Presentations are listed in reverse chronological order
Abstract Selecting the close-to-optimal collective algorithm based on the parameters of the collective call at run time is an important step in achieving good performance of MPI applications. In this paper, we focus on MPI collective algorithm selection process and explore the applicability of the quadtree encoding method to this problem. We construct quadtrees with different properties from the measured algorithm performance data and analyze the quality and performance of decision functions generated from these trees. The experimental data shows that in some cases, the decision function based on a quadtree structure with a mean depth of 3 can incur as little as a 5\% performance penalty on average. The exact, experimentally measured, decision function for all tested collectives could be fully represented using quadtrees with a maximum of 6 levels. These results indicate that quadtrees may be a feasible choice for both processing of the performance data and automatic decision function generation.
Available in
PDF (1.5MB)
format.
Abstract
Previous studies of application usage show that the performance of
collective communications are critical for high-performance computing and are
often overlooked when compared to the point-to-point performance.
In this paper, we analyze and attempt to improve intra-cluster collective communication in
the context of the widely deployed MPI programming paradigm by extending accepted models
of point-to-point communication, such as Hockney, LogP/LogGP, and PLogP.
The predictions from the models were compared to the experimentally gathered data and
our findings were used to optimize the implementation of collective operations in the FT-MPI
library.
MPI Collective Operation Optimization
FT-MPI Project: Graham E. Fagg, Edgar Gabriel, George Bosilca, Jelena Pjesivac-Grbovic, Thara Anguskun
Available in PDF (3.1MB) format
Abstract
Collective operations are a frequently used MPI component.
Numerical libraries as well as user applications rely on
the collective operations to synchronize, distribute, reduce and gather data.
Suboptimal implementation of collective operations can have significant
negative impact on overall application or library performance.
In recent years, major public domain MPI implementations have began to pay
attention to their collective operation algorithms.
Our research goals are twofold. First, we want to improve FT-MPI's
overall collective operation performance. We have conducted extensive
experimental testing of performance of Barrier, Broadcast, Reduce, Scatter,
Scatter(v), Gather and Gatherv algorithms. The experimental data allow us
to select group of algorithms that perform best overall and to come up with
static run-time decision functions that select optimal algorithm based on number
of processors and message size of the data. The experimental data
also confirmed that considering only these two parameters is often not enough to
correctly make decision. The optimal collective operation algorithm
also depends on underlying network topology, load balance among the nodes, and
application specific communication patterns.
Our second research goal is the development of analytical model that would be able
to take into account more parameters. The model would be able to make correct
predictions of algorithm performance based on the description of the system
and possible, application.
Currently, we are still in the process of experimenting with various known
analytical models, and trying to compare their predictions to the
experimental data.
The talk will describe the algorithms we use for some of
the most frequently used collective operations, introduce new
asynchronous broadcast algorithm developed by George Bosilca and myself, compare
FT-MPI collective operation performance with MPICH 2 and LAM 7, and show some of
the results of the ongoing analysis of the collective operation
performance models.
Multiscale Modeling of Avascular Tumor Growth
Co-authors: Yi Jiang, James P. Freyer, Jelena Pjesivac-Grbovic, Charles Cantrell
Available in
PPT (1.3Mb) format
Abstract
The microenvironment inside a tumor is extremely complex and adaptive, involving spatial
and temporal variations in nutrient and waste gradients, cellular physiology, metabolism,
the expression patterns of genes and proteins as well as the malignant progression.
The sum of all these elements defines the response of a tumor to treatment.
The multicellular tumor spheroid system has been a primary example of in vitro models of
the tumor microenvironment, which has provided numerous insights into tumor dynamics and progression.
We develop a multiscale model to study spheroid tumor growth, which includes at
the subcellular level a protein expression network that controls the possibility for
cell cycle arrest, and at the cellular level a hybrid of cellular dynamics
(lattice Monte Carlo) and reaction-diffusion dynamics of chemicals (PDEs).
This integrated subcellular and cellular model provides a realistic representation of
both structure and dynamics over a large range of time and length scales.
Our simulations show favorable comparisons with spheroid experiments. The combination of
a data-rich experimental system with sophisticated mathematical modeling holds the promise
of an improved basic understanding of malignant progression and therapeutic response in
humans
LA-UR-03-5486
Modeling Avascular Tumor Growth
Co-authors: Yi Jiang, James P. Freyer, Jelena Pjesivac-Grbovic
Available in PPT (1.5Mb) format
Abstract
The current understanding of mechanisms behind the early stage of tumor development is
far from complete. We have developed a cellular model to study this process. Many of
the relevant experimental data came from in vitro tumor models such as multicellular
tumor spheroids that exhibit most of the properties of avascular spherical tumors.
An avascular spherical tumor grows into a layered structure consisting of
a necrotic core, a quiescent layer, which consists of living cells arrested mainly in G1 phase due to
stressful environment, and the outermost proliferating layer of cells.
We have developed a cellular model to study tumor growth. This model combines a lattice
Monte Carlo model for cells and continuous reaction-diffusion dynamics for nutrients,
wastes, and growth and inhibitory factors. The cell cycle dynamics is governed by a
protein regulatory network which uses a group of proteins that determine a cell's
transition from G1 to S phase. We have obtained cell population and distribution data
consistent with spheroid experiments.
LA-UR-02-4364
Presented at LACSI symposium, Santa Fe, NM, October 2004.
Co-authors: Graham E. Fagg, Edgar Gabriel, George Bosilca,
Jelena Pjesivac-Grbovic, Thara Anguskun.
Available in PDF
(3,288KB) format.
Multiscale Model of Tumor Growth
Presented at
Presented at