HippoBaro commented on PR #11278:
URL: https://github.com/apache/arrow-rs/pull/11278#issuecomment-5976249333

   > ```string_dictionary/bloom_filter                                  1.00    
101.5±1.26ms     2.5 GB/sec    1.01    102.5±1.49ms     2.5 > GB/sec
   > string_dictionary/cdc                                           1.19     
55.2±0.78ms     4.7 GB/sec    1.00     46.3±0.39ms     5.6 GB/sec
   > string_dictionary/default                                       1.00     
57.2±1.24ms     4.5 GB/sec    1.00     57.4±1.02ms     4.5 >GB/sec
   > string_dictionary/number_distinct_values                        1.00    
115.8±0.67ms     2.2 GB/sec    1.03    119.2±1.42ms     2.2 GB/sec
   > string_dictionary/parquet_2                                     1.00     
48.6±3.96ms     5.3 GB/sec    1.12     54.6±5.43ms     4.7 GB/sec
   > string_dictionary/zstd                                          1.00    
215.5±0.64ms  1225.8 MB/sec    1.02    219.0±1.02ms  1206.0 MB/sec
   > string_dictionary/zstd_parquet_2                                1.00    
193.0±4.40ms  1368.8 MB/sec    1.02    196.5±4.96ms  1343.9 MB/sec
   > ```
   
   These results are strange. I can’t find an algorithmic explanation for the 
opposite movements in `/cdc` and `/parquet_2`. All the changes here are in 
shared schema/type-checking setup, so I’d expect any direct overhead to affect 
the variants similarly rather than move them in opposite directions.
   
   I reran the benchmark multiple times on my Graviton4 benchmark box 
([configured like 
this](https://github.com/apache/arrow-rs/pull/11267#issuecomment-5876792873)). 
It uses the same Neoverse V2 cores, and I couldn’t reproduce either effect:
   
   ```text
   group                                                           base         
                          changed
   -----                                                           ----         
                          -------
   string_dictionary/bloom_filter                                  1.00     
95.9±0.16ms     2.7 GB/sec    1.00     96.1±0.41ms     2.7 GB/sec
   string_dictionary/cdc                                           1.00     
58.1±0.15ms     4.4 GB/sec    1.01     58.6±1.36ms     4.4 GB/sec
   string_dictionary/default                                       1.00     
58.7±0.45ms     4.4 GB/sec    1.01     59.1±0.61ms     4.4 GB/sec
   string_dictionary/number_distinct_values                        1.00    
119.4±2.07ms     2.2 GB/sec    1.01    120.7±0.21ms     2.1 GB/sec
   string_dictionary/parquet_2                                     1.00     
59.2±5.42ms     4.4 GB/sec    1.00     59.3±5.19ms     4.4 GB/sec
   string_dictionary/zstd                                          1.00    
222.8±0.67ms  1185.3 MB/sec    1.00    223.7±0.58ms  1180.7 MB/sec
   string_dictionary/zstd_parquet_2                                1.01    
210.0±4.14ms  1258.0 MB/sec    1.00    208.8±4.45ms  1265.2 MB/sec
   string_dictionary_low_cardinality_100/bloom_filter              1.00     
23.0±0.05ms   174.2 MB/sec    1.00     23.0±0.05ms   173.9 MB/sec
   string_dictionary_low_cardinality_100/cdc                       1.00     
25.2±0.06ms   158.7 MB/sec    1.01     25.5±0.37ms   156.9 MB/sec
   string_dictionary_low_cardinality_100/default                   1.00     
22.9±0.04ms   175.0 MB/sec    1.00     23.0±0.06ms   174.6 MB/sec
   string_dictionary_low_cardinality_100/number_distinct_values    1.00     
46.8±0.07ms    85.5 MB/sec    1.00     47.1±0.09ms    85.1 MB/sec
   string_dictionary_low_cardinality_100/parquet_2                 1.00     
22.8±0.05ms   176.0 MB/sec    1.01     22.9±0.06ms   175.0 MB/sec
   string_dictionary_low_cardinality_100/zstd                      1.00     
23.2±0.05ms   172.4 MB/sec    1.00     23.3±0.06ms   171.7 MB/sec
   string_dictionary_low_cardinality_100/zstd_parquet_2            1.00     
23.1±0.05ms   173.3 MB/sec    1.01     23.3±0.07ms   172.2 MB/sec
   string_dictionary_low_cardinality_20/bloom_filter               1.00     
23.3±0.05ms   171.6 MB/sec    1.00     23.4±0.05ms   171.3 MB/sec
   string_dictionary_low_cardinality_20/cdc                        1.00     
25.6±0.07ms   156.4 MB/sec    1.02     26.1±0.73ms   153.6 MB/sec
   string_dictionary_low_cardinality_20/default                    1.00     
23.3±0.06ms   172.3 MB/sec    1.00     23.3±0.05ms   171.9 MB/sec
   string_dictionary_low_cardinality_20/number_distinct_values     1.00     
48.0±0.07ms    83.5 MB/sec    1.00     47.8±0.07ms    83.8 MB/sec
   string_dictionary_low_cardinality_20/parquet_2                  1.00     
23.2±0.05ms   172.4 MB/sec    1.00     23.2±0.05ms   172.5 MB/sec
   string_dictionary_low_cardinality_20/zstd                       1.00     
23.4±0.06ms   171.3 MB/sec    1.00     23.4±0.05ms   170.9 MB/sec
   string_dictionary_low_cardinality_20/zstd_parquet_2             1.00     
23.3±0.05ms   171.8 MB/sec    1.01     23.4±0.05ms   170.9 MB/sec
   string_dictionary_low_cardinality_400/bloom_filter              1.01     
23.8±0.07ms   168.9 MB/sec    1.00     23.6±0.07ms   169.8 MB/sec
   string_dictionary_low_cardinality_400/cdc                       1.00     
25.9±0.07ms   154.9 MB/sec    1.00     26.0±0.41ms   154.2 MB/sec
   string_dictionary_low_cardinality_400/default                   1.01     
23.7±0.07ms   169.6 MB/sec    1.00     23.5±0.06ms   170.6 MB/sec
   string_dictionary_low_cardinality_400/number_distinct_values    1.00     
47.9±0.09ms    83.7 MB/sec    1.00     47.9±0.11ms    83.8 MB/sec
   string_dictionary_low_cardinality_400/parquet_2                 1.01     
23.7±0.06ms   169.1 MB/sec    1.00     23.6±0.08ms   170.1 MB/sec
   string_dictionary_low_cardinality_400/zstd                      1.00     
24.0±0.08ms   167.0 MB/sec    1.00     24.1±0.06ms   166.8 MB/sec
   string_dictionary_low_cardinality_400/zstd_parquet_2            1.00     
24.1±0.05ms   166.2 MB/sec    1.00     24.1±0.07ms   166.3 MB/sec
   ```


-- 
This is an automated message from the Apache Git Service.
To respond to the message, please log on to GitHub and use the
URL above to go to the specific comment.

To unsubscribe, e-mail: [email protected]

For queries about this service, please contact Infrastructure at:
[email protected]

Reply via email to