Presentation is loading. Please wait.

Presentation is loading. Please wait.

Parallel File System. Outline Working Progress Distributed Metadata Cluster  Subtree Partitioning  Pure Hash.

Similar presentations


Presentation on theme: "Parallel File System. Outline Working Progress Distributed Metadata Cluster  Subtree Partitioning  Pure Hash."— Presentation transcript:

1 Parallel File System

2 Outline Working Progress Distributed Metadata Cluster  Subtree Partitioning  Pure Hash

3 Parallel File System For Windows Porting our MRPVFS to Windows platform  Still client/server model Based on TCP/IP  A centralized metadata server  Separates data/metadata operations Clients directly get files from I/O nodes after getting the metadata of the files  A media player agent at client side Do not need a centralized VOD server to gather striped files

4 Parallel I/O Accesses Use our libwpvfs  Recompiled needed POSIX-compliant interface is under construction  Through Redirect I/O  Existing applications can benefit from our WPVFS

5 Playing Striped Multimedia Files Playing streaming instead of a complete file  Streaming HTTPS A thin web server  Gathers striped files from I/O nodes  Feeds the streaming to client

6 Read Performance

7 Write Performance

8 A Centralized Metadata Server ? A Single Point of Failure..

9 Also a Performance Bottleneck ? One client, one MDS, one I/O node Postmark (1000 files, 10 directories, random access)

10 The Objectives of Metadata Server Cluster POSIX-compliant APIs  Standard UNIX-style file and directory semantics High Performance  Efficient metadata access  Efficient directory operations  Efficient access control  High degree of parallelism Scalability  # of metadata servers 、 namespace 、 load balancing  Addition and removal of metadata servers

11 Directory Subtree Partitioning Hierarchical namespace partitioned by directory subtrees (e.g. NFS) Pros  Supports standard directory semantics  Efficient access to multiple files in same directory Cons  Bottlenecks with high concurrent accesses  Coarse granularity of load balancing  Adding or removing metadata servers is costly Difficulty to manage May have to move a significant amount of metadata

12 Pure Hashing Namespace widely distributed among the metadata servers based on hash of file or pathname  Full name  Vesta, File name  Lustre Pros  One-request metadata lookup  Bottleneck avoidance Cons  Hard to support standard directory semantics Permission, list files in a directory,…  Adding or removing metadata servers is costly

13 Mirrored Distributed Metadata Cluster 0 – 3FFENode 0, Node 4 3FFF - 7FFDNode 1, Node 2 7FFE - BFFCNode 2, Node 3 BFFD - FFFBNode 3, Node 4 Full filename 16-bit hash IO Node IOD Real File IO Node IOD Real File IO Node IOD Real File IO Node IOD Real File IO Node IOD Real File

14 Issues Directory Renamed or Moved Node Add or Leave


Download ppt "Parallel File System. Outline Working Progress Distributed Metadata Cluster  Subtree Partitioning  Pure Hash."

Similar presentations


Ads by Google