wip: sql: dynamically disable vectorized execution

[mgartner]

No description provided.

[blathers-crl]

It looks like your PR touches production code but doesn't add or edit any test code. Did you consider adding tests to your PR?

_{🦉 Hoot! I am a Blathers, a bot for CockroachDB. My owner is dev-inf.}

[cockroach-teamcity]

This change is 

[mgartner]

Preliminary results:

name                                  old time/op    new time/op    delta
Sysbench/SQL/3node/oltp_point_select     166µs ± 7%     161µs ± 5%    ~     (p=0.095 n=5+5)
Sysbench/SQL/3node/oltp_read_only       2.89ms ± 4%    2.75ms ± 5%    ~     (p=0.056 n=5+5)
Sysbench/SQL/3node/oltp_read_write      7.16ms ± 3%    6.87ms ± 2%  -4.00%  (p=0.008 n=5+5)

name                                  old alloc/op   new alloc/op   delta
Sysbench/SQL/3node/oltp_point_select    25.3kB ± 0%    25.2kB ± 0%  -0.44%  (p=0.032 n=5+5)
Sysbench/SQL/3node/oltp_read_only       1.15MB ± 8%    1.10MB ± 0%  -4.97%  (p=0.016 n=5+5)
Sysbench/SQL/3node/oltp_read_write    2.16MB ± 0%    2.18MB ±19%    ~     (p=1.000 n=4+5)

name                                  old allocs/op  new allocs/op  delta
Sysbench/SQL/3node/oltp_point_select       208 ± 0%       207 ± 0%  -0.48%  (p=0.029 n=4+4)
Sysbench/SQL/3node/oltp_read_only        4.06k ± 3%     3.96k ± 0%  -2.33%  (p=0.008 n=5+5)
Sysbench/SQL/3node/oltp_read_write       10.1k ± 2%      9.8k ± 6%    ~     (p=0.548 n=5+5)

I need to run with a higher --count.

[mgartner]

name                                  old time/op    new time/op    delta
Sysbench/SQL/3node/oltp_read_only       2.84ms ± 5%    2.75ms ± 4%  -3.40%  (p=0.019 n=10+10)
Sysbench/SQL/3node/oltp_write_only      4.07ms ± 5%    3.89ms ± 5%  -4.51%  (p=0.011 n=10+10)
Sysbench/SQL/3node/oltp_read_write      7.05ms ± 2%    6.87ms ± 1%  -2.54%  (p=0.000 n=10+8)
Sysbench/SQL/3node/oltp_point_select     162µs ± 5%     163µs ± 3%    ~     (p=0.842 n=10+9)
Sysbench/SQL/3node/oltp_begin_commit    43.4µs ± 3%    43.0µs ± 6%    ~     (p=0.579 n=10+10)

name                                  old errs/op    new errs/op    delta
Sysbench/SQL/3node/oltp_read_only         0.00           0.00         ~     (all equal)
Sysbench/SQL/3node/oltp_write_only        0.00           0.00         ~     (all equal)
Sysbench/SQL/3node/oltp_read_write        0.00           0.00         ~     (all equal)
Sysbench/SQL/3node/oltp_point_select      0.00           0.00         ~     (all equal)
Sysbench/SQL/3node/oltp_begin_commit      0.00           0.00         ~     (all equal)

name                                  old alloc/op   new alloc/op   delta
Sysbench/SQL/3node/oltp_read_only       1.10MB ± 0%    1.10MB ± 0%    ~     (p=0.645 n=8+8)
Sysbench/SQL/3node/oltp_write_only       909kB ±13%     932kB ±14%    ~     (p=0.853 n=10+10)
Sysbench/SQL/3node/oltp_read_write      2.16MB ± 3%    2.25MB ± 8%    ~     (p=0.601 n=7+10)
Sysbench/SQL/3node/oltp_point_select    25.2kB ± 0%    25.2kB ± 0%    ~     (p=0.289 n=10+10)
Sysbench/SQL/3node/oltp_begin_commit    10.4kB ± 0%    10.4kB ± 0%    ~     (p=0.868 n=10+10)

name                                  old allocs/op  new allocs/op  delta
Sysbench/SQL/3node/oltp_read_only        3.97k ± 1%     3.98k ± 1%    ~     (p=0.702 n=8+8)
Sysbench/SQL/3node/oltp_write_only       5.00k ± 6%     4.87k ± 4%  -2.72%  (p=0.022 n=10+10)
Sysbench/SQL/3node/oltp_read_write       10.0k ± 7%      9.8k ± 5%    ~     (p=0.143 n=10+10)
Sysbench/SQL/3node/oltp_point_select       207 ± 0%       207 ± 0%    ~     (p=1.000 n=10+10)
Sysbench/SQL/3node/oltp_begin_commit      55.0 ± 0%      55.0 ± 0%    ~     (all equal)

[mgartner]

Upon closer inspection, this change isn't working as intended and vectorization is not disabled. So the results are very suspect. I don't know what would explain these gains.

[mgartner]

I've fixed the issue, and am getting similar results, though slightly better. Probably time for a full sysbench run...

name                                  old time/op    new time/op    delta
Sysbench/SQL/3node/oltp_point_select     165µs ± 2%     152µs ± 6%  -7.85%  (p=0.000 n=8+10)
Sysbench/SQL/3node/oltp_read_only       2.84ms ± 4%    2.75ms ± 6%  -3.12%  (p=0.035 n=10+9)
Sysbench/SQL/3node/oltp_write_only      4.06ms ± 6%    3.89ms ± 5%  -4.07%  (p=0.023 n=10+10)
Sysbench/SQL/3node/oltp_read_write      7.11ms ± 3%    6.91ms ± 2%  -2.76%  (p=0.003 n=10+9)

name                                  old alloc/op   new alloc/op   delta
Sysbench/SQL/3node/oltp_point_select    25.2kB ± 0%    26.6kB ± 1%  +5.23%  (p=0.000 n=9+10)
Sysbench/SQL/3node/oltp_read_only       1.10MB ± 1%    1.11MB ± 0%  +0.84%  (p=0.008 n=8+9)
Sysbench/SQL/3node/oltp_write_only       926kB ±14%     891kB ±14%    ~     (p=0.247 n=10+10)
Sysbench/SQL/3node/oltp_read_write      2.27MB ±10%    2.28MB ±13%    ~     (p=0.529 n=10+10)

name                                  old allocs/op  new allocs/op  delta
Sysbench/SQL/3node/oltp_point_select       207 ± 0%       206 ± 0%  -0.39%  (p=0.009 n=10+10)
Sysbench/SQL/3node/oltp_read_only        4.03k ± 4%     3.96k ± 0%  -1.78%  (p=0.018 n=10+9)
Sysbench/SQL/3node/oltp_write_only       5.01k ± 6%     4.73k ± 5%  -5.47%  (p=0.007 n=10+10)
Sysbench/SQL/3node/oltp_read_write       9.82k ± 6%     9.58k ± 5%    ~     (p=0.109 n=10+9)

[mgartner]

I only measured a 1.64% increase in throughput from this.

-------------------------------------
 BEFORE
-------------------------------------
 Run 20250508_174235 - QPS: 22061.40
 Run 20250508_180340 - QPS: 21914.64
 Run 20250508_181509 - QPS: 21826.42
 Run 20250508_182638 - QPS: 21772.46
 Run 20250508_183806 - QPS: 22737.97
 Run 20250508_184934 - QPS: 22714.30
 Mean: 22171.19
 Median: 21988.02
 Standard Deviation: 402.47
 Coefficient of variance: .0183

-------------------------------------
 AFTER
-------------------------------------
 Run 20250508_192033 - QPS: 22315.28
 Run 20250508_193204 - QPS: 22631.58
 Run 20250508_194332 - QPS: 22830.21
 Run 20250508_195459 - QPS: 22460.44
 Run 20250508_200626 - QPS: 22593.65
 Run 20250508_201754 - QPS: 22378.25
 Mean: 22534.90
 Median: 22527.04
 Standard Deviation: 172.39
 Coefficient of variance: .0076

[mgartner]

Here's two runs I did without the changes in this PR, for comparison:

-------------------------------------
 vectorize=on
-------------------------------------
 Run 20250509_173918 - QPS: 22518.43
 Run 20250509_180010 - QPS: 22722.36
 Run 20250509_181144 - QPS: 22220.65
 Run 20250509_182318 - QPS: 22265.60
 Run 20250509_183451 - QPS: 22517.04
 Run 20250509_184625 - QPS: 21757.49
 Mean: 22333.59
 Median: 22391.32
 Standard Deviation: 307.74
 Coefficient of variance: .0137

-------------------------------------
 vectorize=off
-------------------------------------
Run 20250509_202802 - QPS: 22175.49
Run 20250509_203934 - QPS: 22227.52
Run 20250509_205107 - QPS: 22104.08
Run 20250509_210240 - QPS: 22290.24
Run 20250509_211414 - QPS: 22104.82
Run 20250509_212547 - QPS: 22126.31
Mean: 22171.41
Median: 22150.90
Standard Deviation: 68.66
Coefficient of variance: .0030