Explore
For Enterprise
Join for Free
Log In

deeplearning.ai
TensorFlow: Data and Deployment Specialization
Join Andrew Ng and Laurence Moroney in the latest Specialization from deeplearning.ai.
Imperial College London
New! Online MSc in Machine Learning and Data Science
Learn the theory of machine learning and its applications from a top 10 university.

Explorez l'intégralité de Coursera

Parcourir
Meilleurs cours
For Enterprise
Connexion
Inscrivez-vous gratuitement

Proof - e

Loading...

Quantitative Formal Modeling and Worst-Case Performance Analysis

4.4 (39 évaluations) | 6.2K étudiants inscrits

S'inscrire gratuitement

Welcome to Quantitative Formal Modeling and Worst-Case Performance Analysis. In this course, you will learn about modeling and solving performance problems in a fashion popular in theoretical computer science, and generally train your abstract thinking skills. After finishing this course, you have learned to think about the behavior of systems in terms of token production and consumption, and you are able to formalize this thinking mathematically in terms of prefix orders and counting functions. You have learned about Petri-nets, about timing, and about scheduling of token consumption/production systems, and for the special class of Petri-nets known as single-rate dataflow graphs, you will know how to perform a worst-case analysis of basic performance metrics, like throughput, latency and buffering. Disclaimer: As you will notice, there is an abundance of small examples in this course, but at first sight there are not many industrial size systems being discussed. The reason for this is two-fold. Firstly, it is not my intention to teach you performance analysis skills up to the level of what you will need in industry. Rather, I would like to teach you to think about modeling and performance analysis in general and abstract terms, because that is what you will need to do whenever you encounter any performance analysis problem in the future. After all, abstract thinking is the most revered skill required for any academic-level job in any engineering discipline, and if you are able to phrase your problems mathematically, it will become easier for you to spot mistakes, to communicate your ideas with others, and you have already made a big step towards actually solving the problem. Secondly, although dataflow techniques are applicable and being used in industry, the subclass of single-rate dataflow is too restrictive to be of practical use in large modeling examples. The analysis principles of other dataflow techniques, however, are all based on single-rate dataflow. So this course is a good primer for any more advanced course on the topic. This course is part of the university course on Quantitative Evaluation of Embedded Systems (QEES) as given in the Embedded Systems master curriculum of the EIT-Digital university, and of the Dutch 3TU consortium consisting of TU/e (Eindhoven), TUD (Delft) and UT (Twente). The course material is exactly the same as the first three weeks of QEES, but the examination of QEES is at a slightly higher level of difficulty, which cannot (yet) be obtained in an online course.

Visualiser le programme de cours

Avis

4.4 (39 évaluations)

5 stars
26 ratings
4 stars
7 ratings
3 stars
3 ratings
2 stars
1 ratings
1 star
2 ratings

À partir de la leçon

Performance analysis

In this module/week you will learn to exploit the structure of single-rate dataflow graphs to perform worst-case analysis of performance metrics like throughput, latency and buffering. After this week, you know how to calculate the maximum cycle mean of a dataflow graph, how to construct a periodic schedule for it, how to optimize this schedule for latency analysis, and how to determine the size of buffers with back-pressure such that the worst-case analysis remains valid. If you understood the material of the previous module/week, the proofs presented in this week will give you a deeper understanding of the mathematical underpinning of these methods.

Throughput is bounded by 1/MCM8:35

The throughput bound is tight5:05

Enseigné par

Dr.ir. Pieter Cuijpers
Assistant Professor
Anne Remke
Prof. dr.

Essayer le cours pour Gratuit USD

Transcription

Right, so we need a step in our proof to show that, if you take the limit of n over p(n), then this implies that, if that limit is bigger than our bounds, then this implies that the actual limit that we're looking for is also bigger than the bound. Okay, let's have a look. This is the theorem. First thing that we're going to do is to assume that our limit of n over p(n) is actually bigger than the bound that we're looking for. And then we want to reach the conclusion that the limit that we're looking for is also bigger than that bound. Okay, so saying that the limit that we are looking for is bigger than the bound. If you remember the definition of the limit, is the same as saying that, for every epsilon bigger than 0, so lets assume that we have some epsilon bigger than 0, for epsilon bigger than 0, there must exist a maximum antichain, so we have to pick that in some way. And, for every point past that maximum antichain, we want to find that the ratio between the number of tokens at that point, w past the antichain divided by the time at that point, is larger than 1 over the MCM. Okay, how can we do that? Well, if we have a point epsilon bigger than 0, then, of course, we can pick n such that, for every m bigger than n, m divided by p of m is bigger than the minimum of one of the MCM minus this epsilon, right? We can do that because we have assumed the limit above. This is basically the definition of the limit above. The fact that there exists such an n, so we can pick such an n. Once we've picked that n, we can also pick a maximum antichain, namely, the antichain, such of points consisting of all the executions at that point in time. So we have a point in time, p(n), that is the worst time at which, at most n tokens have come out, and then, if we pick an antichain of all executions at that time, so we have an antichain here saying this is a certain point in time. Namely, that point in time, then if we pick any point after that antichain, so if we pick w over here, then what do we know for w? Well, by construction of the antichain, we know that the number of tokens at w is going to be bigger than n because this was a certain point in time, after which we are certain to have n tokens. So, at this point, there must be at least n tokens, and it's also to see that the time at this point is bigger than p(n), or better, p of sigma, p of w. It's the same thing. This is also p of n. Okay, so if we now pick m equal to sigma pw, we call this the number of tokens there we call m. Then we find that sigma p of w divided by sigma t of w is bigger than sigma p of w divided by p of sigma p of w. Basically, because the number of tokens is the worst case number of tokens, the worst time is bigger than this point in time, sorry. That must be bigger than, well, it's actually equal to m divided by pm by the definition of m, and that is bigger than the throughput bound because, well, that's what we already computed over here. Right, so now let's get rid of all these drawings again. This is a proof such as saying that if n over p goes above 1 over the maximum, then also our output will rise above that limit bound. Okay, okay, apparently, I forgot something. What did I forget? There are actually two assumptions that you have to make before this proof actually works out. Can you find them? The most important one, I'll give it to you, is over here. We have secretly assumed that I could find a maximum antichain, such that for all points in the antichain, you are at a certain time. Well, this assumes that, if you go through the prefix order, time is not only a monotone, it's not only rising all the time. But there are also no jumps in time because, if there would be in one execution a jump in the amount of time, then the points in between cannot be touched by a crosscut, so this is a continuity principle. We assume that time is continuous throughout the execution of our interpretation of the graph, and we didn't so before. So, we have to add this assumption before we can accept this proof. Okay, and the other one, well, the other one actually mentioned briefly in the beginning is that p of n is actually defined. This is also not always the case. Suppose that we have, for example, an input and along the different branches this input gives tokens later, and later, and later. Its first token, in one branch, the first token comes in at one second, and the second branch, it comes in at two seconds, and three seconds, four seconds. If we have an infinite number of branches, then the first token actually comes in later and later, and maybe in each of the branches, the average throughput is going to be fine, but p of 1 is not even going to be defined. So, we have to be a bit careful on what kind of inputs we accept, and if we accept the right inputs, then actually the event will be defined for all productions p and all consumptions in the graph. So, in resume, time needs to be continuous, and this is the definition of continuous. For every point in time, there is an antichain that hits that point of time, and for the input, the worst case response time for n tokens need also be defined for all n. If we have that, then we can continue with our proof.

Explorer notre catalogue

Rejoignez-nous gratuitement et obtenez des recommendations, des mises à jour et des offres personnalisées.

Coursera

À propos
Direction
Carrières
Catalogue
Certificats
Certificats MasterTrack™
Diplômes
For Enterprise
For Government
Pour le campus

Communauté

Étudiants
Partenaires
Développeurs
Testeurs bêta
Traducteurs
Blog
Blog Tech

Plus

Conditions
Confidentialité
Aide
Accessibilité
Presse
Contact
Répertoire
Filiales

Apprendre partout

Suivez-nous

© 2020 Coursera Inc. Tous droits réservés.