fix(bitmatrix): explicitly compute diagonal entries for self-similarity

The pairwise matrix functions now explicitly calculate and overwrite diagonal entries using `f(i,i)`, replacing previous implicit symmetric mirroring or default values. Documentation has been updated to clarify that diagonals represent self-comparison weights, ensuring accurate self-similarity calculations. Additionally, the obikmer crate version has been bumped to 1.1.34.

docmd/architecture/numa_partition_runner.md

@@ -162,14 +162,158 @@ A single `PartitionRunner` instance can be built once per command invocation
 and reused across multiple `run()` calls (e.g. `merge` runs
 `merge_partitions` then `pack_matrices`).
 
+## Known issue: CPU-only activation signal stalls on I/O-bound stages
+
+Observed on a real `filter` run (109 genomes, 256 partitions, 8×24-core NUMA):
+`rebuild` (CPU-bound — k-mer construction) scales cleanly from 9 to 43 active
+workers as `CpuSample::do_i_activate` (`obisys::lib.rs`) sees efficiency climb.
+`pack_matrices` (I/O-bound — reopens and recomposes per-genome column files
+into `.pbmx`/`.pcmx`) activates one extra worker then flatlines at 10/192 for
+the rest of the stage, even though 256 partitions keep completing over several
+minutes. This matches the documented intent (§ Adaptive mechanism — "avoids
+over-provisioning ... I/O-bound ... workloads") but conflates two different
+things: *"CPU is not the bottleneck"* and *"more workers would not help"*. On
+storage with real queue depth (NVMe, RAID, parallel FS) the second stage could
+still benefit from more concurrent workers even with flat CPU usage — a signal
+the current mechanism cannot see.
+
+A one-off artefact was also found in the same log: right after a stage
+transition, `do_i_activate` produced a physically impossible spike (efficiency
+~94 cores on a 192-core box) because it has no minimum-window guard — unlike
+its sibling `cpu_efficiency`, which returns `0.0` if `wall < 0.1s`
+(`obisys::lib.rs:260`). `do_i_activate` unconditionally overwrites
+`self.wall`/`self.user_secs`/`self.sys_secs` even when the elapsed window is
+too short to be meaningful, so a burst of rapid completions right after
+activating a worker can divide a real CPU delta by a near-zero wall delta.
+
+### Implemented: I/O signal + shared debounce guard
+
+`IoSample` (`obisys::lib.rs`, alongside `CpuSample`) is fed by
+`read_bytes`/`write_bytes` from `/proc/self/io` on Linux (actual bytes
+submitted to the block layer — not `rchar`/`wchar`, which also count
+page-cache hits, and not `ru_inblock`/`ru_oublock`, unreliable on macOS), with
+a `proc_pid_rusage(RUSAGE_INFO_V4)` fallback on macOS
+(`ri_diskio_bytesread`/`ri_diskio_byteswritten`, FFI only via `libc`, no new
+dependency — same pattern as the existing `getrusage` bindings). Any other
+target degrades gracefully to a signal that never triggers (falls back to
+CPU-only activation), same pattern as `cgroup_v2_available`.
+
+`maybe_activate` (`numa.rs`) activates a worker if *either* signal still shows
+headroom, making `PartitionRunner` adapt to whichever resource is actually the
+bottleneck without per-call configuration. Both samplers are called
+unconditionally — no `||` short-circuit — so neither window starves behind
+whichever signal fires first:
+
+```rust
+let cpu_threshold = CPU_SPAWN_THRESHOLD * activation.last_step() as f64;
+let cpu_wants_more = cpu_sample.do_i_activate(cpu_threshold);
+let io_wants_more  = io_sample.do_i_activate(IO_SPAWN_THRESHOLD);
+if cpu_wants_more || io_wants_more {
+    activation.grow(GROWTH_DIVISOR, n_total);
+}
+```
+
+The CPU threshold is *not* the flat absolute delta it started as: it scales
+with `activation.last_step()` — the number of workers activated in the last
+growth step, tracked by `NodeActivation` (`numa.rs`) and updated every time
+`grow()` actually grows something. Growing by 8 workers should add ~8 cores of
+efficiency if the workload is truly CPU-bound; requiring only
+`CPU_SPAWN_THRESHOLD` (20 %) of that expected gain confirms the growth was
+useful without demanding perfect linear scaling. Scaling by the *last step's
+size* rather than the cumulative total keeps the bar equally meaningful
+whether it's the 2nd growth step or the 20th — a flat absolute threshold
+(0.2 core) is a strong signal at 8 active workers but pure noise at 150; a
+threshold scaled by the *cumulative* total instead (considered and rejected)
+would have made the bar essentially impossible to clear late in the ramp,
+strangling exactly the CPU-bound saturation the mechanism exists to allow.
+
+Unlike the CPU signal (an absolute delta in cores — a bounded, portable unit),
+raw I/O throughput has no natural scale across devices, so `IoSample` uses a
+**relative** growth threshold instead of an absolute one:
+
+```rust
+pub fn do_i_activate(&mut self, threshold: f64) -> bool {
+    let elapsed = self.wall.elapsed().as_secs_f64();
+    if elapsed < 0.1 { return false; }        // state untouched — window keeps accumulating
+
+    let n = Self::read_bytes();
+    let rate = n.saturating_sub(self.bytes) as f64 / elapsed;
+    let activate = if self.previous_rate == 0.0 {
+        rate > 0.0                            // bootstrap: any measured throughput is signal
+    } else {
+        (rate - self.previous_rate) / self.previous_rate >= threshold
+    };
+
+    self.bytes = n;
+    self.wall  = Instant::now();              // reset only on a real sample
+    activate
+}
+```
+
+The `elapsed < 0.1s → return false without mutating state` guard was also
+back-ported into `CpuSample::do_i_activate` (previously missing — source of
+the ~94-core artefact above) — one fix for both problems, and it removes the
+need for any arbitrary I/O-rate floor: a short/noisy window is rejected
+outright rather than papered over with a hardware-dependent constant.
+
+Both spawn thresholds (`CPU_SPAWN_THRESHOLD`, `IO_SPAWN_THRESHOLD`, module-level
+`const` in `numa.rs`, both `0.2`) are a starting point, not a derived value:
+`0.2` (20 % relative growth) for `IoSample` was chosen to match the CPU
+threshold's *implicit* relative sensitivity (in the observed log, an 8→9
+worker step raised efficiency by ~12 %) — but I/O throughput is lumpier than
+CPU time (buffered writes flush in bursts), so it needs empirical validation
+against a real `pack` run before being considered final.
+
+## Known issue: ramp-up too slow, and confused with node count
+
+The original design started `n_nodes` workers (one per node) and grew one
+worker at a time. On a real `filter` run this took ~10 minutes to climb from
+9 to ~40 active workers even on the CPU-bound `rebuild` stage — most of a
+35-minute stage spent under-provisioned while waiting for evidence to
+accumulate one worker at a time. There is no scale-down mechanism (`n_active`
+only grows), so the original caution was deliberate — but a quarter of
+available cores is still far from saturation, and the real risk zone (over-provisioning
+a memory-bandwidth-bound stage) only shows up much later in the ramp, near
+full occupancy — not at 25 %.
+
+The fix decouples ramp speed from node *count*: both the initial size and the
+growth step are a fraction of `workers_per_node` (node *size*), applied
+identically on every node. A single-NUMA-node (UMA) machine ramps exactly as
+fast as an 8-node one — growing by `n_nodes` per step, as first considered,
+would have degenerated to "grow by 1" on UMA, reproducing the original
+problem for exactly the machines that need the fix most.
+
+```rust
+// NodeActivation::grow — called both at startup (activate_initial) and on
+// every CPU/IO-triggered growth step, with a different divisor each time.
+let wanted = (self.caps[idx] / divisor).max(1);   // INITIAL_DIVISOR=4 at startup, GROWTH_DIVISOR=8 per step
+let room   = self.caps[idx].saturating_sub(self.active[idx]);
+let grow   = wanted.min(room).min(n_total.saturating_sub(self.total));
+```
+
+This also fixed a latent correctness gap: the original single shared
+`activate_tx`/`activate_rx` pair had *no* per-node addressing — sending one
+activation signal woke up whichever dormant worker (from any node) happened
+to win the race on that channel. `crossbeam_channel` gives no fairness
+guarantee across competing receivers, so "round-robin across nodes" was an
+assumption the code never actually enforced. `PartitionRunner::run` now opens
+one activation channel per node (`activate_txs`/`activate_rxs`, one pair per
+`NodeConfig`); `NodeActivation` (`numa.rs`) tracks how many of each node's
+dormant workers have been woken and grows every node by the same amount per
+step, capped by that node's remaining dormant workers and by the run's total
+budget (`n_total`) — balance across nodes is now guaranteed by construction,
+not incidental to channel implementation details.
+
 ## Open questions
 
 - **Error handling**: `run` currently returns the first error; remaining errors
   are dropped.  A `Vec<E>` return would give complete diagnostics.
 
-- **`workers_per_node` tuning**: currently `(cpus / 8).max(3).min(8)`, calibrated
+- **`INITIAL_DIVISOR` / `GROWTH_DIVISOR` tuning**: currently `4` and `8`
-  for merge on BeeGFS.  I/O-bound commands (`dump`, `select`) may benefit from
+  (start at 1/4 of a node's cores, grow by 1/8 per step), chosen to fix an
-  a higher value.  A per-call override could be added to the API.
+  observed too-slow ramp — not yet validated against a real `pack` (I/O-bound)
+  run, where over-provisioning risk is different from the CPU-bound `rebuild`
+  case this was tuned against.
 
 - **`on_done` ordering**: the runner serialises calls to `on_done` via an
   internal `Arc<Mutex<C>>`.  `Send` is required (the Arc clone crosses thread

src/Cargo.lock

@@ -1704,7 +1704,7 @@ dependencies = [
 
 [[package]]
 name = "obikmer"
-version = "1.1.30"
+version = "1.1.34"
 dependencies = [
  "clap",
  "csv",

src/obicompactvec/src/bitmatrix.rs

@@ -500,17 +500,26 @@ where T: Clone + Default {
 }
 
 /// Compute a symmetric `n×n` matrix in parallel by evaluating `f(i,j)` for
-/// all upper-triangle pairs. `T: Copy` avoids the `.clone()` needed for the
+/// all upper-triangle pairs, plus `f(i,i)` for the diagonal. `T: Copy` avoids
-/// lower-triangle mirror.
+/// the `.clone()` needed for the lower-triangle mirror.
+///
+/// The diagonal is *not* generally `T::default()`: for a self-comparison,
+/// `f(i,i)` is often the column's own weight (e.g. intersection-with-self —
+/// see `pairwise2_matrix`), not zero. Distance finalisations that need a
+/// zero diagonal (self-distance) already overwrite it explicitly.
 pub(crate) fn pairwise_matrix<T>(n: usize, f: impl Fn(usize, usize) -> T + Sync) -> Array2<T>
 where T: Copy + Default + Send {
     let results: Vec<(usize, usize, T)> = upper_pairs(n)
         .into_par_iter().map(|(i, j)| (i, j, f(i, j))).collect();
-    fill_symmetric(n, results.into_iter().map(|(i, j, v)| (i, j, v, v)))
+    let mut m = fill_symmetric(n, results.into_iter().map(|(i, j, v)| (i, j, v, v)));
+    for i in 0..n { m[[i, i]] = f(i, i); }
+    m
 }
 
 /// Same as `pairwise_matrix` but `f` returns two values that fill two
 /// symmetric matrices simultaneously (e.g. intersection + union for Jaccard).
+/// The diagonal is `f(i,i)` (e.g. a genome's kmer count intersected with
+/// itself), not `T::default()` — see `pairwise_matrix` for why that matters.
 pub(crate) fn pairwise2_matrix<T>(n: usize, f: impl Fn(usize, usize) -> (T, T) + Sync) -> (Array2<T>, Array2<T>)
 where T: Copy + Default + Send {
     let results: Vec<(usize, usize, T, T)> = upper_pairs(n)
@@ -523,5 +532,10 @@ where T: Copy + Default + Send {
         m0[[i, j]] = a; m0[[j, i]] = a;
         m1[[i, j]] = b; m1[[j, i]] = b;
     }
+    for i in 0..n {
+        let (a, b) = f(i, i);
+        m0[[i, i]] = a;
+        m1[[i, i]] = b;
+    }
     (m0, m1)
 }

src/obikindex/src/numa.rs

@@ -20,7 +20,7 @@ use hwlocality::cpu::binding::CpuBindingFlags;
 use hwlocality::cpu::cpuset::CpuSet;
 #[cfg(feature = "numa")]
 use hwlocality::object::types::ObjectType;
-use obisys::CpuSample;
+use obisys::{CpuSample, IoSample};
 use tracing::debug;
 
 // ── Public interface ──────────────────────────────────────────────────────────
@@ -70,7 +70,10 @@ pub fn build() -> NumaSetup {
                     nodes.len(),
                     nodes.first().map_or(0, |v| v.len()),
                 );
-                return NumaSetup { pools, cpus_per_node: nodes };
+                return NumaSetup {
+                    pools,
+                    cpus_per_node: nodes,
+                };
             }
         }
     }
@@ -102,7 +105,9 @@ pub fn build() -> NumaSetup {
 /// Silently returns on any error so the thread still runs, just unbound.
 #[cfg(feature = "numa")]
 pub fn pin_current_thread(cpu_indices: &[usize]) {
-    let Ok(topology) = Topology::new() else { return };
+    let Ok(topology) = Topology::new() else {
+        return;
+    };
     let mut cpuset = CpuSet::new();
     for &idx in cpu_indices {
         cpuset.set(idx);
@@ -132,7 +137,22 @@ fn build_pool(cpus: &[usize]) -> Option<rayon::ThreadPool> {
         .ok()
 }
 
-// ── PartitionRunner ───────────────────────────────────────────────────────────
+// ── PartitionRunner ─────────────────────────────────────────────────────────
+
+/// Growth step (fraction of a node's worker capacity added per activation
+/// event, see [`NodeActivation::grow`]).
+const GROWTH_DIVISOR: usize = 8;
+/// Minimum CPU efficiency growth to activate more workers, as a fraction of
+/// the size of the *last growth step* (e.g. `0.2` after adding 8 workers
+/// requires the next check to show at least +1.6 cores of growth — 20 % of
+/// the ~8 cores those 8 workers should contribute if the workload is truly
+/// CPU-bound). Scaling by the last step's size — not the cumulative total —
+/// keeps the bar meaningful regardless of how many workers are already
+/// active, instead of demanding an ever-larger absolute jump as the pool
+/// grows.
+const CPU_SPAWN_THRESHOLD: f64 = 0.2;
+/// Minimum I/O throughput growth (relative) to activate more workers.
+const IO_SPAWN_THRESHOLD: f64 = 0.2;
 
 struct NodeConfig {
     pool: Option<Arc<rayon::ThreadPool>>,
@@ -142,19 +162,23 @@ struct NodeConfig {
 
 /// Generic NUMA-aware runner for partition-level parallel work.
 ///
-/// Workers are distributed round-robin across NUMA nodes and pinned to their
+/// Workers are distributed evenly across NUMA nodes and pinned to their
 /// node's CPUs. UMA is the degenerate case: one node, no pinning.
 ///
-/// Workers are pre-spawned dormant and activated one by one as CPU efficiency
+/// Workers are pre-spawned dormant, one activation channel per node so
-/// falls below `SPAWN_THRESHOLD`. This avoids over-provisioning on I/O-bound
+/// growth always targets a specific node rather than whichever dormant
-/// or memory-bandwidth-bound workloads while saturating CPU-bound ones.
+/// worker happens to wake up first on a shared channel. Growth (both the
+/// initial count and each subsequent step) is expressed as a fraction of
+/// `workers_per_node`, applied identically to every node, so the pace of
+/// ramp-up depends on node size rather than node count — a single-NUMA-node
+/// (UMA) machine ramps just as fast as an 8-node one.
 ///
 /// # Termination
 ///
 /// ```text
 /// drop(part_tx)      → part_rx drains → workers exit → drop their result_tx
 /// drop(result_tx)    → result_rx closes → controller loop exits
-/// drop(activate_tx) → dormant workers exit cleanly
+/// drop(activate_txs) → dormant workers exit cleanly
 /// ```
 pub struct PartitionRunner {
     nodes: Vec<NodeConfig>,
@@ -175,7 +199,8 @@ impl PartitionRunner {
             ns.pools.len(),
             wpn,
         );
-        let nodes = ns.pools
+        let nodes = ns
+            .pools
             .into_iter()
             .zip(ns.cpus_per_node)
             .map(|(pool, cpu_ids)| NodeConfig {
@@ -189,23 +214,24 @@ impl PartitionRunner {
 
     /// Run `f(i)` for every index in `order`.
     ///
-    /// Workers are pre-spawned dormant and activated adaptively. A timer thread
+    /// Workers are pre-spawned dormant and activated adaptively, per node:
-    /// fires a CPU-efficiency check every `TIMER_SECS` seconds; each completed
+    /// `(workers_per_node / INITIAL_DIVISOR).max(1)` are woken immediately on
-    /// partition resets that timer (forcing an immediate check) and also
+    /// every node, then `(workers_per_node / GROWTH_DIVISOR).max(1)` more per
-    /// triggers its own inline check. A new worker is activated whenever
+    /// node each time the check below fires. A timer thread fires that check
-    /// efficiency falls below `SPAWN_THRESHOLD`.
+    /// every `TIMER_SECS` seconds; each completed partition resets that timer
+    /// (forcing an immediate check) and also triggers its own inline check. A
+    /// growth step happens whenever CPU efficiency grows by at least
+    /// `CPU_SPAWN_THRESHOLD` of what the last growth step should have
+    /// contributed, or I/O throughput grows by at least `IO_SPAWN_THRESHOLD`
+    /// (relative) since the last check — whichever resource is the actual
+    /// bottleneck still shows headroom.
     ///
     /// `on_done(i, result, elapsed)` is called from the controller thread as
     /// each partition completes — suitable for progress bars and result
     /// aggregation.
     ///
     /// Returns the first error produced by `f`, if any.
-    pub fn run<F, R, E, C>(
+    pub fn run<F, R, E, C>(&self, order: &[usize], f: F, mut on_done: C) -> Result<(), E>
-        &self,
-        order:      &[usize],
-        f:          F,
-        mut on_done: C,
-    ) -> Result<(), E>
     where
         F: Fn(usize) -> Result<R, E> + Send + Sync,
         R: Send,
@@ -217,22 +243,28 @@ impl PartitionRunner {
             return Ok(());
         }
 
-        const SPAWN_THRESHOLD: f64 = 0.2;
         const TIMER_SECS: u64 = 30;
+        const INITIAL_DIVISOR: usize = 4;
 
         // ── Channels ──────────────────────────────────────────────────────────
         let (part_tx, part_rx) = unbounded::<usize>();
-        let (activate_tx, activate_rx) = unbounded::<()>();
         // reset_tx: controller → timer ("reset the 30 s window")
         let (reset_tx, reset_rx) = unbounded::<()>();
         // event_tx: workers + timer → controller (unified event stream)
         let (event_tx, event_rx) = unbounded::<WorkerEvent<R, E>>();
+        // One activation channel per node: growth always targets a specific
+        // node, rather than whichever dormant worker happens to win the race
+        // on a channel shared across all nodes.
+        let (activate_txs, activate_rxs): (Vec<_>, Vec<_>) =
+            (0..self.nodes.len()).map(|_| unbounded::<()>()).unzip();
 
-        for &i in order { part_tx.send(i).ok(); }
+        for &i in order {
+            part_tx.send(i).ok();
+        }
         drop(part_tx);
 
         let max_workers = self.max_workers();
-        let n_nodes     = self.nodes.len();
+        let node_caps: Vec<usize> = self.nodes.iter().map(|n| n.max_workers).collect();
         let f = &f;
 
         let mut first_err: Option<E> = None;
@@ -256,18 +288,23 @@ impl PartitionRunner {
                 }
             });
 
-            // ── Pre-spawn workers dormant, round-robin across NUMA nodes ──────
+            // ── Pre-spawn workers dormant, grouped by node ────────────────────
-            for w in 0..max_workers {
+            // Each worker listens on its own node's activation channel only.
-                let node    = &self.nodes[w % n_nodes];
+            for (node, arx) in self.nodes.iter().zip(activate_rxs.iter()) {
+                let cpu_ids = &node.cpu_ids;
+                for _ in 0..node.max_workers {
                     let prx = part_rx.clone();
                     let etx = event_tx.clone();
-                let arx     = activate_rx.clone();
+                    let arx = arx.clone();
                     let pool = node.pool.clone();
-                let cpu_ids = &node.cpu_ids;
 
                     s.spawn(move || {
-                    if arx.recv().is_err() { return; }
+                        if arx.recv().is_err() {
-                    if !cpu_ids.is_empty() { pin_current_thread(cpu_ids); }
+                            return;
+                        }
+                        if !cpu_ids.is_empty() {
+                            pin_current_thread(cpu_ids);
+                        }
                         for i in &prx {
                             let t = Instant::now();
                             let r = match &pool {
@@ -278,15 +315,17 @@ impl PartitionRunner {
                         }
                     });
                 }
+            }
             // Drop controller's event_tx: event_rx closes when all workers +
             // timer have exited.
             drop(event_tx);
 
             // ── Controller ────────────────────────────────────────────────────
-            let initial_workers = n_nodes.min(max_workers).min(n_total);
+            let mut activation = NodeActivation::new(&activate_txs, &node_caps, max_workers);
-            for _ in 0..initial_workers { activate_tx.send(()).ok(); }
+            activation.activate_initial(INITIAL_DIVISOR, n_total);
-            let mut n_active              = initial_workers;
+
             let mut cpu_sample = CpuSample::now();
+            let mut io_sample = IoSample::now();
             let mut completed = 0usize;
 
             while completed < n_total {
@@ -295,28 +334,39 @@ impl PartitionRunner {
                     WorkerEvent::Completed(i, r, dur) => {
                         match r {
                             Ok(v) => on_done(i, v, dur),
-                            Err(e) => { if first_err.is_none() { first_err = Some(e); } }
+                            Err(e) => {
+                                if first_err.is_none() {
+                                    first_err = Some(e);
+                                }
+                            }
                         }
                         completed += 1;
                         // Reset the 30 s timer.
                         reset_tx.send(()).ok();
                         // Inline check: same logic as a timer tick.
                         maybe_activate(
-                            &activate_tx, &mut n_active, max_workers,
+                            &mut activation,
-                            &mut cpu_sample, SPAWN_THRESHOLD, completed, n_total,
+                            &mut cpu_sample,
+                            &mut io_sample,
+                            completed,
+                            n_total,
                         );
                     }
                     WorkerEvent::TimerTick => {
                         maybe_activate(
-                            &activate_tx, &mut n_active, max_workers,
+                            &mut activation,
-                            &mut cpu_sample, SPAWN_THRESHOLD, completed, n_total,
+                            &mut cpu_sample,
+                            &mut io_sample,
+                            completed,
+                            n_total,
                         );
                     }
                 }
             }
 
-            // Dormant workers exit when activate_tx closes.
+            // Dormant workers exit once every sender for their node's channel
-            drop(activate_tx);
+            // is dropped — `activate_txs` holds the only ones.
+            drop(activate_txs);
             // Timer thread exits when reset_tx closes.
             drop(reset_tx);
         });
@@ -335,20 +385,113 @@ enum WorkerEvent<R, E> {
     TimerTick,
 }
 
+/// Tracks how many of each node's dormant workers have been woken, and
+/// grows every node by the same amount at each step (capped by that node's
+/// remaining dormant workers and by the run's total budget) so load stays
+/// balanced across nodes at every point in time — never just "one more
+/// worker somewhere". Also remembers the size of the last real growth step
+/// (`last_step`), used to scale the CPU activation threshold to what that
+/// step could plausibly have contributed (see `maybe_activate`).
+struct NodeActivation<'a> {
+    txs: &'a [crossbeam_channel::Sender<()>],
+    caps: &'a [usize],
+    active: Vec<usize>,
+    total: usize,
+    max: usize,
+    last_step: usize,
+}
+
+impl<'a> NodeActivation<'a> {
+    fn new(txs: &'a [crossbeam_channel::Sender<()>], caps: &'a [usize], max: usize) -> Self {
+        Self {
+            txs,
+            caps,
+            active: vec![0; txs.len()],
+            total: 0,
+            max,
+            last_step: 0,
+        }
+    }
+
+    fn total(&self) -> usize {
+        self.total
+    }
+    fn last_step(&self) -> usize {
+        self.last_step
+    }
+    fn max(&self) -> usize {
+        self.max
+    }
+    fn is_full(&self) -> bool {
+        self.total >= self.max
+    }
+
+    /// Wake up to `(node_cap / divisor).max(1)` dormant workers on every
+    /// node, capped by `n_total`. Called once at startup, unconditionally.
+    fn activate_initial(&mut self, divisor: usize, n_total: usize) {
+        self.grow(divisor, n_total);
+    }
+
+    /// Same per-node sizing as [`activate_initial`](Self::activate_initial),
+    /// applied as a growth step. Returns the number of workers actually
+    /// activated (may be less than requested once a node or the total
+    /// budget is exhausted). Updates `last_step` when it actually grew.
+    fn grow(&mut self, divisor: usize, n_total: usize) -> usize {
+        let before = self.total;
+        for idx in 0..self.txs.len() {
+            let wanted = (self.caps[idx] / divisor).max(1);
+            let room = self.caps[idx].saturating_sub(self.active[idx]);
+            let grow = wanted.min(room).min(n_total.saturating_sub(self.total));
+            for _ in 0..grow {
+                self.txs[idx].send(()).ok();
+            }
+            self.active[idx] += grow;
+            self.total += grow;
+        }
+        let grew = self.total - before;
+        if grew > 0 {
+            self.last_step = grew;
+        }
+        grew
+    }
+}
+
 fn maybe_activate(
-    activate_tx: &crossbeam_channel::Sender<()>,
+    activation: &mut NodeActivation,
-    n_active:    &mut usize,
-    max_workers: usize,
     cpu_sample: &mut CpuSample,
-    threshold:   f64,
+    io_sample: &mut IoSample,
     completed: usize,
     n_total: usize,
 ) {
-    if *n_active >= max_workers || completed >= n_total { return; }
+    if activation.is_full() || completed >= n_total {
+        return;
+    }
 
-    if cpu_sample.do_i_activate(threshold) {
+    // Expect roughly 1 core of extra efficiency per worker activated in the
-        activate_tx.send(()).ok();
+    // last growth step (CPU-bound case); require at least CPU_SPAWN_THRESHOLD
-        *n_active += 1;
+    // (20 %) of that expected gain before growing again. Scaling by the last
-        debug!("activated worker {}/{}", n_active, max_workers);
+    // step's size — not the cumulative total — keeps the bar meaningful
+    // regardless of how many workers are already active: growing by 8 should
+    // always take ~+1.6 cores to confirm, whether that's the 2nd growth step
+    // or the 20th.
+    let cpu_threshold = CPU_SPAWN_THRESHOLD * activation.last_step() as f64;
+
+    // Call both unconditionally (no `||` short-circuit): each sampler must
+    // advance its own window every tick, regardless of what the other one
+    // reports, or it would starve behind whichever signal fires first.
+    let cpu_wants_more = cpu_sample.do_i_activate(cpu_threshold);
+    let io_wants_more = io_sample.do_i_activate(IO_SPAWN_THRESHOLD * activation.last_step() as f64);
+    if !(cpu_wants_more || io_wants_more) {
+        return;
+    }
+
+    let grew = activation.grow(GROWTH_DIVISOR, n_total);
+    if grew > 0 {
+        debug!(
+            "activated {} worker(s) — {}/{} active",
+            grew,
+            activation.total(),
+            activation.max()
+        );
     }
 }

src/obikmer/Cargo.toml

@@ -1,6 +1,6 @@
 [package]
 name = "obikmer"
-version = "1.1.30"
+version = "1.1.34"
 edition = "2024"
 
 [[bin]]

src/obisys/src/lib.rs

@@ -266,14 +266,19 @@ impl CpuSample {
     }
 
     pub fn do_i_activate(&mut self, threshold: f64) -> bool {
+        let delta_wall = self.wall.elapsed().as_secs_f64();
+        if delta_wall < 0.1 {
+            // Window too short to be meaningful — leave state untouched so it
+            // keeps accumulating until a real sample can be taken.
+            return false;
+        }
+
         let n = CpuSample::now();
         let delta_ru = (n.user_secs - self.user_secs) + (n.sys_secs - self.sys_secs);
-        let delta_wall = self.wall.elapsed().as_secs_f64();
 
         let efficiency = delta_ru / delta_wall;
         let activate = 0f64.max(efficiency - self.previous) >= threshold;
 
-        if activate {
         debug!(
             "Do I activate : {} -> {} = {} Activate: {}",
             self.previous,
@@ -285,7 +290,93 @@ impl CpuSample {
         self.user_secs = n.user_secs;
         self.sys_secs = n.sys_secs;
         self.wall = n.wall;
+
+        activate
     }
+}
+
+// ── IoSample ──────────────────────────────────────────────────────────────────
+
+/// Snapshot of process-wide block I/O (bytes read + written) + wall clock.
+///
+/// Same activation protocol as [`CpuSample`], but the growth check in
+/// [`do_i_activate`](Self::do_i_activate) is *relative* rather than absolute:
+/// raw I/O throughput has no portable scale across storage devices, unlike a
+/// core count.
+pub struct IoSample {
+    wall: Instant,
+    bytes: u64,
+    previous_rate: f64,
+}
+
+impl IoSample {
+    pub fn now() -> Self {
+        Self {
+            wall: Instant::now(),
+            bytes: Self::read_bytes(),
+            previous_rate: 0.0,
+        }
+    }
+
+    /// Bytes actually submitted to the block layer (read + write), summed
+    /// process-wide. Returns 0 if unavailable — degrades gracefully to a
+    /// signal that never triggers activation (CPU-only heuristic).
+    #[cfg(target_os = "linux")]
+    fn read_bytes() -> u64 {
+        let Ok(io) = std::fs::read_to_string("/proc/self/io") else {
+            return 0;
+        };
+        io.lines()
+            .filter_map(|l| {
+                l.strip_prefix("read_bytes: ")
+                    .or_else(|| l.strip_prefix("write_bytes: "))
+            })
+            .filter_map(|v| v.trim().parse::<u64>().ok())
+            .sum()
+    }
+
+    #[cfg(target_os = "macos")]
+    fn read_bytes() -> u64 {
+        use libc::{RUSAGE_INFO_V4, getpid, proc_pid_rusage, rusage_info_v4};
+        let mut info: rusage_info_v4 = unsafe { std::mem::zeroed() };
+        let ret =
+            unsafe { proc_pid_rusage(getpid(), RUSAGE_INFO_V4, &mut info as *mut _ as *mut _) };
+        if ret != 0 {
+            return 0;
+        }
+        info.ri_diskio_bytesread + info.ri_diskio_byteswritten
+    }
+
+    #[cfg(not(any(target_os = "linux", target_os = "macos")))]
+    fn read_bytes() -> u64 {
+        0
+    }
+
+    /// Same protocol as [`CpuSample::do_i_activate`] (0.1 s minimum window,
+    /// state untouched on early return), but growth is measured relative to
+    /// the previous rate. `threshold` is a fraction, e.g. `0.2` for a 20 %
+    /// increase in throughput since the last real sample.
+    pub fn do_i_activate(&mut self, threshold: f64) -> bool {
+        let elapsed = self.wall.elapsed().as_secs_f64();
+        if elapsed < 0.1 {
+            return false;
+        }
+
+        let n = Self::read_bytes();
+        let rate = n.saturating_sub(self.bytes) as f64 / elapsed;
+        let activate = if self.previous_rate == 0.0 {
+            rate > 0.0 // bootstrap: any measured throughput is signal enough
+        } else {
+            (rate - self.previous_rate) / self.previous_rate >= threshold
+        };
+
+        debug!(
+            "Do I activate (I/O) : {} -> {} Activate: {}",
+            self.previous_rate, rate, activate
+        );
+        self.previous_rate = rate;
+        self.bytes = n;
+        self.wall = Instant::now();
 
         activate
     }