Skip to content

Experiments.md

Latest commit

 

History

History
239 lines (202 loc) · 4.29 KB

Experiments.md

File metadata and controls

239 lines (202 loc) · 4.29 KB
Raw

Expertiments Register

Changes on old code

  • Date: 2026-02-12
  • Git Hash: fd5b2273ad7f51e50a5432a1154c757561650753
  • Command: sbatch jobfile.slurm
  • Logs Prefix: monaco/logs/frollino

Motivation

Implemented the lookup of the last hash prefix to avoid comparisons in the rehashes, plus some small random tweaks

Data

dataset,exact,approx
census,74.48,11.63
ht,638.34,555.61
pamap2,,
chem,,
fashion,6.14,4.54
nytimes,22.19,14.33
glove,321.78,97.14
sift,93.32,35.29
gist,,
simplewiki,83.02,26.93
landmark,260.61,53.42
imagenet,238.27,70.62

Result

Great results, assertion fails on some datasets

Some modifications to the code

  • Date: 2026-02-12
  • Git Hash: d79309db26925ed8e7aa6eadf057e8bea54ed02c
  • Command: sbatch jobfile.slurm
  • Logs Prefix: monaco/logs/spring

Motivation

Implemented the lookup of the last hash prefix to avoid comparisons in the rehashes, plus some small random tweaks

Data

dataset,exact,approx
census,344.30,119.68
ht,589.74,587.32
pamap2,3103.90,1723.80
chem,,
fashion,41.23,6.17
nytimes,1809.88,106.79
glove,11481.90,7344.73
sift,1443.00,1001.30
gist,,
simplewiki,16654.59,1001.30
landmark,3061.66,464.62
imagenet,1443.00,162.48

Result

It's finally working on all datasets, but it's kinda slow

So there's this very old version

  • Date: 2026-01-30
  • Git Hash: 66f47413d136228bbbc1d6d173c864f8c1e32d91
  • Command: sbatch jobfile.slurm
  • Logs Prefix: monaco/logs/cefixima

Motivation

Bisect to see where the performance degradation started

Data

dataset,exact,approx
census,25.14,26.22
ht,NOT CONNECTED,
pamap2,371.82,365.28
chem,,
fashion,15.26,8.25
nytimes,1758.25,118.59
glove,,
sift,2323.33,142.49
gist,,
simplewiki,,
landmark,,
imagenet,,

Result

[still running]

So there's this old version

  • Date: 2026-01-29
  • Git Hash: 7bbcd6ee171154af87b1bcf13a9659183892d750
  • Command: sbatch jobfile.slurm
  • Logs Prefix: monaco/logs/azitromicina

Motivation

Bisect to see where the performance degradation started

Data

dataset,exact,approx
census,,
ht,,
pamap2,,
chem,,
fashion,32.93,5.78
nytimes,1709.60,370.79
glove,TIMEOUT,
sift,5126.96,165.68
gist,,
simplewiki,15747.65,1221.39
landmark,,
imagenet,,

Result

Bad version again

Do we really need 1000 repetitions?

  • Date: 2026-01-28
  • Git Hash: 66bae24a855027aa0159ce7e08d2979b31f1ba13
  • Command: sbatch jobfile.slurm
  • Logs Prefix: monaco/logs/amoxicillina

Motivation

See what happends if we reduce the number of repetitions from 1000 to 500 on the current version of the code

Data

dataset,exact,approx
census,,
ht,,
pamap2,,
chem,,
fashion,221.79,7.66
nytimes,4293.34,312.10
glove,,
sift,,
gist,,
simplewiki,,
landmark,,
imagenet,,

Result

Yes we do, the rehashes are too costly [still running]

Is the main working?

  • Date: 2026-01-28
  • Git Hash: 66bae24a855027aa0159ce7e08d2979b31f1ba13
  • Command: sbatch jobfile.slurm
  • Logs Prefix: monaco/logs/balenot

Motivation

Check the point of the current version of the code

Data

dataset,exact,approx
census,1137.41,218.95
ht,5704.30,2789.55
pamap2,,
chem,,
fashion,31.59,9.21
nytimes,2794.01,164.72
glove,FAILED,
sift,1861.01,245.68
gist,FAILED,
simplewiki,13615.06,996.67
landmark,4368.69,153.68
imagenet,,

Result

Fails on gist, glove

Bloom Filter to avoid duplicate comparisons

  • Date: 2026-01-27
  • Git Hash: 4b5da83e6bc53151d43c106efa62065179bbae7c
  • Command: sbatch jobfile.slurm
  • Logs Prefix: monaco/logs/baleno

Motivation

Can we use a Bloom filter to avoid the comparison of pairs we already seen in other rehashes?

Data

dataset,exact,approx
census,388.86,86.22
ht,,
pamap2,,
chem,,
fashion,28.69,4.16
nytimes,,
glove,,
sift,,
gist,,
simplewiki,,
landmark,7370.41,159.12
imagenet,,

Result

I guess not!

[Title of Experiment]

  • Date: YYYY-MM-DD
  • Git Hash: [hash]
  • Command: sbatch jobfile.slurm
  • Logs Prefix: [name]

Motivation

[Description of why this experiment is being conducted]

Data

dataset,exact,approx
census,,
ht,,
pamap2,,
chem,,
fashion,,
nytimes,,
glove,,
sift,,
gist,,
simplewiki,,
landmark,,
imagenet,,

Result

[Sintesi dei risultati e conclusioni]