Skip to content

Reading the CubicalRipser benchmark data

The CubicalRipser comparison measures barcodes from images and volumes. Its data contain all 48 requests, including the smaller sizes behind the report's scaling plots. Each request fixes an array, a way to turn its values into a growing cubical complex, and the direction in which those values are traversed.

Choose a file

Start with results.csv, a spreadsheet table with one row per request. The JSON files preserve nested records, including the repeated measurements behind each row and the intervals of each barcode.

FileWhat you can learn from it
results.csvArray settings, output counts, times and speed ratios for each request.
analysis.jsonThe pass summaries under records, with combined results under groups and balanced_ratio.
samples.jsonIndividual accepted measurements, including compilation and input-reset counters.
formal_manifest.jsonArray shapes, random seeds, construction rules and input fingerprints.
verification.json and barcodes.jsonThe barcode comparisons and the intervals that were compared.

Read a row

The row 2d_V_noise_size1 uses a 4096 × 4096 array of noise values. Its V convention places array values at vertices; a square enters when all its vertices have entered. With order = sublevel, increasing the threshold admits vertices whose values are at most that threshold.

FieldValue and meaning
dimension, shape2, 4096x4096: A two dimensional square array.
elements16777216: the number of scalar values in the array.
tamerop_seconds3.982081403: TamerOp's summary time, about 3.98 seconds.
cubicalripser_seconds4.906119814…: CubicalRipser's summary time, about 4.91 seconds.
ratio1.25997…: the paired measurements give TamerOp about 1.26 times CubicalRipser's speed on this request.
intervals4437562: the number of barcode intervals, counting repeated intervals separately.

Ratios divide CubicalRipser time by TamerOp time, so values above 1 favor TamerOp. The ratio summarizes paired measurements; it need not equal the quotient of the two displayed summary times.

Understand the timings

Times are in seconds. The timer starts from the supplied array and includes building the cubical input, required conversions, negating values for a superlevel request, and producing the complete result. Loading the input file, checking its fingerprint and comparing barcodes happen outside the timer.

There are three passes, each using a fresh process for each tool and request. After warmup, each process measures three fresh computations. The middle of those times is its pass median; the middle of the three pass medians is the displayed tool time. In analysis.json, each request's tools.tamerop.passes and tools.cubicalripser.passes retain both the individual samples and their median, seconds.

The reported ratio is the geometric mean of the three paired ratios: multiply them and take the cube root. ratio_min and ratio_max in the CSV, or ratio_range in the JSON, give the observed extremes across passes. Each tool's range_seconds similarly gives the extremes of its pass medians. These are observed ranges rather than confidence intervals.

Identify and combine requests

FieldMeaning
case in CSV, id in JSONThe request identifier shared by the measurement and input records.
dimension, shape, elementsThe number of array axes, their lengths and the total number of scalar entries.
conventionV places values at vertices. T places them on the largest cells, such as squares or cubes; their faces enter with the earliest incident cell.
patternThe rule producing the values: noise, binary or smooth data.
ordersublevel grows from smaller values upward; superlevel grows from larger values downward.
reservedWhether the input was held back until the computation and measurement procedure had been fixed.
verifiedWhether its complete barcode agreed between the programs.
classification in CSV, similarity in JSONThe practical comparison label, using a 10% margin around equal speed.

The manifest supplies the random seed and exact generator for each request. analysis.json groups requests by dimension, convention and pattern. The overall balanced_ratio is the geometric mean of the 48 request ratios. Each of the twelve groups has four requests, giving the groups equal weight. The report explains how the two cubical conventions change the mathematical input.

Read a barcode

In barcodes.json, each interval record has four entries:

[degree, birth, death, multiplicity]

The degree identifies the kind of hole: 0 for components, 1 for loops and 2 for enclosed voids. Birth and death are the parameter values where the class appears and disappears. Multiplicity counts how many times that same interval occurs, allowing repeated intervals to be stored together.

For example, the illustrative row [1, 2, 5, 3] means three degree 1 intervals born at 2 and dying at 5. A null death denotes an essential interval, which continues indefinitely. Superlevel requests store grades for the negated scalar function, so negate those grades to recover thresholds in the original values.

Read the memory and verification records

The per-pass records in analysis.json use bytes for sizes and memory:

FieldMeaning
allocation_bytesMemory requested during the call, including temporary objects; null for CubicalRipser, where this counter was not collected.
input_bytes, result_bytesStorage measured for the input array and returned result.
process_peak_rss_bytes, query_peak_rss_bytesPeak recorded resident memory, or physical RAM use, over the process and query observation periods respectively.
import_seconds, warmupPackage loading and warmup measurements, kept separate from the accepted query times.
worker_wall_secondsTotal elapsed time for the worker, including preparation and repeated computations.

The machine record and build record identify the computer and software. Qualification checks describe independent controls, including failures outside the timed requests; calibration.json records the size trials used to select feasible inputs. The fixed task specification, public-call inspection, source fingerprints and study binding connect those definitions to the measured implementation.

The checksum list identifies the archived files by SHA-256 fingerprints. Follow Reproduce a benchmark report with the study name cubicalripser-v1 to verify those records, generate the arrays and collect new measurements.