==================== Performance overview ==================== A path under ``benchmarks/`` on this page names a record of the d-SEAMS 2.0 benchmark campaign. The records and the Snakemake rules that write them are in the private paper repository and the Hugging Face dataset ``d-SEAMS/dseams2-benchmark-records`` (``docs/records/README.md`` names the revision); the scripts under ``scripts/`` that time a run stay here. Four sittings sit next to each other. Every number on this page is from the frozen tables in ``docs/records``, rendered from the benchmark records dataset. A ratio is the threads=1 median divided by the threads=8 median. The script marks a row flat when that ratio is between 0.95 and 1.05. A PLUMED row in the release-candidate table was measured on one thread because ``PLUMED_NUM_THREADS`` was unset. The 884736-molecule sitting sets that variable, so its PLUMED q6 row is timed at the width the column names. .. table:: +----------------------------------------------+-------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------+ | file | what it times | +==============================================+===========================================================================================================================================================================================+ | ``benchmarks/rc3-sitting.csv`` | The release-candidate head-to-head, one thread and eight | +----------------------------------------------+-------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------+ | ``benchmarks/trace-threads.csv`` | The hydrate signature search at 1, 2, 4 and 8 threads | +----------------------------------------------+-------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------+ | ``benchmarks/paper-cells/sitting-large.csv`` | Ice Ic and a TIP4P/ice liquid at 884736 molecules, paired with freud, mdapy, vesin, pyscal3, PLUMED, OVITO and rings-statistics-matter at 1, 2, 4 and 8 threads, recorded at ``2ed49e87`` | +----------------------------------------------+-------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------+ | ``benchmarks/pool-audit.csv`` | The pool-sized kernels on ice Ic, retimed at ``1ce3be3b`` | +----------------------------------------------+-------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------+ | ``benchmarks/cutoff-rings.csv`` | The full cutoff ring call at 884736 molecules, linkcell against the cell list, recorded by ``cad887b2`` | +----------------------------------------------+-------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------+ The full tables are `the release-candidate sitting `_, `the thread controls `_, and `the literature reproduction <../howto/literature-reproduce.rst>`_. ``meson test docs_sitting``, ``docs_trace_threads``, ``docs_pool_audit``, and ``docs_sitting_large`` recompute the stored medians. The same operation name is not the same call across files. rc3 :math:`q_4` is the guest's local parameter. The pool-audit file splits local and cutoff. The large-cell file names the call in its ``detail`` column. The tables keep the operation string the CSV stored. The rings row in the large-cell file is the full cutoff call. It is ``seams_frame_rings``: the cutoff list and the ring search together. The rings kernel in the pool-audit file is ``ringNetwork`` alone, and that timer starts after the neighbour graph exists. The two numbers are not interchangeable. The large-cell ``detail`` column is the agreement check on the ice frame and on the TIP4P/ice liquid. q6 is compared with freud and with mdapy, and with pyscal3 and with PLUMED. On the liquid the freud gap is 1.525e-04, from hard-cutoff rounding, and the PLUMED gap is 1.679e-01, from the rational switch. CHILL+ on that liquid matches 830736/884736 molecules. The neighbour count is the directed pairs left by the cutoff list. d-SEAMS is compared with vesin and freud, and with pyscal3. OVITO coordination is the same pair sum. Both CHILL+ codes are compared on the cubic label. The classifier is compared with OVITO PTM restricted to the cubic and hexagonal diamond templates, where ice Ic is cubic diamond. The six-ring count is compared with rings-statistics-matter. PLUMED q6 in this file set ``PLUMED_NUM_THREADS``. It is not the one-thread rc3 row. .. include:: performance-overview-tables.rst The 884736-molecule sitting --------------------------- ``2ed49e87`` is the sitting of ice and the TIP4P/ice liquid. One warm-up, then the calls stored in ``n``, at 1, 2, 4 and 8 threads. Each width is its own process. The rivals are freud, mdapy, vesin, pyscal3 (neighbours and q6), PLUMED q6 with ``PLUMED_NUM_THREADS`` set to that width, OVITO (CHILL+, q4, q6, coordination, and PTM with the default templates and with only the diamond templates), and rings-statistics-matter. ``scripts/plot_sitting_large_time.py`` draws time against threads on a log axis, one panel per operation. The PDF is not stored. The tables recompute each median from the samples and carry the agreement notes. ``docs_table_csv`` checks the copied numbers. The rings row is still the full cutoff call. The separate ``cad887b2`` table times that call with the cell-list builder. .. include:: sitting-large-tables.rst