G lock token appoints replacements for any special

Info iconThis preview shows page 1. Sign up to view the full content.

View Full Document Right Arrow Icon
This is the end of the preview. Sign up to access the rest of the document.

Unformatted text preview: oking node (not Token Manager) sends revoke messages to all nodes that are holding the token –  Revoking node collects replies –  Revoking node forwards these as a single message to the token manager •  Acquiring a token will never require more than 2 messages to the token manager 4 2/4/13 CS480 Principles of Data Management Spring 2013 Optimizing Token Protocol (2/3) •  Token prefetch and token request batching Spring 2013 Optimizing Token Protocol (3/3) •  When a file is deleted on a node –  Allows acquiring mul,ple tokens in a single message from the token manager •  When a file is accessed for the first ,me, mul+ple tokens are acquired with a single token manager request: –  The necessary inode token –  Metanode token –  Byte ­range token to read or write CS480 Principles of Data Management CS480 Principles of Data Management –  The node does not immediately relinquish the tokens –  The next file created by the same node can re ­use the old inode and will not need to acquire new tokens •  A workload where users on different nodes create and delete files under their home directories –  Generates liFle or no token traffic Spring 2013 CS480 Principles of Data Management Spring 2013 Fault Tolerance •  Node Failures –  One or more node(s) fails •  Communica,on Failures –  Bad network adapter or a loose cable –  A failure in the switching fabric Fault Tolerance GPFS CS480 Principles of Data Management Spring 2013 Node Failures (1/3) •  When a node fails, –  Restores metadata updated by the failed node –  Releases resources held by the failed node •  E.g. lock token –  Appoints replacements for any special roles played by the failed node •  Metanode •  Alloca,on manager •  Token manager •  Disk Failures –  Failure at one or more disks CS480 Principles of Data Management Spring 2013 Node Failures (2/3) •  Token manager releases tokens held by failed node •  Distributed locking protocol –  Ensures the metadata it had updated in its cache but had not yet wriFen back to disk at the ,me of the failure? •  Are released only a_er log recovery is complete •  Metadata modified by the failed node will not be accessible to other nodes un,l the state is consistent •  Recovery logs on shared disks –  Run log recovery from the failed node’s log on one of the surviving nodes 5 2/4/13 CS480 Principles of Data...
View Full Document

This note was uploaded on 02/11/2014 for the course CS 480 taught by Professor Staff during the Spring '08 term at Colorado State.

Ask a homework question - tutors are online