Skip to content

Latest commit

 

History

History
144 lines (113 loc) · 6.69 KB

File metadata and controls

144 lines (113 loc) · 6.69 KB

BlackSTAR Performance

Performance claims apply only to the stated workload, host, binaries, thread count, storage placement, and cache definition. They are not universal guarantees.

Direct Official STAR Comparison

The cumulative qualification compared official STAR 2.7.11b at b1edc1208d91a53bf40ebae8669f71d50b994851 with released BlackSTAR 2.7.11b-blackstar.2 at d6fbf932ae2b155ce4f689bce106429ab2bc07f6.

The benchmark ran on dedicated Slurm node ca2, using Linux x86-64 GCC 13 binaries, 96 logical CPUs for the primary high-thread measurements, and node-local SSD. Every timing used seeded, order-balanced pairs.

Full CHM13 Index

CHM13v2 plus ERCC and the production CHM13/Ensembl/HPRC/ERCC GTF used three pairs:

Metric Official STAR BlackSTAR
Median wall time 1,256.03 s 607.18 s
Paired reduction 51.44%
Speedup 2.06x
Median peak RSS 52.78 GiB 77.15 GiB

All 42 substantive index-file comparisons passed byte-for-byte. The speedup therefore carries a material 46.18% memory cost that must remain visible in capacity planning.

Alignment Scaling

The public corpus contained 12,768,316 paired 76-base ENCODE reads and used a common GRCh38/Ensembl 114 index with GeneCounts, NoSharedMemory, and no alignment file.

Input Threads Official STAR BlackSTAR Reduction
Uncompressed 1 1,613.41 s 1,524.61 s 6.17%
Uncompressed 32 101.50 s 99.22 s 2.38%
Uncompressed 64 83.35 s 61.07 s 26.73%
Uncompressed 96 82.49 s 55.20 s 33.11%
zcat 1 1,633.61 s 1,522.24 s 6.68%
zcat 32 101.14 s 98.89 s 2.20%
zcat 64 84.96 s 64.88 s 23.30%
zcat 96 90.22 s 63.15 s 30.75%

The 1-thread control shows a modest serial gain. The larger 64- and 96-thread differences establish improved multicore efficiency. The approximately 2% 32-thread result prevents a general claim that BlackSTAR is 30% faster at all thread counts.

All 24 mapping-pair comparisons passed. A separate 2-million-pair unsorted-BAM check passed canonical record identity. Median peak RSS at 96 threads was approximately 5.7% lower for BlackSTAR.

Named-Sequence Addition

The test added public GFP and glutathione S-transferase FASTA records plus insert-only GTF annotations to GRCh38/Ensembl 114.

Input-file cache Official full rebuild BlackSTAR Delta Speedup
Warm 1,172.64 s 23.21 s 50.52x
Verified cold 1,225.62 s 39.04 s 31.50x

All six cold gates measured zero resident input-file pages before timing. This controls file-page residency, not powered-off hardware; filesystem metadata, libraries, and device caches can remain warm.

All 57 cold-versus-warm artifact checks passed. Full and Delta alignment matched timing-independent metrics, junctions, gene counts, and canonical BAM records, with exactly 100 GFP and 100 GST fragments counted in each mode.

BlackSTAR 1.1.0 Cross-Workload Qualification

Q02 compared official STAR 2.7.11b with the BlackSTAR 1.1.0 hardening source across common modes beyond paired gene-count-only alignment. These are cumulative workload results, not incremental effects of the 1.1.0 hardening changes alone; most reflect the complete BlackSTAR runtime stack.

Workload Official STAR BlackSTAR Median paired change 95% interval Qualification
Paired 76-base fragmented 72.02 s 58.02 s +20.05% +16.32 to +23.96% Compatibility pass
Paired 150-base fragmented 89.77 s 78.66 s +12.67% +11.38 to +13.46% All gates pass
Single-end 150-base 61.52 s 58.95 s +1.67% -1.87 to +13.42% CV gate failed; no speed claim
Two-pass Basic 223.91 s 177.05 s +20.25% +15.44 to +20.93% Compatibility pass
BySJout 91.36 s 62.16 s +31.45% +31.22 to +35.64% All gates pass
Chimeric detection 89.55 s 60.09 s +31.76% +30.98 to +34.00% All gates pass
Coordinate-sorted BAM 106.56 s 85.31 s +18.42% +16.93 to +24.26% Compatibility pass
Genomic plus transcriptome BAM 217.33 s 213.58 s +1.55% -0.45 to +2.63% Noninferior; no speed claim
STARsolo 10x v3 Gene 153.45 s 149.23 s +3.78% -1.43 to +10.49% Noninferior; no speed claim
STARlong direct RNA 72.79 s 58.61 s +19.48% +17.11 to +22.38% All gates pass with symmetric seed override

Positive change means less BlackSTAR wall time. Every one of the 36/36 mode-specific pair comparisons passed its applicable output oracle, including the pairs in the nonaccepted single-end timing series. Candidate peak RSS was lower in every row.

The paired 76-base, two-pass, and sorted-BAM control arms had CV between 3 and 5 percent. Their positive intervals are useful generalization evidence, but they do not satisfy the separate Labs preference for less than 3 percent arm CV for a new speed claim. STARsolo and TranscriptomeSAM establish only noninferiority.

The STARlong comparison used --seedPerReadNmax 100000 in both arms. Both official STARlong and BlackSTAR STARlong abort on this real direct-RNA fixture under the inherited default. Single-end performance also remains unresolved: the dedicated-host five-pair result passed correctness, RSS, and 2 percent noninferiority, but candidate CV was 5.22 percent.

See Q02 for exact binary identities, fixtures, resource results, negative evidence, and the TranscriptomeSAM determinism correction. Its bounded machine-readable receipts are tracked with an independent checksum manifest.

Evidence

Bounded machine-readable receipts are committed under:

  • docs/benchmarks/official-star-2.7.11b-vs-blackstar.2/ for the stable cumulative comparison; and
  • docs/benchmarks/Q02-cross-workload-20260726/ for the 1.1.0 cross-workload qualification source.

Large raw outputs remain outside Git. The receipts identify measured binaries and inputs and preserve the pair data, aggregate values, negative evidence, and gate decisions needed to audit the published claims. A retained absolute path inside a benchmark contract records the original ephemeral run environment; it is provenance, not a promise that the path exists in a fresh clone.

Claim Rules

  • Do not add incremental experiment percentages to cumulative results.
  • Do not generalize high-thread results to all thread counts.
  • Report memory changes alongside wall-time changes.
  • State local, network, cold, and warm storage conditions explicitly.
  • Require output-equivalence gates before accepting a speed result.
  • Requalify after performance-sensitive code, compiler, dependency, or host topology changes.