Summary: Can now replicate TPC-H SF10 tuning results on dev 4. Note that embedding has not yet been replicated.
Demo:
After running HPO on dev 4 for 18 hours, we get to a runtime of 110s. We run with --duration=4, and 110s matches the performance in the Proto-X paper after 4 hours of tuning.
Details:
Because I was only replicating tuning, I used an already-trained embedder. To replicate, use the dbgym_manual/embedding/ folder on dev4.
The command I used was python3 task.py --no-startup-check tune protox agent hpo tpch --scale-factor 10 --max-concurrent 4 --duration 4 --intended-pgdata-hardware ssd --pgdata-parent-dpath /mnt/nvme1n1/phw2/dbgym_tmp/
Summary: Can now replicate TPC-H SF10 tuning results on dev 4. Note that embedding has not yet been replicated.
Demo: After running HPO on dev 4 for 18 hours, we get to a runtime of 110s. We run with
--duration=4
, and 110s matches the performance in the Proto-X paper after 4 hours of tuning.Details:
dbgym_manual/embedding/
folder on dev4.python3 task.py --no-startup-check tune protox agent hpo tpch --scale-factor 10 --max-concurrent 4 --duration 4 --intended-pgdata-hardware ssd --pgdata-parent-dpath /mnt/nvme1n1/phw2/dbgym_tmp/