TITLE:

Sustainable Multi-level Tuning for Exascale Science Applications

DATE:

Friday, December 2nd, 2016

TIME:

3:30 PM

LOCATION:

GMCS 314

SPEAKER:

Dr. Tony Drummond. Senior Scientist at Global Footprint Network.

ABSTRACT:

Enabling performance sustainability for complex parallel computer

applications and high-end simulations requires the coherent integration

of software programming practices and environments, static tuning and

automatic tuning, both on-line and off-line. The end goal is to produce

code that maximizes the resource utilization, runs at desired

computational scales and produces meaningful results that advance the

science. Here, we look at a multi-stage software tuning workflow, which

covers from algorithmic implementations to full computational

applications and large-scale simulations.

This talk will summarize some lessons learned from a computational

tuning workbench where we first use tracing and profiling tools to

identify key hotspots in the code, which are later optimized depending

on the nature of their computational intensity and the available

hardware. The next level up in tuning, is the use of different

optimized algorithmic implementations or functionalities that are

available in the form of computational kernels. Implementations of

these kernels are part of general-purpose libraries, software tools or

other application codes. Lastly, we look automating the selection of

best kernel in a way that it is sustainably integrated into the

application tuning.

HOST:

Dr. Jose Castillo

DOWNLOAD: