• Share
  • Email
  • Embed
  • Like
  • Save
  • Private Content
Licentiate Defense Slide
 

Licentiate Defense Slide

on

  • 206 views

Slide I used for my licentiate degree defense

Slide I used for my licentiate degree defense

Statistics

Views

Total Views
206
Views on SlideShare
206
Embed Views
0

Actions

Likes
0
Downloads
0
Comments
0

0 Embeds 0

No embeds

Accessibility

Categories

Upload Details

Uploaded via as Microsoft PowerPoint

Usage Rights

© All Rights Reserved

Report content

Flagged as inappropriate Flag as inappropriate
Flag as inappropriate

Select your reason for flagging this presentation as inappropriate.

Cancel
  • Full Name Full Name Comment goes here.
    Are you sure you want to
    Your message goes here
    Processing…
Post Comment
Edit your comment
  • Good morning every body, welcome to my licentiate seminar on the topic “Contributions to Performance Modeling and Management of Data Centers”So, let begin with the introduction to the topic.
  • Let me start with the introduction to the topic by explaining to you about the networked service, which is the central theme of this thesis.1.) I believe many of you here have used or heard some networked services before, for example, the search-engine services like Google or a social network services like Facebook.===These services have become parts of our daily life, because we interact with them very often, e.g., when we want to search for information, or follow what our friends are doing.2.) We all agree that performance of such services are important. An example of such performance is the response time of services. For example, when you search for information, you want to be interactive, you want it to be fast! You don’t want to wait for half an hour for your query to come back.3.) In order to understand the performance of these services, let look into how these services work. They typically adopt two main components.3.1) The first one is the frontend component are the components that interact directly with the users like interfaces in your laptop, web technologies, mobile applications like in Android or iPhone. 3.2) The second component, which is equally important, which are not very visible to the user but also quite important is the backend components typically residing in a data center. The backend component performs operations for sharing states between applications, for example, web servers that serve data from data bases.4.) Performance management of this second component is the main focus of this thesis. In particular, this thesis focuses on two key problems in management of such backend components residing in a data center.I will go through each of them in subsequent slides.4.)
  • Here is an outline of the talk today.For each problem, I will go through motivation, problem, approach, and results what we worked on.Then, I will summarize related research for both problems, conclude the presentation, and outline some open requests.
  • This work is about large-scale cloud. So, let begin with the motivation of the work, or why should we consider a large-scale cloud?The key reason is that data centers are getting bigger. This is the quote of IDC Research last year that “US Data center is increasing in size but decreasing in numbers.” If we think about this, it also makes senses because it is more cost effectiveness when we have an “economy of scale”.What we see here is the Icloud data center from Apple. With a massive size of 500,000 square feet, it expects to host at least 150,000 machines.It is not just Apple. Many other Internet-based companies like Amazon, Microsoft, Facebook and eBay. All have plan for these kind of massive-scale data centers.So the key message here is that the large-scale center is the trend. Let move on to see what kind of problem that this thesis deal with.
  • In particular, the first problem in this thesis relates to the resource allocation. SoWhat do I mean by resource allocation?I particularly focus on the “application placement problem”or answering the questionHow to select the machine for executing an application that satisfies - Resource demands of all applications in the cloud -> Application need resources to operation for example some CPU cores or some memory for temporary computation- Management objectives of the cloud provider -> The easiest ways to do this is to
  • In particular, the first problem we consider is designing and evaluating the resource allocation system that
  • So, this is our optimization problem. Essentially, it minimizes f(t) under the following constraints.
  • A round-based protocol that relies on pairwise interactions between nodes to accomplish a global objectiveEach node executes the same code and the size of the exchanged message is limited
  • Here is an outline of the talk today.For each problem, I will go through motivation, problem, approach, and results what we worked on.Then, I will summarize related research for both problems, conclude the presentation, and outline some open requests.

Licentiate Defense Slide Licentiate Defense Slide Presentation Transcript

  • Contributions to Performance Modeling and Management of Data Centers Rerngvit Yanggratoke Advisor: Prof. Rolf Stadler Opponent: Prof. Filip De Turck Licentiate Seminar October 25, 2013 1
  • Introduction Networked services – search and social-network services Performance of such services is important Data center Internet Backend component Performance management of backend components in a data center 1. Resource allocation for a large-scale cloud environment 2. Performance modeling of a distributed key-value store 2
  • Outline 1. Resource allocation for a large-scale cloud environment • Motivation • Problem • Approach • Evaluation results • Related research / publications 2. Performance modeling of a distributed key-value store • Motivation • Problem • Approach • Evaluation results • Related research / publications – Contributions – Open questions 3
  • Motivation – Large-scale Cloud “U.S. Data centers Growing in Size But Declining in Numbers” IDC Research, 2012. • Apple's data center in North Carolina, USA (500’000 feet2) • Expected to host 150’000+ machines • Amazon data center in Virginia, USA (about 300’000 machines) • Microsoft's mega data center in Dublin, Ireland (300’000 feet2) • Facebook planned for a massive data center in Iowa, USA (1.4M feet2) • eBay's data center in Utah, USA (at least 240’000 feet2) 4
  • Resource Allocation Select the machine for executing an application that satisfies: • Resource demands of all applications in the cloud • Management objectives of the cloud provider Resource allocation system computes the solution to the problem 5
  • Problem & Approach Resource allocation system that supports – Joint allocation of compute and network resources – Generic and extensible for different management objectives – Scalable operation (> 100’000 machines) – Dynamic adaptation to changes in load patterns Approach – We assume the full-bisection-bandwidth network – Formulate the problem as an optimization problem – Distributed protocols - gossip-based algorithms 6
  • An Objective Function f(t) An objective function f(t) expresses a management objective “Smaller f(t) moves the system closer to achieve the objective” Balance load objective f (t) = å k q q Î{w ,g , l } Energy efficiency objective åq q c u n n (t) 2 nÎN Fairness objective æ ö d v f (t) = - ç å kq å q m q m (t) ÷ ç ÷ q Î{w ,g } mÎM (t ) è ø f (t) = å pn (t) Pn(t) = energy of machine n at time t Service differentiation objective æ ö q d v f (t) = - ç å k å q m q m (t) ÷ ç ÷ èq Î{w ,g } mÎb (t ) ø q Î {w, g, l} = {CPU, Memory, Network} N = set of machines, M = set of applications 7
  • Optimization Problem Minimize f(t) Subject to: net (1) q n (t) + å r q m (t) £ qnc for q Î {w, g , l } and n Î N mÎAn (t ) d m r (2) qm (t) ³ q for q Î {w, g, l } and m Î M • (1) are capacity constraints • (2) are demand constraints • Choose a specific f(t) states the problem for a specific management objective • This problem is NP-hard • We apply a heuristic solution • How to distribute the computation? 8
  • Gossip Protocol • A round-based protocol that relies on pairwise interactions between nodes to accomplish a global objective • Each node executes the same code and the size of the exchanged message is limited “During a gossip interaction, move an application that minimizes f(t)” Balanced load objective 9
  • Generic Resource Management Protocol (GRMP) A generic and scalable gossip protocol for resource allocation <- f(t) is applied here 10
  • Evaluation Results Balanced load objective Energy efficiency objective Protocol Ideal Protocol Ideal Relative power consumption 1 Unbalance 0.8 0.6 0.4 0.2 0 0.95 0.75 0.55 0.35 0.15 CPU Load Factor 0.3 0.6 0.9 1.2 1.5 1 0.8 0.6 0.4 0.2 0 0.95 0.75 0.55 0.35 0.15 CPU Load Factor Network Load Factor 0.3 0.6 0.9 1.2 1.5 Network Load Factor Scalability: • System size up to 100’000 • Other parameters do not change 11
  • Related Research • Centralized resource allocation system – (D. Gmach et al., 2008), (A. Verma et al., 2009), (C. Guo et al., 2010), (H. Hitesh et al., 2011), (V. Shrivastava et al., 2011), (D. Breitgand and A.Epstein, 2012), (J.W. Jiang et al., 2012), (O. Biran et al. 2012), … • Distributed resource allocation system – (G. Jung et al., 2010), (Yazir et al., 2010), (D. Jayasinghe et al., 2011), (E. Feller et al., 2012), this thesis • Initial placement – (A. Verma et al., 2009), (M. Wang et al., 2011), (D. Jayasinghe et al., 2011), (X. Zhang et al., 2012) • Dynamic placement – (F. Wuhib et al., 2012), (D. Gmach et al., 2008), (Yazir et al., 2010), (J. Xu and Jose Fortes, 2011) , this thesis 12
  • Publications • R. Yanggratoke, F. Wuhib and R. Stadler, “Gossip-based resource allocation for green computing in large clouds,” In Proc. 7th International Conference on Network and Service Management (CNSM), Paris, France, October 24-28, 2011. • F. Wuhib, R. Yanggratoke, and R. Stadler, “Allocating Compute and Network Resources under Management Objectives in Large-Scale Clouds,” Journal of Network and Systems Management (JNSM), pages 126, 2013. 13
  • Outline 1. Resource allocation for a large-scale cloud environment • Motivation • Problem • Approach • Evaluation results • Related research / publications 2. Performance modeling of a distributed key-value store • Motivation • Problem • Approach • Evaluation results • Related research / publications – Contributions – Open questions 14
  • Spotify Backend A distributed key-value store • Motivation: low latency is key to the Spotify service 15
  • Problem & Approach • Development of analytical models for – Predicting response time distribution – Estimating capacity under different object allocation policies • Random policy • Popularity-aware policy • The models should be – Accurate for Spotify’s operational range – Obtainable – Simple – Efficient • Approach – Simplified architecture – Probabilistic and stochastic modeling techniques 16
  • Simplified Architecture for a Particular Site • • • • Model only Production Storage AP selects a storage server uniformly at random Ignore network and access point delays Consider steady-state conditions and Poisson arrivals 17
  • Model for Response Time Distribution Model for a single storage server Probability that a request to a server is served below a latency t is ( Pr(T £ t) = q +(1- q) 1- e - md (1-(1-q)l /md nd )t ) Model for a cluster of storage servers Probability that a request to the cluster is served below a latency t is 1 l Pr(Tc £ t) = f (t, nd , m d,s , c , qs ) å S sÎS S 18
  • Model for Estimating Capacity under Different Object Allocation Policies Capacity Ω : Max request rate to a server before a QoS is not satisfied Ωc: Max request rate to the cluster such that the request rate to each server is at most Ω Capacity of a cluster under the popularity-aware policy Wc (popularity) = W× S Capacity of a cluster under the random policy Oa 1/a O + a -1 Wc (random) = W × Om a 1/a Om + a -1 When Om O » + S 2 O ln S æ log log S ç1ç S 2 log S è ö ÷ ÷ ø ( , Assuming O >> S log S 19 ) 3
  • Evaluation Results – Response Time Dist. Spotify Server Spotify Cluster 20
  • Evaluation Results – Estimating Capacity Number Random policy of objects Popularity-aware policy Measurement Model Error (%) Measurement Model Error (%) 1250 166.50 176.21 5.83 190.10 200 5.21 1750 180.30 180.92 0.35 191.00 200 4.71 2500 189.70 182.30 3.90 190.80 200 4.82 5000 188.40 186.92 0.78 191.00 200 4.71 10000 188.60 190.39 0.95 192.40 200 3.95 21
  • Related Research • Distributed key-value store – Amazon Dynamo (G. Decandia et al., 2007), Cassandra (A. Lakshman, 2010), Riak, HyperDec, Scalaris • Performance modeling of a storage device – (J.D. Garcia et al., 2008), (A.S. Lebrecht., 2009), (F. Cady et al., 2011), (S. Boboila and P. Desnoyers, 2011), (P. Desnoyers, 2012), (S. Li and H.H. Huang, 2010) • White-box vs. black-box performance model for a storage system – Whitebox: (A.S. Lebrecht., 2009), (F. Cady et al., 2011), (S. Boboila and P. Desnoyers, 2011), (P. Desnoyers, 2012), this thesis – Blackbox: (J.D. Garcia et al., 2008), (S. Li and H.H. Huang, 2010), (A. Gulati et al., 2011) 22
  • Publications • R. Yanggratoke, G. Kreitz, M. Goldmann and R. Stadler, “Predicting response times for the Spotify backend,” In Proc. 8th International conference on Network and service management (CNSM), Las Vegas, NV, USA, October 2226, 2012. Best Paper award. • R. Yanggratoke, G. Kreitz, M. Goldmann, R. Stadler and V. Fodor, “On the performance of the Spotify backend,” accepted to Journal of Network and Systems Management (JNSM), 2013. 23
  • Contributions • We designed, developed, and evaluated a generic protocol for resource allocation that – supports joint allocation of compute and network resources – Is generic and extensible for different management objectives – enables scalable operation (> 100’000 machines) – supports dynamic adaptation to changes in load patterns • We designed, developed, and evaluated performance models for predicting response time distribution and capacity of a distributed key-value store – Simple yet accurate for Spotify’s operational range – Obtainable – Efficient 24
  • Open Questions for Future Research • Resource allocation for a large-scale cloud environment – Centralized vs. decentralized resource allocation – Multiple data centers & telecom clouds • Performance modeling of a distributed key-value store – Black-box models for performance predictions – Online performance management using analytical models 25