diff options
| author | Pjotr <wrk@bosa.lan> | 2026-11-11 13:21:05 +0000 |
|---|---|---|
| committer | Pjotr <wrk@bosa.lan> | 2026-11-11 13:21:05 +0000 |
| commit | ca7efa231bbd818b5166e38b59e3c81ea30821b3 (patch) | |
| tree | ca02a03ebbf576afc1ea45ff3e251bd1c967056b /src | |
| parent | f4d2d719f90774e85bc47eca041bc469f0f0951d (diff) | |
| download | genenetwork4-ca7efa231bbd818b5166e38b59e3c81ea30821b3.tar.gz | |
bulklmm: BULKLMM_BLOCKS as a config option, default 4
The marker blocks a single-trait scan is split into were the driver's
default (Threads.nthreads()) and changed with --threads. They are a
separate knob now:
BULKLMM_THREADS Julia threads (--threads) default 4
BULKLMM_BLOCKS independent marker blocks, scanned default 4
in parallel (1 = serial scan)
Blocks are exact either way -- LOD 9.589938 and the same permutation
thresholds at 1 and 4 -- so the knob only buys wall clock. The reply says
what ran ("split into 4 independent blocks and scanned on 4 Julia
thread(s) (BULKLMM_BLOCKS / BULKLMM_THREADS ...)"), because the answer to
"is it multicore?" should be in the response and not in the log.
Diffstat (limited to 'src')
| -rw-r--r-- | src/bulklmm.rs | 4 | ||||
| -rw-r--r-- | src/config.rs | 11 | ||||
| -rw-r--r-- | src/routes/bulklmm.rs | 13 |
3 files changed, 26 insertions, 2 deletions
diff --git a/src/bulklmm.rs b/src/bulklmm.rs index 46f6102..0b167d9 100644 --- a/src/bulklmm.rs +++ b/src/bulklmm.rs @@ -189,6 +189,7 @@ pub async fn compute( reml: bool, assumption: &str, threads: u32, + blocks: u32, ) -> Result<Output, BulklmmError> { let unique = format!( "{}-{}", @@ -201,6 +202,7 @@ pub async fn compute( let result_path = work_dir.join(format!("{unique}-bulklmm.csv")); let perm_path = work_dir.join(format!("{unique}-bulklmm-perm.csv")); let threads_str = threads.to_string(); + let blocks_str = blocks.to_string(); let threads_for_env = threads.to_string(); let nperm_str = nperm.to_string(); let seed_str = seed.to_string(); @@ -223,6 +225,8 @@ pub async fn compute( result_path.to_string_lossy().into_owned(), "--assumption".to_string(), assumption.to_string(), + "--blocks".to_string(), + blocks_str, "--seed".to_string(), seed_str, ]; diff --git a/src/config.rs b/src/config.rs index 42b74e7..cf5476e 100644 --- a/src/config.rs +++ b/src/config.rs @@ -158,6 +158,10 @@ pub struct Config { pub bulklmm_depot: String, pub bulklmm_script: String, pub bulklmm_threads: u32, + /// Independent marker blocks scanned in parallel (`BULKLMM_BLOCKS`, + /// default 4). `1` is a plain serial scan; the LODs and thresholds are + /// identical either way. + pub bulklmm_blocks: u32, pub bulklmm_cache_dir: String, /// The script that draws a BulkLMM scan (`BULKLMM_PLOTTER`). pub bulklmm_plotter: String, @@ -341,6 +345,10 @@ impl Config { "scripts/bulklmm_scan.jl".to_string(), ); values.insert("BULKLMM_THREADS".to_string(), "4".to_string()); + // Independent marker blocks scanned in parallel: a BulkLMM `scan()` is + // serial in Julia, so this is where a single-trait scan gets its + // parallelism. 1 is a plain serial scan and gives the same answer. + values.insert("BULKLMM_BLOCKS".to_string(), "4".to_string()); values.insert( "BULKLMM_PLOTTER".to_string(), "scripts/plot-bulklmm.py".to_string(), @@ -425,6 +433,7 @@ impl Config { "BULKLMM_DEPOT", "BULKLMM_SCRIPT", "BULKLMM_THREADS", + "BULKLMM_BLOCKS", "BULKLMM_CACHE_DIR", "BULKLMM_PLOTTER", ] { @@ -478,6 +487,7 @@ impl Config { bulklmm_depot: values["BULKLMM_DEPOT"].clone(), bulklmm_script: values["BULKLMM_SCRIPT"].clone(), bulklmm_threads: values["BULKLMM_THREADS"].parse().unwrap_or(4), + bulklmm_blocks: values["BULKLMM_BLOCKS"].parse().unwrap_or(4), bulklmm_cache_dir: values["BULKLMM_CACHE_DIR"].clone(), bulklmm_plotter: values["BULKLMM_PLOTTER"].clone(), genotype_files_dir: values["GENOTYPE_FILES"].clone(), @@ -734,6 +744,7 @@ mod tests { "BULKLMM_DEPOT", "BULKLMM_SCRIPT", "BULKLMM_THREADS", + "BULKLMM_BLOCKS", "BULKLMM_CACHE_DIR", "BULKLMM_PLOTTER", ] { diff --git a/src/routes/bulklmm.rs b/src/routes/bulklmm.rs index f244e5d..098ddb3 100644 --- a/src/routes/bulklmm.rs +++ b/src/routes/bulklmm.rs @@ -73,7 +73,8 @@ pub struct BulklmmRequest { /// `null` (default: variance components estimated once, under the null) or /// `alt` (per marker, slower and more accurate). pub assumption: Option<String>, - /// Cores for Julia (`--threads`), default from `BULKLMM_THREADS`. + /// Julia threads (`--threads`), default `BULKLMM_THREADS` (4). The marker + /// blocks scanned in parallel come from `BULKLMM_BLOCKS` (4). pub cores: Option<u32>, /// Normalisation of the trait values, as on `/compute/rqtl`. pub normalize: Option<String>, @@ -145,6 +146,7 @@ pub async fn compute_bulklmm( let seed = request.seed.unwrap_or(0); let reml = request.reml.unwrap_or(false); let threads = request.cores.unwrap_or(config.bulklmm_threads).clamp(1, 32); + let blocks = config.bulklmm_blocks.clamp(1, 64); let mut notes: Vec<String> = Vec::new(); // A basket is another source for `samples`. @@ -514,6 +516,7 @@ pub async fn compute_bulklmm( reml, &assumption, threads, + blocks, ) .await .map_err(|error| match error { @@ -587,7 +590,13 @@ pub async fn compute_bulklmm( "Rows are positional: marker, chr, pos (Mb), lod, beta, h2. BulkLMM fits one LMM per \ marker with a global kinship matrix (not GEMMA's leave-one-chromosome-out), by ML \ or REML, and reports the effect size `beta` and the null heritability `h2`.", - ); + ) + .with_note(&format!( + "Parallelism: BulkLMM's single-trait scan() is serial in Julia, so the markers are \ + split into {blocks} independent blocks and scanned on {threads} Julia thread(s) \ + (BULKLMM_BLOCKS / BULKLMM_THREADS; blocks 1 is a plain serial scan and gives the \ + same LODs). Where the cores really pay is many traits at once." + )); if let Some(group) = request.group.clone() { envelope = envelope.with_next( &format!("/groups/{}", path_segment(&group)), |
