The LOCUS Distributed Operating System

Transcription

1 The LOCUS Distributed Operating System Bruce Walker, Gerald Popek, Robert English, Charles Kline and Greg Thiel University of California at Los Angeles 1983 Presented By Quan(Cary) Zhang

2 LOCUS is not just a paper

3 History on Distributed System Roughly speaking, we can divide the history of modern compu9ng into the following eras: 1970s: Timesharing (1 computer with many users) 1980s: Personal compu9ng (1 computer per user) 1990s: Parallel compu9ng (many computers per user) - Andrew S. Tanenbaum(in Amoeba Project)

4 LOCUS Distributed Operating System Network Transparency Looks like one system, though each site runs a copy of kernel Distributed File system with name transparency and replication Remote process creation & Remote IPC Can be across heterogeneous CPU

5 Overview Distributed File System Flexible and automatic replication Nested transaction Remote Processes Recovery from conflicts Dynamic Reconfiguration

6 Distributed File System Naming similar to UNIX: single directory tree Transparent Naming: not mapping name to the location Transparent Replication: Glue filegroups to the directory tree

7 Distributed File System(Cont.)

8 Replication Distributed File System Availability, Yet complicate the update Directory entries stored `nearby the site the child file store Performance High level of the naming tree should be highly replicated, vice versa Essential to the Environment Set up files (Q)

9 Mechanism Support for replication Physical containers store subset of filegroup s files Copies of a file samely resolves to <filegroup number, inode> Version vector for the complicated update

10 File System: Operation All operations keep intact to the unix system call Access Using Site (US) Storage Site (SS) Current Synchronization Site (CSS) <filegroup number, inode> US US

11 File System: Operation(cont.)

12 File System: Operation(cont.) Open/Read SS CSS

13 Read Sharing

14 File System: Operation(cont.) Close If it is the last close, because the US can open the file for several 9mes.

15 Name resolution Search for the pathname iteratively starting from working directory or root Finds a <filegroup, inode> at the end of each search that can match the pathname Possible optimization: No synchronization on CSS Because directory never sees an inconsistent picture ( directory is just a pointer ) Instead of iterating through a tree which is remotely located, try to use migration

16 File System: Operation (Contd ) Modification Modifications are written to new pages/old pages, followed by atomic commit to SS, and close Commit & Abort( use undo log ) One copy of file ( shadow page ) is updated and committed Notify to the CSS( change the version vector ) and to SS Updated file propagation - Pulling by other SS, and also use the commit mechanism Creation Storage locations/copy number for new file determined at create time. Attempts to use same storage sites as parent directory/ local site Remote sites inode allocated by physical container decided by filegroup

17 File System: Operation (Contd ) Machine Dependent File Different Versions of the same file (Process Context based) Remote device and IPC pipe(distributed memory)

18 Remote Processes Supports remote fork and exec (a special module) Copies entire process memory to destination machine: can be slow run system call performs the equivalent to local fork and remote exec Shared data protected using token passing (e.g. file descriptors, handle) Child Parent notified upon failures

19 Recovery Partitioning will occur Strict Synchronization in a partition (independently/transaction) Merging: Directories and mailboxes are easy Issue: File is both deleted and updated Solution: Propagate changes/delete, whichever is later Name conflict: Rename and Automatic - CSS Else, pass to filetype-specific recovery program Else, mail owner by massage

20 Dynamic Reconfiguration Principles for Reconfiguration Delays (internal table reconstruction) should be negligible Same version availability even upon failure - Consistency Clean up when failure detected Close files and sessions & issue error return values Partition: Find fully connected component Synchronize site tables Polls all suspected-good machines then broadcast result Merge: Forms large partitions - Centralized Polls all sites asynchronously, merges partitions if found. Finds new CSS for all filegroups and global tables. Protocol Synchronization Passive site periodically checks active site Passive site can restart protocol

21 Experiment Bla, bla, bla, yet with no essen9al result

22 Conclusion Transparent Distributed System Behavior with high performance is feasible Performance Transparency Except remote fork/exec Not much experience with replicated storage Protocol for membership to a partition with high performance works

23 My perspective on LOCUS Not done: Parallel Compu9ng: because no thread concept currently Process/Thread Migra9on: Load balance Security: SSH Distributed File System VS File System Service in Distributed environment

24 Discussion Why not simply store the Kernel, Shells etc on "local storage space"? I think you are right, not all the resource/file should be distributed into the network of computers e.g. FITCI

25 Can the RPC techniques we discussed earlier be implemented in this framework and help? Probably true: That RPC paper was in 1984, yet this paper was finished in 1983

26 The loca9on transparency required/implemented in this work may incur imbalance of performance cost, is this a problem? Can centralized solu9on help for this problem? I think it is possible to add some policies to a`ain the load balance, e.g., set the physical container for the file group respec9vely to the sites, yet there is real challenge in geang the informa9on about the topology of the network.

27 Compare and contrast "The Mul9kernel" that we discussed on last Systems Research group mee9ng and LOCUS? Could we say LOCUS is a very early stage of a Mul9kernel approach I think the distributed system(amoeba) itself is planning to design the mul9kernel, so, the mul9kernel concept is not new, nonetheless, the mul9kernel we discussed last week can apply to one mechine with different execu9on unit(isa)

28 If the node responsible for a file is very busy and unable to handle network requests, then what happens to my request? I don t think this can happen, because the replica9on never stops.

29 It seems that distributed filesystems don't tend to be very popular in the real world. Instead, networked organiza9ons who need to make files available in mul9ple loca9ons tend to concentrate all storage on one server (or bank of servers), using something like NFS. Why has distributed storage, like in this paper, not become more popular? Because of the scalability issue

30 As number of nodes and par99oning events increase I think that the paper's approach of manual conflict resolu9on won't scale!

31 What do they mean by the 'guess' provided for the incore inode number? Is this sent by the US or the CSS and what happens if the guess is incorrect? And why should guessing be necessary at all - shouldn't the CSS know exactly which logical file it needs to access from the SS

32 Is it possible to store a single file over mul9ple nodes, not in a replicated form, but in a striped form? That would mean that block A of a file is on PC 1, while block B of the file is on PC 2