Slash n: Technical Session 3 - Storage @ Scale: Quest for the mythical silver bullet - Phaneesh...
-
Upload
slashn -
Category
Technology
-
view
449 -
download
0
description
Transcript of Slash n: Technical Session 3 - Storage @ Scale: Quest for the mythical silver bullet - Phaneesh...
17/01/13
Storage @ ScaleQuest for " My#ical Silver Bullet
Wednesday, 30 January 131. Introduce ourselves2. Tell what we do3. Give rundown on flow4. Why storage is a key piece
Stora% &sues & Solutions...
Wednesday, 30 January 13Background: Meta:! - Seeming non-overlap between processing and delivery - Meta-management! ! - Versioning! ! - State/lifecycle! ! - Tagging/Variants Backend:! - Different QoS and SLAs:! ! - Processing needs high writes! ! - Delivery needs high reads! ! - Archival needs cheap! - File-size varianceNeed for separating the meta and storage - our meta >> block level metaSolution: Storage as a framework
Self-healing*
Support of various content types
(Video, music, text etc etc)
Scales to 100s of petabytes
Adaptive read/writescaling*
Supports 1000s of Concurrent clients
Metadata Management & Versioning
QOS & SLA Policy management
One Stora% Solution to rule #em all...
We will not utter any more of Mordor's dark language here....
Ops & Admin Tools
Wednesday, 30 January 13Problem areas
1) Metadata management - Left Side2) Storage - Right side
Clients
Stora% Sy'em in Shir(..
Metadata Server
Storage Service
Client Client Client
NFS MountHTTP Service
HTTP API
ClientClient
Wednesday, 30 January 131) Metadata server
1.1) Versioning 1.2) Tagging
1.3) Lifecycle (Archive, Delete..., )2) Storage Service
2.1) Deployed over ZFS, OpenSolaris, Storage Controller with 16 RAID zones 2.2) Manage files over computable layouts in filesystem
2.3) Duplicate detection & de-duplication3) Clients
3.1) NFS mount 3.2) HTTP client for path transparency
4) Problems 4.1) Capacity scaling
4.2) HA 4.3) Coupling with client systems/applications
Handoff to phaneesh
Stora% Sy'em in Rivendell...
Metadata Servers
HTTP Service
Clients
Client Client Client ClientClient
Storage Infrastructure
HTTP API
Storage Service 1
Storage Service 2
Storage Service 3
Storage Service 3
Optimized for READ
Optimized for WRITE
Optimized for CAPACITY
Optimized for X
Discovery Service
Read Cache Read Cache Read Cache Read Cache Read Cache Read Cache Read Cache
HTTP READ
HTTP WRITE
QOS
Wednesday, 30 January 131) Discovery Service
1.1) Publish capabilities & QOS characteristics 1.2) Maintain availability information
1.3) Leader election (Most optimal for queries returning multiple instances)2) Metadata Server
2.1) Use discovery metadata: QOS3) Storage Infrastructure
3.1) Hybrid 3.2) Optimized for a specific application, usecase, operation
4) Read cache - LRU, Top X files
Storage Infrastructure
StorageStats Query
Palantir Stone sees it all...
Policy & QOS
Metadata Repository
StorageAPI
Discovery Service
Versions
Metadata
Lifecycle
Storage Service 1 Storage Service n
Clients
Client Client Client ClientClient
Metadata
Storage Stats
Actual I/O
Storage System Query
Service Request Service response
Wednesday, 30 January 131) Metadata - Versioned
2) Lifecycle tracking3) Policies -> Referenced/managed with metadata
Discuss call flow
1) Write2) Read
3) MoveQOS -> Write/Move (Lifecycle change)
Read -> Will use QOS applied during Write/Move
Every#ing & commo)ty in Rivendell...
7
X X X•Quad Core•8 GB or 16 GB RAM•2T X 12 SATA
•Quad Core•8 GB or 16 GB RAM•2T X 12 SATA
•Quad Core•8 GB or 16 GB RAM•2T X 12 SATA
Wednesday, 30 January 13
1) Easily replaceable2) Optimized configurations for specific operations, applications3) Economics @ scale
Meanwhile In Dexter's lab...
•5 Node Setup (2 Clients, 3 Storage Nodes)•Desktop grade machines•Connected to a 1G 8-port switch•Various deployment/configurations•Standard read/write load tests using dd, iozone etc..
Subjects•GlusterFS•Ceph
Wednesday, 30 January 131) GlusterFS
1.1) Distributed mode characteristics 1.2) Distributed-replicated
1.3) Striping - Read charecteristics, Write charecteristics
2) Ceph??
Self-healing*
Support of various content types
(Video, music, text etc etc)
Scales to 100s of petabytes
Adaptive read/writescaling*
Supports 1000s of Concurrent clients
Metadata Management & Versioning
QOS & SLA Policy management
One Stora% Solution to rule #em all...
We will not utter any more of Mordor's dark language here....
Ops & Admin Tools
Wednesday, 30 January 131) Problems that we need not solve
2) Solutions that are generic to LOB3) Specific to digital content
Questions?
Wednesday, 30 January 13
Wednesday, 30 January 13