DualFS: a New Journaling Journaling File System without File System - PowerPoint PPT Presentation

International Conference on Supercomputing International Conference on Supercomputing (ICS 2002) (ICS 2002) New York, New York, June 2002 New York, New York, June 2002 DualFS: a New Journaling Journaling File System without File System without DualFS: a New DualFS: a New Journaling File System without Meta- -Data Duplication Data Duplication Meta Meta-Data Duplication Juan Piernas 1 , Toni Cortes 2 , and José M. García 1 1 Dpt. Ingeniería y Tecnología de Computadores Universidad de Murcia 2 Dpt. d’Arquitectura de Computadors U P Universitat Politècnica de Catalunya C D A C e-mail: piernas@ditec.um.es GACOP

Introduction Introduction Introduction � Meta-data management is a key design issue � Especially important for recovery after a system crash � Traditional file systems: � Write meta-data in a synchronous way � Use fsck-like tools � New approaches: � Log of last meta-data updates (e.g. XFS, JFS) � Asynchronous meta-data writes (e.g. Soft Updates) � New approaches treat data and meta-data U somewhat differently P C � But they are completely different. D A C GACOP

Objective Objective Objective � Introduce a new high performance journaling file system, DualFS � Present a working prototype � Evaluate the prototype and ideas U P C D A C GACOP

Outline Outline Outline � Introduction � DualFS � Experimental Methodology and Results � Conclusions U P C D A C GACOP

Key Issues Key Issues Key Issues � Main idea: � Manage data and meta-data in completely different ways. � Each type of blocks will be located on different devices: – meta-data blocks in the meta-data device – data blocks in the data device – can share the same disk � Previous work � Muller and Pasquale (SOSP’91) � Anderson, Chase, and Vahdat. (OSDI’00) U P C D A C GACOP

Data Device Data Device Data Device � Like Ext2 without meta-data blocks � Standard linux file system � Structure based on the concept of group of data blocks � Related blocks are kept together. � Grouping is performed in a per directory basis. � Data blocks are not written synchronously U P C D A C GACOP

Meta- -Data Data Device Device Meta Meta-Data Device � We understand meta-data as all these items: � i-nodes � indirect blocks � directory “data” blocks � symbolic links � bitmaps � superblock copies � Organized as a log-structured file system � Similar to BSD-LFS. U � Similar cleaner P C D A C GACOP

Meta- -Data Data Device Structure Device Structure Meta Meta-Data Device Structure U P C D A C GACOP

IFile IFile IFile U P C D A C GACOP

Recovery Recovery Recovery � DualFS is considered consistent when information about meta-data is correct. � We can recover the file system consistency very quickly from the last checkpoint. � The length of time for recovery is proportional to the inter-checkpoint interval. U P C D A C GACOP

Potential and limitations Potential and limitations Potential and limitations � Advantages of DualFS � One-copy meta-data elements – Traditional journaling file systems write meta-data twice � In the log � In the regular file system structure � Different structures for the meta-data and the data devices � Potential Problem: � Read-only workloads U – Possible solution: to put meta-data blocks on another disk P C D A C GACOP

Outline Outline Outline � Introduction � DualFS � Experimental Methodology and Results � Conclusions U P C D A C GACOP

Experimental Methodology Experimental Methodology Experimental Methodology � DualFS has been compared with: � Ext2, the default file system in Linux, and � Ext3, a journaling file system derived from Ext2. � Tested applications � Microbenchmarks (in the paper) � Macrobenchmarks U P C D A C GACOP

System Under Test Test System Under System Under Test Linux Platform Processor Two 450 Mhz Pentium III Memory 256 MB, PC100 SDRAM Disk Two 4 GB IDE 5,400 RPM Seagate ST- 34310A. Two 4GB SCSI 10,000 RPM FUJITSU MAC3045SC. SCSI disk 1: Operating system, swap and trace log. SCSI disk 2: trace log. IDE disks: test disks U P OS Linux 2.2.19 C D A C GACOP

Macrobenchmarks Macrobenchmarks Macrobenchmarks � Kernel Compilation for 1 Process (KC-1P) � make dep, make bzImage, make modules � Kernel Compilation for 8 Processes (KC-8P) � make dep, make -j8 bzImage, make -j8 modules � Video Compression (VC) � Specweb99 (SW99) � PostMark (PM) U P C D A C GACOP

Macrobenchmarks Results Macrobenchmarks Results Macrobenchmarks Results Macrobenchmarks (1 disk) 3 2.5 Normalized Disk I/O Time 2 Ext2 1.5 Ext3 2.42 DualFS 1 1.78 309.06 sec 982.19 sec 51.65 sec 84.87 sec 62.36 sec 1.26 1.14 1.11 1.01 0.5 0.24 0.63 0.57 0.56 U 0 P C D A C KC-1P KC-8P VC SW99 PM Benchmark GACOP

Cleaner Evaluation Cleaner Evaluation Cleaner Evaluation � Experiment for evaluating the cleaner : � copy a directory tree � delete 87.5% (7/8) of its regular files � Repeat the two previous steps 20 times � Tested DualFS configurations: � DualFS without cleaner � DualFS with cleaner – cleaning a segment every five seconds. U P C D A C GACOP

Cleaner Results Cleaner Results Cleaner Results 70 60 50 I/O Time 40 64.04 30 44.41 44.9 42.16 20 10 0 U Ext2 Ext3 DualFS-cleaner DualFS+cleaner P C File System D A C GACOP

Outline Outline Outline � Introduction � DualFS � Experimental Methodology and Results � Conclusions U P C D A C GACOP

Conclusions Conclusions Conclusions � DualFS is a new journaling file system � data and meta-data managed in very different ways � one-copy meta-data blocks � We have compared DualFS against Ext2 and Ext3: � DualFS reduces total I/O time in most cases (up to 97%) � DualFS increases total I/O time (up to 17%) – Only in a few and limited cases � DualFS has a very good performance in – Internet U – Development environments P C D A C GACOP

International Conference on Supercomputing International Conference on Supercomputing (ICS 2002) (ICS 2002) New York, New York, June 2002 New York, New York, June 2002 DualFS: a New Journaling Journaling File System without File System without DualFS: a New DualFS: a New Journaling File System without Meta- -Data Duplication Data Duplication Meta Meta-Data Duplication Juan Piernas 1 , Toni Cortes 2 , and José M. García 1 1 Dpt. Ingeniería y Tecnología de Computadores Universidad de Murcia 2 Dpt. d’Arquitectura de Computadors U P Universitat Politècnica de Catalunya C D A C e-mail: piernas@ditec.um.es GACOP