Performance Model & Tools Summary

Performance Model & Tools Summary

Hung-Hsun Su

UPC Group, HCS lab

2/5/2004

Models

Amdahl’s law, Scaled-speedup, LogP, cLogP, BSP Parametric micro-level (PM, 1994)

Predict execution time, identify bottleneck, compare machines Incorporate precise details of interprocessor communication, memory

operations, auxiliary instructions and effects of communication and computation schedules

Derive analytical formulas experimental measurement of sample cases estimate misc. overhead refine formula predict execution time using formula

ZPL (1998) Model incorporated into language design Scalar performance, concurrency and interprocessor communication Identify interacting regions to determine how the data/processor is

mapped Once mapping is know, the cost is calculated by Also try to compare alternative solutions through formula

Models “Analytical Modeling of Parallel Programs”

Execution time, Total Parallel Overhead, Speedup, Efficiency, Cost Isoefficiency function

Determines the ease with which it can achieve speedups increasing in proportion to number of processors (small highly scalable)

Determine if system is “cost-optimal” if [(Num. Proc) * Tp] vs Ts is proportional to each other Calculation of lower bound is use to determine the degree of concurrency

Minimum execution time and cost-optimal execution time Asymptotic Analysis

Analyzing performance using kernel performance Define coupling (interaction) between kernels that tries to improve the accuracy

Overhead Model Generalized Amdahl’s law model Lost Cycles Analysis

Agarwal network model Wire, switch delays, message size, communication latency (contention not considered)

Closed queueing network model Task graph that gives the synchronization constraints and use a closed queuing model to describe

contention delay Predict mean response time and resource utilization

Anita W. Tam Model Application – establishes a relationship between message generation rate and communication latency Network Model – provide average message latency as function of message generation rate of nodes

together with other system parameters

EPPA*

*All information regarding EPPA taken from http://parallel.vub.ac.be/research/parallel_performance/

EPPA

Information Retained The different phases of the program, like useful computation,

partitioning (Cost of each phase, its impact on the performance) The experiment parameters, like #processors, work size,

hardware, … (Multiple experiment analysis: measurements in function of parameters)

The #quantums processed and communicated in each phase (Time of the phases in function of #quantums)

The #operations that are computed in each phase#operations per quantum (Time of phases in function of #basic operations)

Does not use hardware counters, give first-order analysis

EPPA

EPPA

PROPHET*

*All information regarding PROPHET taken from http://www.par.univie.ac.at/project/prophet/

PROPHET

prediction of the performance behavior of parallel and distributed applications on cluster and grid architectures

Based on a UML model of an application and a simulator for a target architecture, one can predict the execution behavior of the application model

SCALEA*

*All information regarding SCALEA taken from http://www.par.univie.ac.at/project/scalea/

SCALEA

Profile/Trace Analysis Inclusive/Exclusive Analysis Load balancing Analysis Metric Ratio Analysis Execution Summary

Overhead analysis Region to Overhead Overhead to region

Analysis functions

AKSUM*

*All information regarding AKSUM taken from http://www.par.univie.ac.at/project/aksum/

AKSUM

Automatic performance bottleneck analysis tool Performance properties are normalized

Performance property name Threshold Reference code region

Prediction Tools

P3T performance estimator for HPF programs closely integrated with

VFCS The core part of P3T is centered around a set of parallel program

parameters (transfer time, number of transfers, computation time, etc.

Carnival attempt to automate the cause-and-effect inference process for

performance phenomena Network Weather Service

uses numerical models and monitored readings of current conditions to dynamically forecast the performance that various network and computational resources can deliver over a given time frame

Knowledge-based Tools

Autopilot aims at dynamically optimizing the performance of parallel

applications. Kappa-PI

knowledge-based performance analyzer for parallel MPI and PVM programs. The basic principle of the tool is to analyse the efficiency of an application and provide the programmer with some indications about the most important performance problem found in the execution

Organizations

APART - IST Working Group on Automatic Performance Analysis: Real Tools http://www.fz-juelich.de/apart-2/

Parallel Tools Consortium http://www.ptools.org/

Interesting Ideas

Tool that facilitate going from one system to another