Problem types: NumPy, Numba and JAX on CPU and GPU, to guide the choice of stack

Aperta
#5 2 commenti 0 reazioni 0 assegnatari Vedi su GitHub

I maintainer di solito rispondono entro 1 giorno

Nessuno ha ancora preso questa issue.

Valutazione

Difficoltà
5/5
Tempo stimato
Più di una settimana
Idoneità per principianti
32/100
Tipo di issue
Funzionalità
Chiarezza
Specificata chiaramente
Stato di attività
Attiva
Stack tecnologico
numpy, python

Direzione di ricerca

Read the existing precision/ harness and results from #1, then review meta#264 and the lecture runs in #3 for the intended crossover questions. Start with P1, P4, P3 and P7, implementing and comparing NumPy, Numba and JAX variants across the stated sizes, precision settings and devices. Done means validated timings, results under results/, a dashboard by problem type and a short guide with recommendations.

Scritto dal modello di indicizzazione a partire dal testo della issue.

Descrizione

A suite of the problem types QuantEcon's lectures and research use, each written idiomatically in NumPy, Numba and JAX, and timed as used on CPUs and GPUs across problem sizes. The aim is data-driven guidance on which stack to use for which kind of problem, at what size, and on what hardware. This is the experiment for meta#264.

Why

  • meta#264 asks for exactly this: crossover points between backends and devices, by problem type and size, to replace rules of thumb.
  • #1 measures JAX only, by kernel type. Its precision matrix (results under results/) shows how differently kernel types behave on a GPU. On a TITAN RTX, float64 costs 1.3× on a latency-bound recursion, about 4× on a memory-bound step and 20× to 27× on compute or transcendental work. It doesn't say whether Numba on a CPU would beat JAX for the same problem.
  • The lecture runs (#3) found that how a problem is written matters as much as the hardware. A vectorised Bellman step and a looped one, or a stored reward array and a recomputed one, land in different regimes.

Problem types

Each type gets a NumPy version as the baseline, a Numba version (CPU, parallel=True where it fits) and a JAX version (CPU and GPU), each written the way a careful author in that stack would write it.

Problem type QuantEcon pattern Shape
P1 Value function iteration on a dense grid Savings and discrete-choice models Vectorised max over a choice grid; loops in Numba
P2 Policy iteration Howard improvement with a linear solve Dense or iterative solves
P3 Endogenous grid method Income fluctuation problems Interpolation
P4 Monte Carlo simulation Many paths of an AR(1) or Markov process Parallel across paths, sequential in time
P5 Markov chains Stationary distributions, simulation Sparse and dense linear algebra
P6 Root finding and optimisation per state Newton or bisection over a grid Many small independent solves
P7 Sequential recursions Kalman filter, lax.scan Latency-bound
P8 Gradient-based estimation Maximum likelihood with autodiff Where JAX's autodiff should matter

Dimensions

  • Problem size, swept to find the crossovers.
  • Precision: float64 throughout, plus float32 where the problem tolerates it.
  • Device: CPU and GPU, for JAX. Numba runs on CPU only; numba.cuda is out of scope unless a problem type calls for it.
  • Dask (in meta#264) is left out, unless a problem type needs data larger than memory or distributed scheduling.

Method

#1's house standard:

  • one fresh process per measurement, with compile time counted, and cold and warm times kept apart;
  • the result dtype asserted;
  • every implementation's answer checked against the NumPy baseline.

The harness extends precision/, with one module per problem type and one function per stack.

Platforms

As in #1: qe-mini1's M6 and qe-wharton1-gpu's Threadripper and TITAN RTX first. Then qe-wharton2's M5 Pro once it is onboarded, and the RunsOn T4 once RunsOn is configured for this repository.

Phases

  1. P1, P4, P3 and P7 first. They span the regimes #1 found, and they shake out the harness across three stacks.
  2. Then P2, P5, P6 and P8.

Output

  • results under results/, as for #1;
  • a dashboard view by problem type, showing the fastest stack and device at each size;
  • for the lecture series, a short guide of the form "for problem type X at size N, use stack S on device D", with the numbers beside it.

🤖 Generated with Claude Code

Lingua principale
Python
Stelle
0
Fork
1
Merge medio
3h 29m
PR unite (30g)
64

Preparare l'ambiente

Questo progetto non fornisce container di sviluppo, Dockerfile né guida per i contributori, quindi l'ambiente è a tuo carico: parti dal suo README e consulta la nostra guida al primo contributo per i passaggi generali.

Come iniziare

  1. Leggi tutta la issue e poi la guida ai contributi del progetto.
  2. Commenta sulla issue per dire che te ne occupi tu — evita che due persone facciano lo stesso lavoro.
  3. Fai un fork del repository e lavora su un branch.
  4. Apri una pull request che faccia riferimento al numero della issue.

Altre issue di QuantEcon/benchmarks

Tutte le issue di QuantEcon/benchmarks

Issue simili

Altre issue su Python

Ricevi le nuove issue nella tua casella

Un breve riepilogo di issue GitHub adatte ai principianti.