Enabling System Transactions via Lightweight Kernel Extensions R.P. - PowerPoint PPT Presentation

Enabling System Transactions via Lightweight Kernel Extensions R.P. Spillane, S. Gaikwad. M. Chinni, C.P. Wright, E. Zadok Stony Brook University http://www.fsl.cs.sunysb.edu/

Summary  What is the design complexity of system transactions implemented in the VFS?  Low  100 lines of code added to page writeback  4000 lines of module code (log implementation)  What is the performance?  Valor: 35% overhead on top of theoretical best, compared to…  104% overhead for an efficient user-level alternative 2/28/2009 FAST 2009 - Enabling System Transactions 2

System Transaction Process 1 FS State: foo ’ FS State: foo System Calls TID←sys_tbegin (...) / / write(TID,...) unlink(TID,...) sys_tabort(TID) f1 f2 f1 f2 2/28/2009 FAST 2009 - Enabling System Transactions 3

The Design Spectrum  Valor side-steps the traditional trade-off by working with the Kernel’s page cache in a general way. Quicksilver, TxF Valor Transparency & Amino Performance Berkeley DB, KBDB Stasis Design Feasibility 2/28/2009 FAST 2009 - Enabling System Transactions 4

Valor’s Process Txn Model  Transactional Model  Supported Operations:  dirtying a page  appending to a file, modifying an inode  modifying a directory  Locking:  directory locks, inode locks  page range locks for overwrites  intent locks for directory renames 2/28/2009 FAST 2009 - Enabling System Transactions 5

Asynchronous By Default  ACI (no D w/o tsync)  Similar to asynchronous write(2) with fsync(2)  Same purpose (performance increase)  Requires page cache for files updated transactionally 2/28/2009 FAST 2009 - Enabling System Transactions 6

Valor Design  Modify page writeback to support simple write ordering  Implement an ARIES style undo/redo log module for FS-operations 2/28/2009 FAST 2009 - Enabling System Transactions 7

Page Dirtying: No Txns LEGEND: OK bad Old Page Process 1 New Page Uh- oh… write(TID,… ) write(TID,… ) 2/28/2009 FAST 2009 - Enabling System Transactions 8

Page Dirtying: With Txns LEGEND: Old Page Process 1 U/R Page New Page log_append (TID,… ) log_append (TID,… ) write(TID,… ) write(TID,… ) 2/28/2009 FAST 2009 - Enabling System Transactions 9

Current Kernel Design LEGEND: Page Cache Old Page U/R Page Ext3 New Page Ext2 Uh- oh… XFS ZFS Page Writeback … Process 2 log_append (TID,…) write(TID,…) 2/28/2009 FAST 2009 - Enabling System Transactions 10

What DBs Do Page Cache II: The Wrath of Khan Disk Cache Flush (fsync) Page Cache Ext2 XFS ZFS 2/28/2009 FAST 2009 - Enabling System Transactions 11

Simple Write Ordering LEGEND: Page Cache Old Page FS1 U/R Page FS2 New Page FS3 FS4 Valor 2/28/2009 FAST 2009 - Enabling System Transactions 12

Log Module State File Log File Process 2 U/R Page 1 Disk tbegin (TID,…) 1 1 U/R,1 U/R,1 U/R,1 tlog (TID,…) 2 1 3 2 U/R Page 2 Valor Module write(TID,…) 3 2 U/R,1 U/R,1 C,1 page writeback 4 U/R Page 3 4 5 6 tlog (TID,…) 5 3 Record Maps write(TID,…) 6 U/R Page 4 U/R,1 U/R,1 U/R,1 tresolve (TID,…) 7 4 1 3 2 8 page writeback U/R Page 5 page writeback 9 5 U/R,1 U/R,1 C,1 4 5 6 6 2/28/2009 FAST 2009 - Enabling System Transactions 13

Atomicity Argument  Transition from pre-writeback to post- writeback disk state atomically iff  All writes preceded by sys_log_append  Simple write ordering is implemented  writes to a single sector are atomic  Valor satisfies the top 2 constraints  A supported hard disk satisfies the third 2/28/2009 FAST 2009 - Enabling System Transactions 14

Performing Recovery  Two kinds of recovery are supported:  System Recovery  Application Recovery (per-process abort)  Standard recovery process:  Reconstruct RAM state from log  In reverse LSN order commit/abort landed transactions  Perform a page writeback 2/28/2009 FAST 2009 - Enabling System Transactions 15

Evaluation  We must compare against traditional asynchronous FSes  benchmark against asynchronous ext3  do serial transfer benchmarks for large files  We turn off synchronous transactions for two other controls (for fairness)  FS built on top of Stasis  FS built on top of Berkeley DB 2/28/2009 FAST 2009 - Enabling System Transactions 16

Mock ARIES Benchmark  Important lower bound (not tight) MT-ow-noread MT-ow MT-ow-finite Disk Disk Disk Log Log Log 2/28/2009 FAST 2009 - Enabling System Transactions 17

Mock ARIES Benchmark 104% 90 66% 80 Elapsed Time (sec) 70 35% 60 16% 50 2% 2x 40 Wait 30 User 20 System 10 0 2/28/2009 FAST 2009 - Enabling System Transactions 18

Serial Overwrite Transaction size: 16 pages 1000 900 22.75 x Ext3 Elapsed Time (sec) 800 700 600 BDB 500 Stasis 400 Valor 300 200 Ext3 100 5.0 x Ext3 0 256 512 1024 2048 2.75 x Ext3 Size of Serial Transfer (MiB) 2/28/2009 FAST 2009 - Enabling System Transactions 19

Transaction Throughput 1200 Valor Heel 1000 Elapsed Time (sec) 23.0 x Ext3 800 BDB 600 BDB Heel Stasis Valor 400 Stasis Heel Ext3 200 4.2 x Ext3 0 2.9 x Ext3 1 4 16 64 256 Size of Transaction (pages) 2/28/2009 FAST 2009 - Enabling System Transactions 20

Conclusions  System transactions are feasible  Valor achieves good overhead  Minimal changes to existing kernels 2/28/2009 FAST 2009 - Enabling System Transactions 21

Limitations/Future Work  Limitations  Locking slows interleaved writes to the same page  Some FSes/Disks do not fsync() when asked to  Future Work  Explore use of logging device as a coordinator in a transactional disk array 2/28/2009 FAST 2009 - Enabling System Transactions 22

Q&A Enabling System Transactions via Lightweight Kernel Extensions R.P. Spillane, S. Gaikwad. M. Chinni, C.P. Wright, E. Zadok Stony Brook University http://www.fsl.cs.sunysb.edu/

TxF  TxF is Microsoft’s transactional file system  Motivation: program installation, system updates, website updates  Pros  Backed by Microsoft  Cons  Specific to NTFS 2/28/2009 FAST 2009 - Enabling System Transactions 24

Isolation  Extended mandatory locking  Allows locking of directories  Do not have to set group exec/setgid bits  Locking permissions  Let users decide if a file can be locked  All processes acquire locks  Regular processes hold only for the syscall  Lock inheritance  Allow multi-process transactions 2/28/2009 FAST 2009 - Enabling System Transactions 25

Valor != Journaling  Journaling FSes good at fast recovery  …but are too special-purpose:  No-Steal Caching  all state modified by a txn. must remain in memory until commit/abort  Non-Modular Design  does not handle rollback of VFS and page caches, just disk-state on boot 2/28/2009 FAST 2009 - Enabling System Transactions 26