Computer Laboratory

Large-Scale Data Processing and Optimisation (2017-2018 Michaelmas Term)

LSDPO - R244

review_log

Open Source Projects

Reading Club Papers

Contact

 

 

 

 

 

 

 

  

 

Overview

This module provides an introduction to large-scale data processing, optimisation, and the impact on computer system's architecture. Large-scale distributed applications with high volume data processing such as training of machine learning will grow ever more in importance. Supporting the design and implementation of robust, secure, and heterogeneous large-scale distributed systems is essential. To deal with distributed systems with a large and complex parameter space, tuning and optimising computer systems is becoming an important and complex task, which also deals with the characteristics of input data and algorithms used in the applications. Algorithm designers are often unaware of the constraints imposed by systems and the best way to consider these when designing algorithms with massive volume of data. On the other hand, computer systems often miss advances in algorithm design that can be used to cut down processing time and scale up systems in terms of the size of the problem they can address. Integrating machine learning approaches for system optimisation will also be explored in this course. On completion of this module, the students should:

  • Understand key concepts of scalable data processing approaches in future computer systems. Obtain a clear understanding of building distributed systems using data centric programming and large-scale data processing.
  • Understand a large and complex parameter space in computer system's optimisation and applicability of Machine Learning approach.

Module Structure

The module consists of 8 sessions, with 5 sessions on specific aspects of large-scale data processing research. Each session discusses 3-4 papers, led by the assigned students. One session is a hands-on tutorial on MapReduce using data flow programming and/or Deep Neural Networks using Google TensorFlow. The 1st session advises on how to read/review a paper together with a brief introduction on different perspectives in large-scale data processing and optimisation. The last session is dedicated to the student presentation of open-source project studies. One guest lecture is planned, covering inspiring current research on stream processing systems.

Schedule and Reading List

We’ll meet in SW01 (or FW11) every Tuesday (from October 10 to November 28) in 2017. The time slot is 15:00-17:00.1

 2017/10/10 Session 1: Introduction to Large-Scale Data Processing and Optimisation

  • Introduction to R244 (Slides)
    • Assignment details
    • Guidance of how to read/review/present a paper
    • Guidance to Open Source Project
  • Overview of technologies for Big Data Processing (Slides)

 2017/10/17 Session 2: Data flow programming: Map/Reduce to TensorFlow

  • Data flow programming, Cluster Computing

1. Yuan Yu, Michael Isard, D. Fetterly, M. Budiu, U. Erlingsson, P.K. Gunda, J. Currey: DryadLINQ: A System for General-Purpose Distributed Data-Parallel Computing Using a High-Level Language, OSDI, 2008.

Lukasz Dudziak (slides)
2. M. Zaharia, M. Chowdhury, T. Das, A. Dave, J. Ma, M. McCauley, M. Franklin, S. Shenker, I. Stoica: Resilient Distributed Datasets: A Fault-Tolerant Abstraction for In-Memory Cluster Computing, NSDI, 2013.

3. Peter Alvaro, Tyson Condie, Neil Conway, Khaled Elmeleegy, Joseph M. Hellerstein, Russell Sears: Boom analytics: exploring data-centric, declarative programming for the cloud, Eurosys 2010.

Stefanos Laskaridis (slides)
4. J. Dean, S. Ghemawat: MapReduce: Simplified Data Processing on Large Clusters, OSDI, 2004.

Ioana Bica (slides)
5. Derek Murray, Malte Schwarzkopf, Christopher Smowton, Steven Smith, Anil Madhavapeddy and Steven Hand: Ciel: a universal execution engine for distributed data-flow computing, NSDI 2011. 

6. Naiad

Frank McSherry's Talk on Differential Dataflow is here.

6.1. Frank McSherry, Rebecca Isaacs, Michael Isard, and Derek G. Murray, Composable Incremental and Iterative Data-Parallel Computation with Naiad, no. MSR-TR-2012-105, 2012. 

Jesse Mu (slides)
6.2. D. Murray, F. McSherry, R. Isaacs, M. Isard, P. Barham, M. Abadi: Naiad: A Timely Dataflow System, SOSP, 2013. 

7. P. Bhatotia, A. Wieder, R. Rodrigues, U. A. Acar, and R. Pasquini: Incoop: MapReduce for incremental computation, ACM SOCC, 2011.

Nathaniel McAleese-Park (slides)
8. M. Abadi et al. Tensorflow: A system for large-scale machine learning. OSDI, 2016.

    M. Abadi et al.: TensorFlow: Large-Scale Machine Learning on Heterogeneous Distributed Systems, Preliminary White Paper, 2015.

9. M. Looks et al.: Deep Learning with Dynamic Computation Graphs, ICLR, 2017.

 2017/10/24 Session 3: Large-scale graph data processing: storage, processing model and parallel processing

  • Scalable distributed processing of graph structured data, processing model, and programming model

George Wort (slides)
1. G. Malewicz, M. Austern, A. Bik, J. Dehnert, I. Horn, N. Leiser, and G. Czajkowski: Pregel: A System for Large-Scale Graph Processing, SIGMOD, 2010.

2. Z. Qian, X. Chen, N. Kang, M. Chen, Y. Yu, T. Moscibroda, Z.Zhang: MadLINQ: large-scale distributed matrix computation for the cloud, EuroSys, 2012.

3. Y. Low,  J. Gonzalez, A. Kyrola, D. Bickson, C. Guestrin, J. Hellerstein: Distributed GraphLab: A Framework for Machine Learning and Data Mining in the Cloud, VLDB, 2012.

Joshua Send (slides)
4. J. Gonzalez, Y. Low, H. Gu, D. Bickson, and C. Guestrin: Powergraph: distributed graph-parallel computation on natural graphs. OSDI, 2012.

Thomas Parks (slides)
5. .J. Shun and G. Blelloch: Ligra: A Lightweight Graph Processing Framework for Shared Memory, PPoPP, 2013. 

6. J.  Gonzalez, R. Xin, A. Dave, D. Crankshaw, M. Franklin, I. Stoica: GraphX: Graph Processing in a Distributed Dataflow Framework, OSDI, 2014. 

7. B. Shao,  H. Wang, Y. Li: Trinity: A Distributed Graph Engine on a Memory Cloud, SIGMOD, 2013.

8. A. Kyrola and G. Blelloch: Graphchi: Large-scale graph computation on just a PC, OSDI, 2012.  

Stella Lau (slides)
9. A. Roy, I. Mihailovic, W. Zwaenepoel:   X-Stream: Edge-Centric Graph Processing using Streaming Partitions, SOSP, 2013.

10. A. Roy, L. Bindschaedler, J. Malicevic and W. Zwaenepoel: Chaos: Scale-out Graph Processing from Secondary Storage , SOSP, 2015.

Alex Gubbay (slides)
11. F. McSherry, M. Isard and D. Murray: Scalability! But at what COST? , HOTOS, 2015.

12. X. Hu, Y. Tao, C. Chung:  Massive Graph Triangulation, SIGMOD, 2013.

13. W. Xie, G. Wang, D.Bindel, A. Demers, J. Gehrke:  Fast Iterative Graph Computation with Block Updates, VLDB, 2014.

Tudor Tiplea (slides)
14. S. Hong, H. Chafi, E. Sedlar, K.Olukotun: Green-Marl: A DSL for Easy and Efficient Graph Analysis, ASPLOS, 2012.

15. D. Prountzos, R. Manevich, K. Pingali: Elixir: A System for Synthesizing Concurrent Graph Programs, OOPSLA, 2012.

16. D. Nguyen, A. Lenharth, K. Pingali: A Lightweight Infrastructure for Graph Analytics, SOSP 2013.

17. D. Merrill, M. Garland, A. Grimshaw: Scalable GPU Graph Traversal, PPoPP, 2012.

18. A. Gharaibeh, E. Santos-Neto, L. Costa, M. Ripeanu:  Efficient Large-Scale Graph Processing on Hybrid CPU and GPU Systems, IEEE TPC, 2014.

 2017/10/31 Session 4: Map/Reduce and Deep Neural Network using TensorFlow Hands-on Tutorial  

 2017/11/7 Session 5: Stream Data Processing and Data/Query Model 

  • Data and continuous query in steam data processing

Guest lecture: Peter Pietzuch (Imperial College London)

Title: Hybrid Data Stream Processing with Accelerators (slides)

Abstract: A stream processing model is well-suited for the processing of large amounts of data with low latency. As modern servers have   become  heterogeneous, often combining multi-core CPUs with many-core GPUs, there is potential to improve the performance of data-intensive stream processing applications. For today's relational stream processing engines to exploit heterogeneous servers, they must execute streaming queries with sufficient data-parallelism to fully utilise all available heterogeneous processors, and decide how to use each in the most effective way. In this talk, I will introduce streaming applications as a new challenging workload for heterogeneous servers. I will then describe SABER, a new design for a hybrid stream processing engine for CPUs and GPUs. Under a hybrid model, SABER executes streaming SQL queries in a data-parallel fashion using all available CPU and GPU cores simultaneously. Instead of statically assigning query operators to heterogeneous processors, SABER employs an adaptive scheduling strategy that opportunistically assigns an operator to the processor that will complete first. Our experiments show that SABER's hybrid stream processing model can aggregate the performance of multiple heterogeneous processors.

  • Reading Club

1. T. Akidau, A. Balikov, K. Bekiroglu, S. Chernyak, J. Haberman, R. Lax, S. McVeety, D. Mills, P. Nordstrom, S. Whittle: MillWheel: Fault-Tolerant Stream Processing at Internet Scale , VLDB, 2013.

2. V. Zaharia, T. Das, H. Li, T. Hunter, S. Shenker, I. Stoica: Discretized Streams: Fault-Tolerant Streaming Computation at Scale, SOSP, 2013.

3. R. Fernandez, M. Migliavacca, E. Kalyvianaki, P. Pietzuch: Making State Explicit for Imperative Big Data Processing, USENIX ATC, 2014.

4. D. Abadi, Y. Ahmad, M. Balazinska et al. : The Design of the Borealis Stream Processing Engine, CIDR, 2005.
 
5. S. Babu, J. Widom: Continuous Queries over Data Streams, SIGMOD Record 30(3), 2001.  

6. B.Gedik, H. Andrade, K. Wu, P. Yu, and M. Doo: SPADE: the system S Declarative Stream Processing Engine , SIGMOD. 2008.  

Thomas Brady (slides)
7. T. Akidau, R. Bradshaw, C. Chambers, S. Chernyak, R.  Fernandez-Moctezuma, R. Lax, S. McVeety, D. Mills, F. Perry, E. Schmidt, S. Whittle: The Dataflow Model: A Practical Approach to Balancing Correctness, Latency, and Cost in Massive-Scale, Unbounded, Out-of-Order Data Processing, VLDB, 2015.

8. R. Cheng, J. Hong, A. Kyrola, Y. Miao, X. Weng, M. Wu, F. Yang, L. Zhou, F. Zhao, E. Chen: Kineograph: Taking the Pulse of a Fast-Changing and Connected World, EuroSys, 2012. 

Yashovardhan Sharma (slides)
9. A. Floratou et al.: Dhalion: self-regulating stream processing in Heron, VLDB, 2017.

 2017/11/14 Session 6: Machine Learning for Optimisation of Computer Systems  

  • Space for using machine learning on optimisation in computer systems

1. Jeffrey Dean, Greg Corrado, Rajat Monga, Kai Chen, Matthieu Devin, Quoc V. Le, Mark Z. Mao, Marc’Aurelio Ranzato, Andrew W. Senior, Paul A. Tucker, Ke Yang, and Andrew Y. Ng. Large scale distributed deep networks. NIPS, 2012.

2. G. Venkates et al.: Accelerating Deep Convolutional Networks using low-precision and sparsity,  ICASSP, 2017.

3. V. Mnih et al.: Asynchronous Methods for Deep Reinforcement Learning, ICML, 2016.

Jesse Mu (slides)
4 V. Dalibard, M. Schaarschmidt, and E. Yoneki: BOAT: Building Auto-Tuners with Structured Bayesian Optimization, WWW, 2017. 

5. J. Ansel et al. Opentuner: an extensible framework for program autotuning. PACT, 2014.

Tudor Tiplea (slides)
6. B. Bodin, L. Nardi, MZ Zia et al.: Integrating Algorithmic Parameters into Benchmarking and Design Space Exploration in 3D Scene Understanding, PACT, 2016. 

7. J. Ansel et al. Petabricks: A language and compiler for algorithmic choice. In Proceedings of the 2009 ACM SIGPLAN Conference on Programming Language Design and Implementation, PLDI, 2009.

Thomas Park (slides)
8. V. Mnih et al.: Playing Atari with Deep Reinforcement Learning, NIPS, 2013.

9. J. Snoek, H. Larochelle, and R. Adams: Practical Bayesian Optimization of Machine Learning Algorithms, NIPS, 2012.

10. B. Teabe et al.: Application-specific quantum for multi-core platform scheduler, EuroSys, 2016.

Nathaniel McAleese-Park (slides)
11. G. Tesauro et al.: A Hybrid Reinforcement Learning Approach to Autonomic Resource Allocation, ICAC, 2006.

Joshua Send (slides)
12. A. Valadarsky et al.: A Machine Learning Approach to Routing, arXiv, 2017.

Alex Gubbay (slides)
13. N. Lane et al.: DeepX : A Software Accelerator for Low-Power Deep Learning Inference on Mobile Devices, IPSN, 2016.

14. K. Arulkumaran et al.: A Brief Survey of Deep Reinforcement Learning, IEEE Journal of Signal Processing, 2017.  

15. F. Hutter et al.: Algorithm runtime prediction: Methods&evaluation, Elsevier J. AI, 2014.

 2017/11/21 Session 7: Task scheduling, Performance, and Resource Optimisation

  • Optimisation examples in coputer systems (e.g. scheduling, resource allocation...)

Stella Lau (slides)
1. A. Mirhoseini et al.: Device Placement Optimization with Reinforcement Learning, ICML, 2017.

2. F. Yang et al.: LFTF: A Framework for Efficient Tensor Analytics at Scale, VLDB, 2017.

3. Y. You et al.: Scaling Deep Learning on GPU and Knights Landing clusters, SC, 2017.

Lukasz Dudziak (slides)
4. I. Gog, M. Schwarzkopf, A. Gleave, R. Watson, S. Hand: Firmament: fast, centralized cluster scheduling at scale, OSDI, 2016.

5. O. Alipourfard et al.: CherryPick: Adaptively Unearthing the Best Cloud Configurations for Big Data Analytics, NSDI, 2017.

6. C. Delimitrou et al.: Quasar: Resource-Efficient and QoS-Aware Cluster Management, ASPLOS, 2014.

7. H. Mao et al.: Neural Adaptive Video Streaming with Pensieve, SIGCOMM, 2017.

8. S. Venkataraman et al.: Ernest: Efficient Performance Prediction for Large-Scale Advanced Analytics, NSDI, 2016.

9. N. Roy et al.: Efficient Autoscaling in the Cloud using Predictive Models for Workload Forecasting, CLOUD, 2011.

10. K. LaCurts et al.: Cicada: Introducing Predictive Guarantees for Cloud Networks, HOTCLOUD, 2014.

11. M. Carvalho et al.: Long-term SLOs for reclaimed cloud computing resources, SOCC, 2014.

12. H. Hoffmann et al.: Dynamic Knobs for Responsive Power-Aware Computing, Asplos, 2011.

13. N.J. Yadwadkar, B. Hariharan, J. Gonzalez and R. Katz: Faster Jobs in Distributed Data Processing using Multi-Task Learning, SDM, 2015.

14. X. Dutreih et al.: Using Reinforcement Learning for Autonomic Resource Allocation in Clouds: Towards a Fully Automated Workflow, ICAS, 2011.

15. J. Eastep et al.: Smart Data Structures: An Online Machine Learning Approach to Multicore Data Structures, ICAC, 2011.

16. H. Hoffmann et al.: SEEC: A Framework for Self-aware Management of Multicore Resources, MIT Technical Report, 2011.

17. E. Ipek et al.: Self-Optimizing Memory Controllers: A Reinforcement Learning Approach, ISCA, 2008.

Stefanos Laskaridis (slides)
18. Y. Kang et al.: Neurosurgeon: Collaborative Intelligence Between the Cloud and Mobile Edge, ASPLOS, 2017.

19. S. Teerapittayanon et al.: Distributed Deep Neural Networks over the Cloud, the Edge and End Devices, ICDCS, 2017.

20. B. Zoph et al.: Learning Transferable Architectures for Scalable Image Recognition, arXiv, 2017.

21. D. Golovin et al.: Google Vizier: A Service for Black-Box Optimization, KDD, 2017.

22. D. Baylor et al.: TFX: A TensorFlow-Based Production-Scale Machine Learning Platform, KDD, 2017.

Yashovardhan Sharma (slides)
23. H. Mao et al.: Resource Management with Deep Reinforcement Learning, HotNets, 2016.

Thomas Brady (slides)
24. M. Raghu et al.: On the Expressive Power of Deep Neural Networks, PMLR, 2017. 

Ioana Bica (slides)
25. D. Aken et al.: Automatic Database Management System Tuning Through Large-scale Machine Learning, SIGMOD, 2017. 

George Wort (slides)
26. A. Pavlo et al.: Self-Driving Database Management Systems, CIDR, 2017.

 2017/11/28 Session 8: Presentation of Open Source Project Study

  • Start @15:00 in SW01
  • Presentation of Open Source Project Study by all (~9 minutes of presentation plus Q&A for each presentation)
    1. 15:00 Ioana Bica (Tensorflow, PyTorch) Distributed Neural Network Training and Data Flow Graph Construction: TensorFlow vs PyTorch (slides)
    2. 15:10 Thomas Brady (TensorForce) Implementing Cross Entropy Method for TensorForce (slides)
    3. 15:20 Łukasz Dudziak (TensorFlow) Automation for distributing computations in tensorflow (slides)
    4. 15:30 Alexander Gubbay (Apache Storm) Real Time Graph Modelling of the London Bus Network (slides)
    5. 15:40 Stefanos Laskaridis (PyTorch and Tensorflow) Text Sentiment Analysis with rNN on the IMDB dataset (slides)
    6. 15:50 Stella Lau (GraphX) Word sense induction using GraphX (slides)
    7. 16:00 Nathaniel McAleese-Park (Naiad) Timely dataflow with the GPU in Rust (slides)
    8. 16:10 Jesse Mu (Snorkel (Stanford)) Learning causal relations from weak supervision in scientific literature (slides)
    9. 16:20 Thomas Parks (Naiad) Dataflow raytracer using Naiad (slides)
    10. 16:30 Joshua Send (PowerGraph) Tradeoffs Between Synchronous and Asynchronous Evaluation in PowerGraph (slides)
    11. 16:40 Yashovardhan Sharma (Spark) Analysing Large-Scale Data Processing Engines : A Comparative Study of Spark and MapReduce
    12. 16:50 Tudor Tiplea (TensorForce) Evolution Strategies using TensorForce (slides)
    13. 17:00 George Wort (Naiad) Using Naiad to Analyze Twitter Data in Batch and Real-time (slides)

     Wrap-up Discussion (slides)              

Coursework 1 (Reading Club)

The reading club will require you to read between 1 and 3 papers every week. You need to fill out simple review_log (MS word format, text format) prior to each session and email me by the end of Sunday. The minimum requirement of review_log is one per session, but you can read as many as you want and fill the review_log for each paper you read. review_log is not marked but 'tick'.

At each session, 3 - 4 papers are selected under the session topic, and if you are assigned to present your review work, please prepare 15-20 minutes slides for presenting your review work. Your presented material should also be emailed by the following day Wednesday. You would present your review work approximately twice during the course. The paper includes following two types and you can focus on the specified aspects upon reviewing the paper.

  1. Full length papers 
    • What is the significant contribution?
    • What is the difference from the existing works?
  2. Short length papers 
    • What is the novel idea?
    • What is required to complete the work?

 Coursework 2 (Reports)

The following three reports are required, which could be extended from the reading assignment of the reading club or a different one within the scope of data centric networking.

  1. Review report on a full length of paper (max 1800 words)
    • Describe the contribution of paper in depth with criticism
    • Crystallise the significant novelty in contrast to the other related work
    • Suggestion for future work
  2. Survey report on sub-topic in data centric networking (~1800 - max 2000 words)
    • Pick up to 5 papers as core papers in your survey scope
    • Read the above and expand your reading through related work
    • Comprehend your view and finish as your survey paper
    • See how to write a survey paper
  3. Project study and exploration of a prototype (max 2500 words)
    • What is the significance of the project in the research domain?
    • Compare with the similar and succeeding projects
    • Demonstrate the project by exploring its prototype
    • Please email your project selection (MS word format or text format <150 words) by November 1, 2017.
    • Project presentation on November 28, 2017.

The report 1 should be handed in by the end of 5th week (November 10, 2017 - 16:00) and the report 2 by 7th week (November 24, 2017 - 16:00). The report 3 should be by the end of the Michaelmas term (January 16,  2018 - 16:00 - but if you could finish by 20th of December, that will be good!).

 Assessment

The final grade for the course will be provided as a letter grade or percentage and the assessment will consist of two parts:

  1. 20%: for a reading club (Presentation and participation + tick of review_log and hands-on tutorial)
  2. 80%: for the three reports
    • 20%: Intensive review report
    • 25%: Survey report
    • 35%: Project study

Open Source Projects

See the candidates of Open Source Projects in data centric networking. The list is not exhausted. If you take anything other than the one in the list, please discuss with me. The purpose of this assignment is to understand the prototype of the proposed architecture, algorithms, and systems through running an actual prototype and present/explain to the other people how the prototype runs, any additional work you have done including your own applications and setup process of the prototype. This experience will give you better understanding of the project. These Open Source Projects come with a set of published papers and you should be able to examine your interests in the paper through running the prototype. Some projects are rather large and may require extensive environment and time; make sure you are able to complete this assignment.

How to Read/Review a Paper

The following papers aid how to read/review a paper.

Further supplement: see ‘how to read/review a paper’ section in Advanced Topics in Computer Systems by Steven Hand.

Presentations

Presentations should be about 15-20 minutes long, where you need to cover the following aspects.

  1. What are the background and the problem domain of the paper? What is the motivation of the presented work? What is the difference from the existing works?  What is the novel idea? How did the paper change/unchange the research in the research community?

  2. What is the significant contribution? How did the authors tackle the problem? Did the authors obtain expected result from their trial?

  3. How do you like the paper and why? What is the takeaway message to you (and to research community)? What is required to complete the work?

The following document aids in presenting a review.

How to write a survey paper

A survey paper provides the readers with an exposition of existing work that is comprehensive and organized. It must expose relevant details associated in the surveying area, but it is important to keep a consistent level of details and to avoid simply listing the different works. Thus a good survey paper should demonstrate a summary of recent research results in a novel way that integrates and adds understanding to work in the field. For example, you can take an approach by classifying the existing literature in your own way; develop a perspective on the area, and evaluate trends. Thus, after defining the scope of your survey, 1) classify and organize the trend, 2) critical evaluation of approaches (pros/cons), and 3) add your analysis or explanation (e.g. table, figure). Also adding reference and pointer to further in-depth information is important (summary from Rich Wolski’s note).

 Papers for OS Principles (Distributed Storage and Deterministic Parallelism)

  • Following papers will help you to understand distributed storage and parallelism.
  • Systems Research and System Design
1.  B. Lampson: Hints for Computer Systems Design (Revised), ACM OSR 1983.

  • Distributed Storage
2. S. Ghemawat, H. Gobioff, and S. Leung: The Google File System, ACM SOSP 2003.
3. F. Chang et al: BigTable: A Distributed Storage System for Structured Data, USENIX OSDI 2006.
4. G. DeCandia et al:  Dynamo: Amazon's Highly Available Key-value Store, ACM SOSP 2007.

  • Deterministic Parallelism
5. J. Devietti et al: DMP: Deterministic Shared Memory Multiprocessing, ACM ASPLOS 2009.
6. A. Aviram, et al: Efficient System-Enforced Determistic Parallelism, USENIX OSDI 2010.
7. T. Liu et al: Dthreads: Efficient and Determistic Multithreading, ACM SOSP 2011.

Contact Email

Please email to eiko.yoneki@cl.cam.ac.uk for your submission of course work or any question.