This folder contains HEP sample anayses in order to measure and compare the behavior of TTree and RNTuple. Every analysis processes a single file with event data -- either in TTree format or in RNTuple format -- and fills control histograms. Every sample analysis is implemented in a single binary that contains the analysis code in three flavors:
- Direct access with TTree, using GetBranchAddress() style code
- Direct access with RNTuple, using type-safe Views (similar to TTreeReader)
- RDataFrame style (TODO: make all analyses available as RDF)
There are corresponding data generation binaries (gen_...) to produce
the input files from publicly available master sources.
The samples are atlas, cms, h1, lhcb.
ATLAS: H --> gg 2020 open data (cf. tutorial df104)
Medium dense reading (~25%, 12/81 features), non-trivial analysis code.
Background channels are skipped for the sake of benchmarking.
7.8 million events, 76k selected events. EDM consists of (non-nested) std::vectors.
CMS: 2019 ttjet nanoAOD (cf. tutorial df102). Sparse reading (<1%, 6/1479 features). 1.6 million events, 141k selected events. EDM with collections
H1: ROOT's H1 standard analysis (cf. df101 tutorial). The data are concatenated 10 times for a larger running time. Medium dense reading (~10%, 16/152 features). 2.8 million events, 75k selected events. EDM with collections.
LHCB: LHCb run 1 open data (B mass spectrum). Dense reading (>75%, 18/26 features). 8.5 million events, 24k selected events. Fully flat EDM.
Use the Makefile to build the binaries and produce benchmark results and plots.
To build only the benchmarks, run make benchmarks.
The Makefile requires ROOT available in the environment. ROOT should be built using
cmake -Droot7=on -DCMAKE_BUILD_TYPE=relwithdebinfo
The input data can be generated or will be automatically downloaded from
https://root.cern/files/RNTuple
Each benchmark takes the same input parameters:
- `-i` input file, ending either in `.root` (tree) or `.ntuple` (ntuple)
- `-s` show the control plot
- `-p` show the tree/ntuple performance statistics
- `-r` run the benchmark with RDataFrame instead of hand-written event loop
- `-m` enable implicit multi-threading (paralle RNTuple page decompression, parallel RDF event loop)
- `-x` cluster bunch size; a value less than 1 will disable the cluster cache
The real-time timing uses std::chrono::steady_clock and starts with the second event (direct access) or with an artificial first filter (RDF).
In order to clear the file system page cache between benchmark runs, use the clear_page_cache utility.
It can be created with make clear_page_cache, which requires sudo privileges.
The clear_page_cache utility is not removed by make clean.
It works on Linux only.
make benchmarks
./atlas -mps -i gg_data~zstd.ntuple
./cms -mps -i ttjet_13tev_june2019~lzma.root # NOTE: the data file is not public. Ask us for further information.
./h1 -mps -i h1dstX10~zstd.ntuple
./lhcb -mps -i B2HHH~zstd.ntuple