Assessing high-performance lightweight compression formats for Geospatial computation

Beating main-memory bandwidths in geospatial pipelines with fast in-memory compression

Cambridge Sensor CDT MRes Project

Author: Omar Tanner (omsst2)

Report: here

Files

codecs/*: core codecs implementing the codecs/generic_codecs.h inteface

Main programs:

• test_comp.cpp: test codecs
• bench_comp.cpp: benchmark codecs
• bench_pipeline.cpp: benchmark geospatial pipelines

Additional files:

• codec_collection.h: bundled codecs
• test_remappings.cpp: verifies Morton remappings
• util.h, transformations.h, remappings.h: CPP utilities
• py/*: Python utilities
• sh/*: Shell performance-monitoring utilities

Setup (Basic)

We assume a Linux environment.

1. install these packages with apt-get (you might need more, debug appropriately)
   1. g++, g++-11, libgdal-dev, python3-gdal, liblz4-dev, libzstd-dev, zlib1g-dev, liblzma-dev
2. obtain the submodules in external and build them
3. make

Setup (Fusing Summing into Decompression)

1. re-build external/FastPFor and external/simdcomp from these forks:
   1. FastPFor: https://github.com/omarathon/FastPFor
   2. simdcomp: https://github.com/omarathon/simdcomp
2. replace codecs/custom_vec_logic_codecs.h with agg/custom_vec_logic_codecs.h (the new file contains the modification to the custom_rle_vecavx512 codec which fuses summing into decompression)
3. replace bench_pipeline.cpp with agg/bench_pipeline.cpp
4. make clean && make

Setup (Running on HPC)

1. source hpc/modules.sh
2. replace Makefile with hpc/Makefile (the new Makefile contains compiler modifications for the HPC)
3. make

Miscellanea

License: CC BY 4.0

Full repo/data/report on request