FAWN: Fast Array of Wimpy Nodes Developed By D. G. Andersen, J. Franklin, M. Kaminsky, A. Phanishayee, L. Tan, V. Vasudevan Presented by Peter O. Oliha.

Slides:



Advertisements
Similar presentations
M AINTAINING L ARGE A ND F AST S TREAMING I NDEXES O N F LASH Aditya Akella, UW-Madison First GENI Measurement Workshop Joint work with Ashok Anand, Steven.
Advertisements

Ion Stoica, Robert Morris, David Karger, M. Frans Kaashoek, Hari Balakrishnan MIT and Berkeley presented by Daniel Figueiredo Chord: A Scalable Peer-to-peer.
1 Parallel Scientific Computing: Algorithms and Tools Lecture #2 APMA 2821A, Spring 2008 Instructors: George Em Karniadakis Leopold Grinberg.
Peer-to-Peer (P2P) Distributed Storage 1Dennis Kafura – CS5204 – Operating Systems.
SILT: A Memory-Efficient, High-Performance Key-Value Store
CSE 486/586, Spring 2012 CSE 486/586 Distributed Systems New Trends in Distributed Storage Steve Ko Computer Sciences and Engineering University at Buffalo.
FAWN: Fast Array of Wimpy Nodes A technical paper presentation in fulfillment of the requirements of CIS 570 – Advanced Computer Systems – Fall 2013 Scott.
Common approach 1. Define space: assign random ID (160-bit) to each node and key 2. Define a metric topology in this space,  that is, the space of keys.
Log-Structured Memory for DRAM-Based Storage Stephen Rumble, Ankita Kejriwal, and John Ousterhout Stanford University.
Energy-efficient Cluster Computing with FAWN: Workloads and Implications Vijay Vasudevan, David Andersen, Michael Kaminsky*, Lawrence Tan, Jason Franklin,
Authors: David G. Andersen et al. Offense: Chang Seok Bae Yi Yang.
Lecture 6 – Google File System (GFS) CSE 490h – Introduction to Distributed Computing, Winter 2008 Except as otherwise noted, the content of this presentation.
FAWN: A Fast Array of Wimpy Nodes Authors: David G. Andersen et al. Offence: Jaime Espinosa Chunjing Xiao.
FAWN: A Fast Array of Wimpy Nodes Presented by: Aditi Bose & Hyma Chilukuri.
Lecture 3: A Case for RAID (Part 1) Prof. Shahram Ghandeharizadeh Computer Science Department University of Southern California.
METU Department of Computer Eng Ceng 302 Introduction to DBMS Disk Storage, Basic File Structures, and Hashing by Pinar Senkul resources: mostly froom.
Large Scale Sharing GFS and PAST Mahesh Balakrishnan.
FAWN: A Fast Array of Wimpy Nodes Presented by: Clint Sbisa & Irene Haque.
Virtual Memory Deung young, Moon ELEC 5200/6200 Computer Architecture and Design Lectured by Dr. V. Agrawal Lectured by Dr. V.
Peer To Peer Distributed Systems Pete Keleher. Why Distributed Systems? l Aggregate resources! –memory –disk –CPU cycles l Proximity to physical stuff.
The Google File System.
Wide-area cooperative storage with CFS
Virtual Memory By: Dinouje Fahih. Definition of Virtual Memory Virtual memory is a concept that, allows a computer and its operating system, to use a.
Storage: Scaling Out > Scaling Up? Ankit Singla Chi-Yao Hong.
INTRODUCTION TO PEER TO PEER NETWORKS Z.M. Joseph CSE 6392 – DB Exploration Spring 2006 CSE, UT Arlington.
Roger ZimmermannCOMPSAC 2004, September 30 Spatial Data Query Support in Peer-to-Peer Systems Roger Zimmermann, Wei-Shinn Ku, and Haojun Wang Computer.
Application-driven Energy-efficient Architecture Explorations for Big Data Authors: Xiaoyan Gu Rui Hou Ke Zhang Lixin Zhang Weiping Wang (Institute of.
Modularizing B+-trees: Three-Level B+-trees Work Fine Shigero Sasaki* and Takuya Araki NEC Corporation * currently with 1st Nexpire Inc.
CSE 486/586, Spring 2013 CSE 486/586 Distributed Systems Distributed File Systems Steve Ko Computer Sciences and Engineering University at Buffalo.
Cooperative File System. So far we had… - Consistency BUT… - Availability - Partition tolerance ?
RAMCloud: A Low-Latency Datacenter Storage System Ankita Kejriwal Stanford University (Joint work with Diego Ongaro, Ryan Stutsman, Steve Rumble, Mendel.
Chord & CFS Presenter: Gang ZhouNov. 11th, University of Virginia.
Chord: A Scalable Peer-to-peer Lookup Protocol for Internet Applications Xiaozhou Li COS 461: Computer Networks (precept 04/06/12) Princeton University.
Design of Flash-Based DBMS: An In-Page Logging Approach Sang-Won Lee and Bongki Moon Presented by Chris Homan.
MapReduce and GFS. Introduction r To understand Google’s file system let us look at the sort of processing that needs to be done r We will look at MapReduce.
Log-structured Memory for DRAM-based Storage Stephen Rumble, John Ousterhout Center for Future Architectures Research Storage3.2: Architectures.
Fast Crash Recovery in RAMCloud. Motivation The role of DRAM has been increasing – Facebook used 150TB of DRAM For 200TB of disk storage However, there.
CPSC 404, Laks V.S. Lakshmanan1 External Sorting Chapter 13: Ramakrishnan & Gherke and Chapter 2.3: Garcia-Molina et al.
Amar Phanishayee,LawrenceTan,Vijay Vasudevan
Paper Survey of DHT Distributed Hash Table. Usages Directory service  Very little amount of information, such as URI, metadata, … Storage  Data, such.
DMBS Internals I. What Should a DBMS Do? Store large amounts of data Process queries efficiently Allow multiple users to access the database concurrently.
File Structures. 2 Chapter - Objectives Disk Storage Devices Files of Records Operations on Files Unordered Files Ordered Files Hashed Files Dynamic and.
GFS. Google r Servers are a mix of commodity machines and machines specifically designed for Google m Not necessarily the fastest m Purchases are based.
Improving Disk Throughput in Data-Intensive Servers Enrique V. Carrera and Ricardo Bianchini Department of Computer Science Rutgers University.
Introduction: Memory Management 2 Ideally programmers want memory that is large fast non volatile Memory hierarchy small amount of fast, expensive memory.
Multilevel Caches Microprocessors are getting faster and including a small high speed cache on the same chip.
Dynamo: Amazon’s Highly Available Key-value Store DAAS – Database as a service.
Unit C-Hardware & Software1 GNVQ Foundation Unit C Bits & Bytes.
CS 540 Database Management Systems
Memory The term memory is referred to computer’s main memory, or RAM (Random Access Memory). RAM is the location where data and programs are stored (temporarily),
DMBS Internals I February 24 th, What Should a DBMS Do? Store large amounts of data Process queries efficiently Allow multiple users to access the.
DMBS Internals I. What Should a DBMS Do? Store large amounts of data Process queries efficiently Allow multiple users to access the database concurrently.
Chapter 5 Record Storage and Primary File Organizations
Querying the Internet with PIER CS294-4 Paul Burstein 11/10/2003.
Chapter 11 System Performance Enhancement. Basic Operation of a Computer l Program is loaded into memory l Instruction is fetched from memory l Operands.
The Google File System Sanjay Ghemawat, Howard Gobioff, and Shun-Tak Leung Presenter: Chao-Han Tsai (Some slides adapted from the Google’s series lectures)
Chord: A Scalable Peer-to-Peer Lookup Service for Internet Applications * CS587x Lecture Department of Computer Science Iowa State University *I. Stoica,
Advanced Topics in Concurrency and Reactive Programming: Case Study – Google Cluster Majeed Kassis.
Lecture 16: Data Storage Wednesday, November 6, 2006.
Peer-to-Peer Data Management
CSC 4250 Computer Architectures
Basic Performance Parameters in Computer Architecture:
FAWN: A Fast Array of Wimpy Nodes
Lecture 11: DMBS Internals
Be Fast, Cheap and in Control
EECS 498 Introduction to Distributed Systems Fall 2017
FAWN: A Fast Array of Wimpy Nodes
THE GOOGLE FILE SYSTEM.
Caching 50.5* + Apache Kafka
Presentation transcript:

FAWN: Fast Array of Wimpy Nodes Developed By D. G. Andersen, J. Franklin, M. Kaminsky, A. Phanishayee, L. Tan, V. Vasudevan Presented by Peter O. Oliha Chengyu Zheng UCL Computer Science COMPM038/COMPGZ06

Can we reduce energy use by a factor of ten? Still serve the same workloads. Avoid increasing capital cost. Motivation

Power Consumption and Computing High amount of energy is required for large amounts of data processing “Energy consumption by data centers could nearly double...(by 2011) to more than 100 billion kWh, representing a $7.4 billion annual electricity cost” [EPA Report 2007]

FAWN System FAWN-KV is a key/value store with per-node datastore built on flash storage. Desires to reduce energy consumption Each node: Single core 500MHz AMD processor, 256MB RAM, 4GB CompactFlash device

FAWN -Components FAWN – Flash – FAWN-DS Log structured data store – FAWN-KV Key/value system Put()/get() interface

FAWN Approach: Why use “wimpy” nodes Match CPU-I/O processing times Using wimpy processors reduce I/O-induced idle cycle while maintaining high performance Fast CPU’s consumes more power Spends longer time idle, so less utilization

FAWN Approach: Why use Flash Storage Fast Random Reads – <<1ms upto 175 times faster than random reads on magnetic disks Efficient I/O – Consumes less than 1W even under heavy load Slow Random writes – influences design of the FAWN-DS Suitable for desired workload; random-access, read-intensive.

FAWN-DS: Datastore Functions: Lookup, store,delete, merge, split, compact Designed specifically for flash characteristics – Sequential writes, single-random-access reads

FAWN-DS: Store, Delete Store: -Appends an entry to the log -Updates hash table entry to point to the offset within the data log -Set valid bit to 1 - If the key written already exists, the old value is now orphaned. Delete: -Invalidates hash entry corresponding to the key -Clears the valid bit -Writes “delete entry” at the end of the file -Delete operations are not applied immediately to avoid random writes. -Deletes are carried out on compact operations

FAWN-DS: Maintenance Split, Merge, Compact Split & Merge  Parses the Data log sequentially  Splits single DS into two, one for each key range  Merge writes every log entry from one DS to the other Compact  Cleans up entries to the data store  It Skips  Entries outside data store key range  Orphaned entries  Delete entries corresponding to the above  Writes all other valid entries to the output data store

FAWN-KV: The key-value system Client Front-end  Services client requests through standard put/get interface.  Passes request to the back-end Back-end  Satisfies requests using its FAWN-DS  Replies front-end

FAWN-KV: Consistent hashing Consistent hashing used to organize FAWN-KV virtual ID’s (similar to Chord DHT) Uses 160-bit circular ID space Does not use DHT routing

FAWN-KV: Replication and Consistency Items stored at successor and R-1 virtual ID’s Put()’s are successful when writes are completed on all virtual nodes. Get()s are directly routed to the tail of the chain

FAWN-KV: Joins and Leaves Joins occur in 2 phases – Datastore pre-copy New node gets data from current tail – Chain insertion, log flush Leaves – Replicas must merge key range owned by departed node – Add a new replica to replace departed node: equivalent to a join

Nodes are assumed to be fail-stop Each front-end exchanges heartbeat messages with nodes FAWN-KV: Failure Detection

FAWN: Evaluation Single core 500MHz AMD processors, 256MB RAM, 4GB CompactFlash device Workload targets small objects that are read- intensive( 256 byte and 1KB) 1. Individual Node Performance 2. FAWN-KV 21-Node System

FAWN: Single Node Lookup and Write Speed 80% lookup Speed of raw flash systems Insert rate 23.2MB/s(~24Kentries/s) is 96% write Speed of raw Flash Systems

FAWN: Read-intensive vs. Write-intensive workload

FAWN: Semi-Random Writes

FAWN: System Power Consumption Measurements shown at peak performance

FAWN: Node Joins and Power Measurements shown at max and low loads Joins take longer to complete at max load

FAWN: Splits and Query Latency For purely get() workloads Split increases query latency

FAWN Nodes vs. Conventional Nodes Traditional systems still have sub-optimal efficiency.

TCO: FAWN vs. Traditional Architecture

FAWN: When to use FAWN? FAWN-Based system can provide lower cost per (GB, QueryRate)

Related Work JouleSort: energy efficiency benchmark developed for disk-based low-power CPU. CEMS, AmdahlBlades, Microblades: advocates low-cost, low-power components as building blocks for Datacenter systems IRAM Project: CPU's and memory into a single unit. IRAM-based CPU could use quarter of the power of conventional system for same workload. Dynamo: distributed hashtable structure providing availability to certain workloads

Consider more failure scenarios Management node replication Use in computationally intensive/large dataset workloads Decrease impact of split on query latency Future Work

FAWN: Conclusion Fast and efficient processing of random read- intensive workloads. More work done with less power FAWN-DS balances read/write throughput FAWN-KV balances workload while maintaining replication and consistency Splits and Joins affect latency at high workload Can it be used for computational intensive workloads?