You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
{{ message }}
Repository navigation
Update ClickBench benchmarks with DataFusion 34 #8789
I re-ran the benchmark and v34 is actually a little bit slower than v33. I filter the benchmark for just type: "stateless" on machine: "c6a.4xlarge, 500gb gp2" to compare against similar systems. You can compare the index.html from the PR to the one on the site: filtered benchmark.
I re-ran the benchmark and v34 is actually a little bit slower than v33. I filter the benchmark for just type: "stateless" on machine: "c6a.4xlarge, 500gb gp2" to compare against similar systems. You can compare the index.html from the PR to the one on the site: filtered benchmark.
Do you have a sense for how reproduceable the numbers are? Many of the reported performance differences are less than 100ms, and at that range I have found significant variation.
I wonder if you ran it more than once and are they reproducable?
Do you have a sense for how reproduceable the numbers are? Many of the reported performance differences are less than 100ms, and at that range I have found significant variation.
I wonder if you ran it more than once and are they reproducible?
I ran this on the same instance that I ran the v33 benchmarks on (it's been powered down since the v33 run). The numbers are consistently slightly slower compared to v33. I ran it twice, since it was unexpected. I also updated to latest Rust and ran again -- that's what I published.
I'll run the v33 benchmark again on the same instance and see what it produces.
I've re-run the v33 benchmarks on the same instance and modified the benchmark so it will display both 33 and 34 at the same time so you can compare the runs:
Activity
kmitchener commentedon Jan 9, 2024
I re-ran the benchmark and v34 is actually a little bit slower than v33. I filter the benchmark for just type: "stateless" on machine: "c6a.4xlarge, 500gb gp2" to compare against similar systems. You can compare the index.html from the PR to the one on the site: filtered benchmark.
alamb commentedon Jan 9, 2024
Thanks @kmitchener !
Do you have a sense for how reproduceable the numbers are? Many of the reported performance differences are less than 100ms, and at that range I have found significant variation.
I wonder if you ran it more than once and are they reproducable?
kmitchener commentedon Jan 9, 2024
I ran this on the same instance that I ran the v33 benchmarks on (it's been powered down since the v33 run). The numbers are consistently slightly slower compared to v33. I ran it twice, since it was unexpected. I also updated to latest Rust and ran again -- that's what I published.
I'll run the v33 benchmark again on the same instance and see what it produces.
kmitchener commentedon Jan 9, 2024
I've re-run the v33 benchmarks on the same instance and modified the benchmark so it will display both 33 and 34 at the same time so you can compare the runs:

You can grab that from -> https://github.com/kmitchener/ClickBench/blob/new-run-of-datafusion-33/index.html
alamb commentedon Jan 10, 2024
🤔 I don't really have any good reason / understanding of what is different
I looked at q13 briefly which has the largest reported change and I don't really have any good explanation about why it is reported 8% slower.
And I profiled running it locally,

It doesn't seem to show anything obvious -- the biggest offenders are managing the group keys which I don't think changed much
alamb commentedon Jan 11, 2024
Thanks @kmitchener -- I filed #8836 to look into the results
alamb commentedon Feb 29, 2024
The results for 34 were merged, and we just release 36 (including #8827) . Let's see track the work for Updating ClickBench for DataFusion 36 in #9404
Thanks again @kmitchener