Eos at 6,500 kilometres wide An Australian Experience

Slides:



Advertisements
Similar presentations
CERN IT Department CH-1211 Geneva 23 Switzerland t Data & Storage Services Technical student CERN IT-DSS-FDO University of Vigo WCSFSS 2014.
Advertisements

Toolbox Mirror -Overview Effective Distributed Learning.
The Google File System. Why? Google has lots of data –Cannot fit in traditional file system –Spans hundreds (thousands) of servers connected to (tens.
Lesson 20 – OTHER WINDOWS 2000 SERVER SERVICES. DHCP server DNS RAS and RRAS Internet Information Server Cluster services Windows terminal services OVERVIEW.
ownCloud Primary Storage
Seafile - Scalable Cloud Storage System
TYPES OF NETWORKS NETWORK CONFIGURATIONS /TOPOLOGIES TRANSMISSION MEDIA By B. Vialva.
Virtual Network Servers. What is a Server? 1. A software application that provides a specific one or more services to other computers  Example: Apache.
Google Distributed System and Hadoop Lakshmi Thyagarajan.
Take An Internal Look at Hadoop Hairong Kuang Grid Team, Yahoo! Inc
Storage Area Networks The Basics. Storage Area Networks SANS are designed to give you: More disk space Multiple server access to a single disk pool Better.
VAP What is a Virtual Application ? A virtual application is an application that has been optimized to run on virtual infrastructure. The application software.
DNN Performance & Scalability Planning, Evaluating & Improving : Part 2.
Chapter 7: Using Windows Servers to Share Information.
CSC 456 Operating Systems Seminar Presentation (11/13/2012) Leon Weingard, Liang Xin The Google File System.
Module 12: Designing High Availability in Windows Server ® 2008.
IRODS performance test and SRB system at KEK Yoshimi KEK Building data grids with iRODS 27 May 2008.
Chapter 8 Implementing Disaster Recovery and High Availability Hands-On Virtual Computing.
Computing on the Cloud Jason Detchevery March 4 th 2009.
G053 - Lecture 08 Hosting Websites Mr C Johnston ICT Teacher
Why GridFTP? l Performance u Parallel TCP streams, optimal TCP buffer u Non TCP protocol such as UDT u Order of magnitude greater l Cluster-to-cluster.
Personal Computer - Stand- Alone Database  Database (or files) reside on a PC - on the hard disk.  Applications run on the same PC and directly access.
LCG Phase 2 Planning Meeting - Friday July 30th, 2004 Jean-Yves Nief CC-IN2P3, Lyon An example of a data access model in a Tier 1.
Basic Computer Knowledge. Outline Notes 1 Notes 2 Assessment.
CENTER FOR HIGH PERFORMANCE COMPUTING Introduction to I/O in the HPC Environment Brian Haymore, Sam Liston,
Future home directories at CERN
SLACFederated Storage Workshop Summary For pre-GDB (Data Access) Meeting 5/13/14 Andrew Hanushevsky SLAC National Accelerator Laboratory.
Grid Technology CERN IT Department CH-1211 Geneva 23 Switzerland t DBCF GT Upcoming Features and Roadmap Ricardo Rocha ( on behalf of the.
ALICE DATA ACCESS MODEL Outline 05/13/2014 ALICE Data Access Model 2  ALICE data access model  Infrastructure and SE monitoring.
Andrea Manzi CERN On behalf of the DPM team HEPiX Fall 2014 Workshop DPM performance tuning hints for HTTP/WebDAV and Xrootd 1 16/10/2014.
Windows Azure poDRw_Xi3Aw.
Tackling I/O Issues 1 David Race 16 March 2010.
Computing Strategies. A computing strategy should identify – the hardware, – the software, – Internet services, and – the network connectivity needed.
Website Deployment Week 12. Software Engineering Practices Consider the generic process framework – Communication – Planning – Modeling – Construction.
Introduction to Computers - Hardware
Federating Data in the ALICE Experiment
Parallel Virtual File System (PVFS) a.k.a. OrangeFS
Chapter 7: Using Windows Servers
Getting the Most out of Scientific Computing Resources
Dynamic Extension of the INFN Tier-1 on external resources
Storage Area Networks The Basics.
File Syncing Technology Advancement in Seafile -- Drive Client and Real-time Backup Server Johnathan Xu CTO, Seafile Ltd.
Integrating Disk into Backup for Faster Restores
DEPARTMENT OF COMPUTER SCIENCE AND ENGINEERING CLOUD COMPUTING
Getting the Most out of Scientific Computing Resources
Netscape Application Server
Cloud Computing.
Large-scale file systems and Map-Reduce
SCHRödinger’s cloud storage
Introduction to Data Management in EGI
Federated Data Storage System Prototype for LHC experiments
Study course: “Computing clusters, grids and clouds” Andrey Y. Shevel
Windows Azure Migrating SQL Server Workloads
Product Datasheet AppSense DataNow 4.1
Introduction to HDFS: Hadoop Distributed File System
Introduction to Networks
Introduction to Computers
Gregory Kesden, CSE-291 (Storage Systems) Fall 2017
CTA: CERN Tape Archive Overview and architecture
Gregory Kesden, CSE-291 (Cloud Computing) Fall 2016
Migration Strategies – Business Desktop Deployment (BDD) Overview
Computer networking In this section of notes you will learn the rudiments of networking, the components of a network and how to secure a network.
Big-Data around the world
Clouds & Containers: Case Studies for Big Data
Cloud computing mechanisms
Outline Announcements Lab2 Distributed File Systems 1/17/2019 COP5611.
CS 345A Data Mining MapReduce This presentation has been altered.
Outline Review of Quiz #1 Distributed File Systems 4/20/2019 COP5611.
THE GOOGLE FILE SYSTEM.
Presentation transcript:

Eos at 6,500 kilometres wide An Australian Experience David Jericho – Solutions Architect, AARNet

We’re geographically large… Australia is a huge continent 47ms from Sydney to Perth 27ms from Brisbane to Melbourne 90ms from Darwin to Perth Bandwidth is not the problem RTT * TCP Window = Goodput Campus networks User equipment choices © AARNet Pty Ltd |

…Because we’re spread thin The dark blue areas show 50% of our 24 million people 98% of the population live within 150 kilometres of the coast line Collaborations occur with groups thousands of kilometres apart © AARNet Pty Ltd |

Cloudstor sites 3 major sites Brisbane -> Melbourne is 22ms Melbourne -> Perth is 43ms At least two geographically disperse replicas Network splits do occur despite redundancy At least 24 servers providing metadata, application, storage and additional compute in each site This is excessive for our current needs It is a large number of servers for a small team © AARNet Pty Ltd |

Solutions we have tried Hadoop MapR, Hortonworks, Apache official XtreemFS Ceph GlusterFS pNFS OrangeFS … and others © AARNet Pty Ltd |

Our primary EOS cluster © AARNet Pty Ltd |

EOS clusters in AARNEt CloudStor CloudStor support infrastructure 2.5PB presently 12 machines Not homogenous 40Gbps connected each 44 or 36 disks per machine Multiple MGMs Encrypted at rest Only 4% of files are larger than 10 megabytes CloudStor support infrastructure Uses Seagate Kinetic FSTs Connects back to developer’s workstations Content Delivery Network 6 FSTs 48 TB, 12 disks each Acts as a canary for CloudStor Extremely read heavy with only one write client Test CloudStor Only 30TB 3 machines Treated as entirely disposable Interest in using EOS elsewhere too © AARNet Pty Ltd |

CloudStor uses EOSD Fastest way to production EOS usage Everything understands files eosd has had many improvements since we’ve been using it > 0.3.222 has been almost entirely trouble free 48 web servers eosd runs in foreground within a container Volume mapped /eos to other containers © AARNet Pty Ltd |

Completely containerised Via Docker Engine All containers bind to the host’s network stack Orchestrated by Rancher Each FST geotag is its own pod/stack MGM Master is own pod/stack Slaves all reside in the same pod/stack Each component containerised MGM, MQ, Sync, eossync on same host FSTs also run on MGM hosts as a container FSTs mount encrypted storage inside the container eosd running foreground in container Extremely quick to deploy and very consistent © AARNet Pty Ltd |

Successes we’ve had IT WORKS! Obvious write latency penalty Stable, server issues have been almost exclusively container related Fast Obvious write latency penalty Users don’t notice Hello all, I know it’s Monday… CERN have been very responsive, THANKYOU! © AARNet Pty Ltd |

observations IO performs at disk speed IO even at 65ms Does require suitable tuning and stable network Latency can be obvious with many small files “Square” groups can mask this through average performance Xroot latency is insignificant with respect to network latency Moving towards direct xroot manipulation for IO FileSender tool streams concatenated chunks Mass ingest tool for ownCloud environment GeoTag functionality works well Will eventually result in imbalanced FST counts in some sites This isn’t a problem © AARNet Pty Ltd |

Problems along the way eosd has matured a lot FST start up can be slow Many user space tools do silly things (rsync) FST start up can be slow Mostly resolved now Better group/fs layouts has helped Documentation is rare Word of mouth training Internal documentation © AARNet Pty Ltd |

Possible improvements Bundling of balancing transfers Many small files limits speed of balancing UDT or similar protocol for FST communications Or at least a software tunable window size Latency sensitive geotag reading AARNet contributed test infrastructure for Citrine Help with maturity of software in non-CERN use cases © AARNet Pty Ltd |

All images CC0 unless otherwise noted