coren

Compute and resource normalization.

Measures what your machine can do. Built for irohds -- to tell you whether a computation is faster to run locally or fetch from the network.

Any two machines looking at the same function can independently agree on how much work it requires (an almost deterministic op count). Each machine knows its own capabilities (benchmarked once at startup). The verdict is local arithmetic: compare estimated compute time against estimated fetch time.

Integration into irohds, a decentralized memoization system for scientific computing, was the initial reason for making this package. It is also useful as a standalone tool for roofline analysis, ETL buffer sizing, and build parallelism decisions.

Install

uv add coren

Or from source:

uv run maturin develop --features python

Usage

from coren import FnCost, MachCap

# Describe the work (deterministic, same on every machine)
cost = FnCost.sort(1_000_000, 64, 64_000_000)

# Measure this machine (benchmarks run once, cached)
cap = MachCap.read()

# Get the answer
v = cap.verdict(cost)
print(v)  # "compute (saves 0.712s)" or "fetch (saves 2.3s)"

if v.should_fetch():
    download_result()
else:
    compute_locally()

How it works

Two layers:

FnCost describes a function's resource requirements in absolute physical units. Four integers: ops (total arithmetic operations), mem_bytes (memory traffic), peak_mem (peak RAM footprint), result_bytes (output size). These are properties of the algorithm and its inputs. Bitwise identical on every machine.

MachCap describes what this machine can do. Measured via micro-benchmarks (FMA throughput, STREAM triad, disk sequential I/O) and OS queries (NIC link speed, battery state, core count, RAM). Produces a roofline model: peak ops/s and memory bandwidth.

The verdict compares estimated compute time (from the roofline model) against estimated fetch time (result_bytes / NIC bandwidth). The score is the difference in seconds: positive means fetch is faster, negative means compute is faster. Infinity means one option is impossible (no RAM, or no network).

FnCost constructors

FnCost.new(ops, mem_bytes, peak_mem, result_bytes)   raw values
FnCost.scan(n_bytes, result_bytes)                   linear scan
FnCost.sort(n, item_bytes, result_bytes)             merge sort
FnCost.hash(n_bytes)                                 crypto hash
FnCost.matmul(m, n, k, result_bytes)                 dense GEMM
FnCost.etl(rows, row_bytes, ops_per_row, result_bytes)  row processing
FnCost.copy(size)                                    file copy (ops=0)

Combinators

a.then(b)    # sequential: ops sum, peak_mem = max, result = b's output
a + b        # same as then
a.par(b)     # parallel: ops = max, peak_mem sums, result sums
a.repeat(k)  # k iterations, peak_mem unchanged

Normalizing wall-clock measurements

When a function is executed and you only know the wall-clock time (not the algorithmic complexity), MachCap.normalize() converts the measurement into a FnCost suitable for local verdict computation.

WARNING: normalize() output is NOT deterministic across machines. Different machines produce different ops/mem_bytes values for the same function. Do NOT use normalize()-produced FnCost as cache keys, content addresses, or any identifier that must match across peers. For cache keys, use the static constructors (sort, hash, matmul, etc.) or FnCost.new() with values derived from the function's definition and parameters.

cost = cap.normalize(compute_ns=15_000_000_000, peak_mem=4_000_000_000, result_bytes=500_000_000)
# cost is a safe overestimate, suitable for verdict() but NOT for cache keys

CLI

$ coren
coren  [desktop]
  cores                    8p / 16l
  frequency                3600 MHz

  roofline (measured)
    peak (all cores)       89.2 GFLOPS
    mem bandwidth          38.1 GB/s
    disk bandwidth         1.52 GB/s
    ridge point            2.34 ops/byte

  verdicts (what should this machine do?)
    task                       ops         Q         R    score       neck action
    sort 1M x 64B          200.0M     1.2GB    10.0MB   -0.712   memory compute
    matmul 1k^3              2.0G    22.9MB    10.0MB   -0.779  compute compute

$ coren --json   # machine-readable output

Rust

use coren::{FnCost, MachCap};

let cost = FnCost::sort(1_000_000, 64, 64_000_000);
let cap = MachCap::read(".");
let v = cap.verdict(&cost);

if v.should_fetch() {
    // download from peer
} else {
    // compute locally
}

License

MIT OR Apache-2.0