メニュー
Run and interpret Comet performance benchmarks comparing batch search vs real-time search (RTS) throughput. Use when measuring search speed, comparing before/after changes, or analyzing Hz and ms/spectrum metrics.
Codex または Claude でインストール この Prompt をコピーして Codex、Claude、または他のアシスタントに貼り付けると、Skill ページを確認してインストールできます。
Build the Comet mass spectrometry search engine from source. Use when building, rebuilding, or diagnosing compile/link errors in the Comet project (Comet.sln).
Navigate the Comet codebase. Use when looking for where a feature is implemented, which file owns a function, how data flows between layers, or what global variables are involved in a code path.
Debug crashes, silent exits, or wrong results in Comet. Use when a search dies unexpectedly, produces no output, crashes during index build or search, or gives suspicious results.
Run Comet batch or real-time searches, interpret output files, and understand search parameters. Use when executing searches, reading results, or configuring search settings.
SOC 職業分類に基づく
| name | comet-benchmark |
| description | Run and interpret Comet performance benchmarks comparing batch search vs real-time search (RTS) throughput. Use when measuring search speed, comparing before/after changes, or analyzing Hz and ms/spectrum metrics. |
RealtimeSearch\bin\x64\Release\)20170103_HelaQC_01.raw (39,970 spectra)human.target-decoy.fasta.idx (410 MB, ~40,858 proteins)cd /mnt/c/Work/Comet-master/RealtimeSearch/bin/x64/Release
../../../x64/Release/Comet.exe -Dhuman.target-decoy.fasta.idx 20170103_HelaQC_01.raw
Key output line: searching "20170103_HelaQC_01" ... 10s (39970 spectra, 0.25ms/spec, 3930Hz)
cd /mnt/c/Work/Comet-master/RealtimeSearch/bin/x64/Release
./RealtimeSearch.exe 20170103_HelaQC_01.raw 20170103_HelaQC_01.raw human.target-decoy.fasta.idx 20
(Or run via PowerShell from that directory.)
Key output lines:
Pre-read N scans in X.XX s
initialize search elapsed time: XX.XX s
search elapsed time: X.XX s, avg Y.YY ms/spectrum (39970 spectra), ZZZZ Hz
| Mode | Time | ms/spec | Hz |
|---|---|---|---|
| Batch (Comet.exe) | 10 s | 0.25 | 3930 |
| RTS original | 33.2 s | 0.83 | 1204 |
| RTS after pre-read fix | ~4 s search | ~0.10 | ~10000 |
search elapsed time = wall-clock time for the parallel search phase only (excludes pre-read and index load).
avg ms/spectrum and Hz are computed from cumulativeElapsedMS2 — the sum of per-call wall time inside DoSingleSpectrumSearchMultiResults only, divided by scansProcessedMS2. This measures pure C++ search speed, excluding raw file I/O and marshaling overhead. It is a single-thread-equivalent rate, not wall-clock throughput.
Wall-clock throughput = scansProcessedMS2 / watchGlobal.Elapsed.TotalSeconds (not printed directly).
The original RTS bottleneck was 20 C# threads sharing one IRawDataPlus rawFile. Thermo's reader serializes all calls internally, so ~87% of thread time was idle waiting for file I/O. Fix: pre-read all scan data sequentially before launching parallel workers (PrereadScan queue). Workers then operate entirely from in-memory arrays.
This is InitializeSingleSpectrumSearch() loading the .idx file and building g_pvProteinNameCache. It is a fixed one-time cost, not part of the search phase timer.