A pulsar-search pipeline dedisperses a radio time series in parallel batches

PHYSICS · Multi-part · Fermi University Physics · Computational Physics · High-performance and data-driven computational physics · High-performance and data-driven computational physics

Scientific Computing for Physicists, Ch Many processors: speed-up, decomposition and the cost of communication

Problem

A pulsar-search pipeline dedisperses a radio time series in parallel batches, and a companion image-domain step is decomposed across a grid of processes. ((a)) A fraction

f=0.99

of the one-process time parallelises; compute Amdahl's speed-up

S(100)

on 100 processes and its ceiling. ((b)) Run instead at fixed work per process, with

f'=0.995

; compute Gustafson's scaled speed-up

S(200)

on 200 processes. ((c)) The image-domain step exchanges one halo strip of

8192\,\mathrm{B}

with each of two neighbours every step. A ping-pong test gives

\alpha_{\rm lat}=0.80\,\mu\mathrm{s}

and

\beta_{\rm net}=1.2\times10^{10}\,\mathrm{B\,s^{-1}}

. Find the time in the two halo messages per step and its percentage of a step update measured at

0.210\,\mathrm{ms}

. ((d)) A triangular pair loop over candidate pairs is split into

P=6

equal row ranges. Using

(2P-1)/P^2

for the first range's share, find that share and the resulting bound on speed-up.

Browse the problem library