Improve parquet reading performance ~35-40%

[ritchie46]

Previous implementation we use memmap and read to a vector. This is not very smart, as we pay double IO. Removing memmap and reading directly to the buffers improved performance by ~30%.

But that data still needs to be zeroed before written, so if we use memmap directly we could skip that step. That improved performance by yet another ~8% compared to master.

Reading the yellowtrip dataset benchmarks:

master

Performance counter stats for './target/release/memcheck':

         22.978,03 msec task-clock                #    6,296 CPUs utilized          
           465.290      context-switches          #   20,249 K/sec                  
            10.129      cpu-migrations            #  440,812 /sec                   
         2.323.129      page-faults               #  101,102 K/sec                  
    75.015.799.667      cycles                    #    3,265 GHz                    
    73.372.673.203      instructions              #    0,98  insn per cycle         
    13.989.811.069      branches                  #  608,834 M/sec                  
        71.684.372      branch-misses             #    0,51% of all branches        

       3,649539611 seconds time elapsed

      10,707208000 seconds user
      13,337545000 seconds sys

read with zeroed

 Performance counter stats for './target/release/memcheck':

         17.159,57 msec task-clock                #    6,073 CPUs utilized          
           393.657      context-switches          #   22,941 K/sec                  
             7.877      cpu-migrations            #  459,044 /sec                   
         1.995.815      page-faults               #  116,309 K/sec                  
    59.192.854.862      cycles                    #    3,450 GHz                    
    52.386.441.413      instructions              #    0,89  insn per cycle         
    10.072.386.163      branches                  #  586,983 M/sec                  
        69.491.528      branch-misses             #    0,69% of all branches        

       2,825714164 seconds time elapsed

       5,988224000 seconds user
      12,028295000 seconds sys

memory map

 Performance counter stats for './target/release/memcheck':

         16.007,82 msec task-clock                #    6,071 CPUs utilized          
           352.604      context-switches          #   22,027 K/sec                  
             6.504      cpu-migrations            #  406,301 /sec                   
         1.747.134      page-faults               #  109,143 K/sec                  
    55.113.085.851      cycles                    #    3,443 GHz                    
    51.144.360.047      instructions              #    0,93  insn per cycle         
     9.883.151.274      branches                  #  617,395 M/sec                  
        67.085.974      branch-misses             #    0,68% of all branches        

       2,636863878 seconds time elapsed

       6,678761000 seconds user
      10,102559000 seconds sys

@jorgecarleitao FYI

[codecov-commenter]

Codecov Report

Merging #3821 (6796be3) into master (ca28e70) will increase coverage by 0.00%.
The diff coverage is 97.91%.

@@           Coverage Diff           @@
##           master    #3821   +/-   ##
=======================================
  Coverage   77.95%   77.95%           
=======================================
  Files         446      447    +1     
  Lines       73865    73946   +81     
=======================================
+ Hits        57578    57645   +67     
- Misses      16287    16301   +14     

Impacted Files	Coverage Δ
polars/polars-io/src/parquet/mod.rs	93.54% <ø> (ø)
polars/polars-io/src/parquet/read.rs	98.90% <ø> (ø)
polars/polars-io/src/parquet/read_impl.rs	94.53% <95.34%> (-0.47%)	⬇️
polars/polars-io/src/parquet/mmap.rs	100.00% <100.00%> (ø)
...core/src/chunked_array/logical/categorical/from.rs	21.62% <0.00%> (-18.92%)	⬇️
...ars/polars-core/src/series/implementations/utf8.rs	71.53% <0.00%> (-1.13%)	⬇️
polars/polars-arrow/src/utils.rs	75.28% <0.00%> (-1.13%)	⬇️
polars/polars-time/src/series/mod.rs	56.00% <0.00%> (-0.67%)	⬇️
polars/polars-core/src/series/from.rs	83.12% <0.00%> (-0.63%)	⬇️
...olars/polars-core/src/frame/groupby/into_groups.rs	60.29% <0.00%> (-0.30%)	⬇️
... and 7 more

---

Continue to review full report at Codecov.

Legend - Click here to learn more
Δ = absolute <relative> (impact), ø = not affected, ? = missing data
Powered by Codecov. Last update ca28e70...6796be3. Read the comment docs.

[ritchie46]

@thobai this will certainly help ;)

[alexander-beedie]

That's quite a gain - especially given that it was hardly slow to start with :)

[ritchie46]

That's quite a gain - especially given that it was hardly slow to start with :)

Yeap, curious for the benchmarks. :)

[thobai]

@ritchie46 This sounds very promising! Did that already make it into the recent python version of polars (0.13.51)? I just tested with that version and I can unfortunately not really see a difference in performance for my datasets and queries. But that could also be because the majority of time is spent somewhere else and reading wasn't a bit fraction of the execution time.