Skip to content

Reading the C++ Ripser benchmark data

The C++ Ripser comparison follows a request from a distance file to its complete barcode. Its data contain ten requests, together with their repeated timings and answer checks. Each row lets us see both how long the request took and how much that comparison varied between runs.

Choose a file

Start with summary.csv, a table with one row per request that you can open in a spreadsheet. The JSON files preserve the measurements and input definitions behind it.

FileWhat you can learn from it
summary.csvEach input's size, times, speed ratio and memory measurements.
summary.jsonThe same results with pass times, input filenames and interval counts by degree.
observations.jsonIndividual observations, including warmups and measurements rejected because they included compilation.
manifest.json and fixtures.jsonThe fixed request settings, timing protocol and input constructions.
verification.jsonThe comparison of complete barcodes.

Read a row

The row circle_1280_f2 uses distances between 1,280 points on a circle. It computes homology through degree 1 with coefficients modulo 2: connected components in degree 0 and loops in degree 1.

FieldValue and meaning
n, edges1280, 818560: the number of vertices and supplied edges.
tamerop_seconds15.527778…: TamerOp's summary time for the full request.
cpp_seconds28.592399…: C++ Ripser's summary time for the full request.
ratio1.83084…: the paired runs give TamerOp about 1.83 times C++ Ripser's speed on this request.
ratio_min, ratio_max1.72793…, 1.92879…: the smallest and largest paired ratios observed across the three passes.
intervals1281: the barcode contains 1,280 degree 0 intervals and one degree 1 interval.

Ratios divide C++ Ripser time by TamerOp time. Values above 1 favor TamerOp; values below 1 favor C++ Ripser. The displayed ratio summarizes the paired runs, so it need not equal the quotient of the two displayed times.

Understand the timings

All times are in seconds. The main timer includes reading the distance file, building the program's own input objects, computing the barcode and emitting the full answer. C++ Ripser launches a new executable for each call, so its launch time is included. TamerOp runs in an already loaded Julia process with compiled methods.

The study uses three passes, or rounds of measurements. Each case starts a fresh Julia process in each pass. A discarded warmup precedes two accepted samples per tool; their median is the average of the two times. The displayed tool time is the middle of its three pass medians. The JSON fields tamerop_pass_seconds and cpp_pass_seconds retain those medians.

The reported ratio is the geometric mean of the three paired C++/TamerOp ratios: multiply them and take the cube root. ratio_min and ratio_max are the observed extremes, rather than confidence intervals. construction_seconds and query_seconds describe parts of TamerOp's main timer. Their separate summaries need not add to its total, which also includes file reading and answer output.

Identify the request and its answer

FieldMeaning
id, familyThe request identifier and its input family.
n, prime, degreeThe vertex count, coefficient prime and highest requested homology degree.
formatIn JSON, how distances are stored, such as a lower triangular distance table.
thresholdIn JSON, the distance cutoff; null means no finite cutoff was supplied for that request.
input, sha256In JSON, the input filename and the SHA-256 fingerprint identifying its bytes.
interval_countsIn JSON, interval counts in degree order, beginning with degree 0. Their sum is intervals.

The barcode checks compare births, deaths and repeated intervals by degree. Small control examples additionally compare against an independent boundary-matrix calculation. The sparse-cutoff record and its four-vertex input describe a separate input on which C++ Ripser's behavior disagreed with the intended finite-cutoff answer. That control is excluded from the timing table.

Read the memory and verification records

All memory fields use bytes:

FieldMeaning
allocation_bytesMemory requested during TamerOp's measured call, including temporary objects.
retained_result_bytesThe size of TamerOp's returned barcode object.
tamerop_peak_rss_bytesThe largest recorded resident memory, or physical RAM use, for the case's Julia process. This includes loading, compilation, warmup and repeated samples of that case.
cpp_peak_rss_bytesThe peak observed by sampling C++ invocations separately.

These process peaks include more than the returned barcode, and the two programs' process lifetimes differ. The machine and build record supplies the software and memory settings. The configuration record explains the changes that preceded the reported runs; default-heap observations and shared-session observations retain those excluded attempts. The recovery record identifies resumed work.

Source fingerprints, measurement-script fingerprints and the checksum list identify the exact code and archived files. Follow Reproduce a benchmark report with the study name ripser-cpp-v1 to check them, regenerate the distance files and collect new measurements.