3.1.0

Release Notes

Faster builds that write far less to the working directory, with the same
graphs. On 149,998 Salmonella assemblies (k = 31), measured back to back on
one host against 3.0.3, with intermediate compression off (3.0.3 has none):

	16 threads	64 threads	written (t16)
uncolored	12:12 to 6:05 (-50%)	4:45 to 2:45 (-42%)	472 to 204 GB
colored	17:46 to 11:56 (-33%)	6:30 to 4:20 (-33%)	616 to 347 GB

Peak memory stays close to 3.0.3: within 0.2 GB in three of the four
configurations, and about 0.4 GB higher colored at 64 threads (18.5 against
18.1 GB). For comparison, C++ Cuttlefish 3 took 14:43 uncolored and 25:06
colored at 16 threads, writing 644 and 970 GB. Output matches 3.0.3: the same
unitigs, and matching color digests (a cyclic unitig may start at a different
position).

Speed

• The partition's minimizer scan is vectorized: AVX2 on x86-64, chosen at
  run time with an identical scalar fallback, and NEON on aarch64. It follows
  simd-minimizers (Groot Koerkamp and Martayan, SEA 2025). Partitioning is
  23-28% faster.
• Reference FASTA records made only of upper-case ACGT skip the per-byte
  fragment scan (SSE2 on x86-64, NEON on aarch64).
• The 32-base label packer gains a NEON form on aarch64.
• At cutoff 1, the reference default, local contraction keeps edges as
  presence bits in the vertex state. Vertex-table slots shrink from 24 to
  20 bytes colored and from 16 to 12 uncolored, making uncolored local
  contraction 11% faster at 150k.
• For k > 31 the vertex table no longer pads every slot to 32 bytes.
  Uncolored local contraction is 14% faster at k = 55, and read mode's peak
  memory falls from 7.1 to 6.1 GB.

Intermediate I/O

• Much less is written to the working directory:
  • partition buckets store each label once and reference it after that;
  • local contraction replays its color pass from memory instead of
    re-reading the bucket;
  • intermediate labels are stored at 2 bits per base;
  • the path-info, coordinate and edge records are narrower.
• --compress-intermediates auto|on|off lz4-compresses the local-unitig,
  coordinate and color streams. auto, the default, decides at startup from
  the storage under --work-dir:
  • network filesystems, rotational disks and FUSE mounts (with a warning)
    are compressed without timing;
  • other storage is timed for about a second, and compressed unless it
    keeps well ahead of the build.
    On a spinning disk, compression made colored builds 24% faster. On fast
    NVMe, forcing it on costs about 2%, and auto leaves it off at moderate
    thread counts. The bar rises with threads (up to about 4.7 GB/s from 64
    threads on), so near it the choice can go either way; it never changes the
    graph.
• cuttlefish probe -w DIR -t N shows that measurement and what auto
  would choose. It is built on a new crate, scratch-probe, supported on
  Linux and macOS.
• cuttlefish cleanup also removes files left by an interrupted probe, but
  skips any modified in the last minute.

Compatibility

• Final outputs are unchanged. The private intermediate formats change, so
  intermediates left by an interrupted 3.0.x build are not readable by 3.1.0.
• Super-k-mers are assigned to subgraphs by a new hash, so subgraph contents
  differ from earlier releases and from C++ Cuttlefish 3. The unitigs do not.
• Library users: BuildParams has a new public field,
  compress_intermediates, so code that builds it with a struct literal must
  set it (BuildParams::new does). Its default, Auto, makes the build entry
  points time the work directory once and print their choice to stderr; set
  On or Off to skip that.

Fixes

• A colored build whose super-k-mers all fall in one subgraph, and so has no
  discontinuity edges, wrote an empty FASTA. This affected only very small
  inputs.

Install cuttlefish-rs-cli 3.1.0

Install prebuilt binaries via shell script

curl --proto '=https' --tlsv1.2 -LsSf https://github.com/COMBINE-lab/cuttlefish/releases/download/v3.1.0/cuttlefish-rs-cli-installer.sh | sh

Download cuttlefish-rs-cli 3.1.0

File	Platform	Checksum
cuttlefish-rs-cli-aarch64-apple-darwin.tar.xz	Apple Silicon macOS	checksum
cuttlefish-rs-cli-x86_64-apple-darwin.tar.xz	Intel macOS	checksum
cuttlefish-rs-cli-aarch64-unknown-linux-gnu.tar.xz	ARM64 Linux	checksum
cuttlefish-rs-cli-x86_64-unknown-linux-gnu.tar.xz	x64 Linux	checksum