seems to be taking a while.
Twitter may be over capacity or experiencing a momentary hiccup. Try again or visit Twitter Status for more information.