[FEA] Add an alternative method of reading from files that uses `mmap`

[vuule]

Memory mapping could be faster in some use cases and/or some systems.
This is a proposal to add an option to replace file reads with copying from the mapped-to buffer.
Some open (AFAIK) design decisions:

• How should this interact with compatibility mode options.
• Should this be used for reads to device memory, or when.

[vuule]

this is addressed now

[kingcrimsontianyu]

Should this issue be reopened?

[kingcrimsontianyu]

Related PRs

Main features

• C++ implementation: Support file-backed mapping (1/n): host/device read C++ implementation #740
• Python binding: Support file-backed mapping (2/n): host/device read Python binding #742

Improvements and fixes

• Example: Support file-backed mapping (3/n): host side read examples #743
• Benchmark: TBA
• Relax mmap read requirement. Improve error message. #781
• Fix C++20 warning in the mmap test #785

Benchmark results

GH200 (C2C)

• SAMSUNG MZ1L21T9HCLS-00A07 PCIe Gen 4.0 x4

t1: PDS all queries

• t1.1: Hot page cache

  • Each query is repeated 11 times. The average of the elapsed time of the last 10 repetitions is used. Same for the case of cold page cache.

  

• t1.2: Cold page cache

  

t2: PDF Q5 parameter sweep (number of threads, task size)

• Query 5 at a certain (number of threads, task size) coordinate is repeated 5 times. THe average of the elapsed time of the last 4 repetitions is used.

t3: cuDF benchmark

• t3.1: Parquet

  PARQUET_READER_NVBENCH -d 0 -b parquet_read_io_compression -a compression_type=NONE -a io_type=FILEPATH -a cardinality=0 -a run_length=1 --min-samples 80 --timeout 60 --csv stdout
  

  	Standard I/O (s)	Mapped I/O (s)
  Hot page cache	0.02365 (0.5381%)	0.02318 (2.163%)
  Cold page cache	0.1651 (5.02%)	0.1672 (2.143%)

• t3.2: ORC

  ORC_READER_NVBENCH -d 0 -b orc_read_io_compression -a compression=NONE -a io=FILEPATH -a cardinality=0 -a run_length=1 --min-samples 80 --timeout 60 --csv stdout
  

  	Standard I/O (s)	Mapped I/O (s)
  Hot page cache	0.03901 (0.2444%)	0.03688 (1.46%)
  Cold page cache	0.1803 (4.862%)	0.1761 (4.605%)

• t3.3: Json

  JSON_READER_NVBENCH -d 0 -b json_read_io -a io=FILEPATH --min-samples 80 --timeout 60 --csv stdout
  

  	Standard I/O (s)	Mapped I/O (s)
  Hot page cache	0.1164 (0.09435%)	0.1136 (0.6991%)
  Cold page cache	0.4927 (4.263%)	0.4962 (4.149%)

• t3.4: CSV

  CSV_READER_NVBENCH -d 0 -b csv_read_io -a io=FILEPATH --min-samples 80 --timeout 60 --csv stdout
  

  	Standard I/O (s)	Mapped I/O (s)
  Hot page cache	0.05214 (0.2064%)	0.05125 (1.113%)
  Cold page cache	0.2672 (5.455%)	0.1804 (0.2069%) ($${\color{green}verified}$$)

A100 (PCIe gen 4)

• SAMSUNG MZQLB3T8HALS-00007 PCIe Gen 3.0 x4

t4: PDS all queries

⚠Currently, on A100, query 1 has an out-of-memory error.

• t4.1: Hot page cache

  

• t4.2: Cold page cache

  

t5: PDF Q5 parameter sweep (number of threads, task size)

t6: cuDF benchmark

• t6.1: Parquet

  	Standard I/O (s)	Mapped I/O (s)
  Hot page cache	0.0678 (1.898%)	0.07598 (3.88%)
  Cold page cache	0.1219 (2.507%)	0.1421 (4.498%)

• t6.2: ORC

  	Standard I/O (s)	Mapped I/O (s)
  Hot page cache	0.07965 (1.619%)	0.09319 (2.965%)
  Cold page cache	0.1688 (3.663%)	0.2401 (5.334%)

• t6.3: Json

  	Standard I/O (s)	Mapped I/O (s)
  Hot page cache	0.273 (2.017%)	0.2862 (1.744%)
  Cold page cache	0.4278 (2.052%)	0.4691 (7.523%)

• t6.4: CSV

  	Standard I/O (s)	Mapped I/O (s)
  Hot page cache	0.1234 (2.335%)	0.1448 (3.236%)
  Cold page cache	0.2477 (1.796%)	0.2853 (1.025%)