Fast File System

2/17/2006

Introduction

• Paper talked about changes to old BSD 4.2 File System (FS)

• Motivation - Applications require greater throughput– Large amounts of paging– Want to retain the good of the old FS

• Abstraction• Backward compatibility with existing software

Old File System

• Block size change from 512 KB to 1024 performance > x2 – WHY?– Disk access had 2x more data– Direct blocks had twice the data so indirect blocks

weren’t needed as much• Performance degradation over time

– 175kb/s 30kb/s due to randomization of block placement on disk

• Fundamental limits:– Small block size– read-ahead in the system– Large seek numbers limits file system throughput.

New File System (FS)

• Drives mapped into partitions

• Each partition has a FS described by redundant superblock

• File system blocks = 4096 bytes

• Cyllinder groups – what’s that?

Cyllinder Groups

• a collection of cylinder groups; Each cylinder group has the following components:

• a backup copy of the superblock • a cylinder group header, with statistics, free lists,

etc, about this cylinder group, similar to those in the superblock

• a number of inodes, each containing file attributes

• a number of data blocks

Storage Utilization

• 4x bigger blocks sizes 4096 bytes throughput

• Problem: unix FS commonly composed of small files wasted space

Utilization Cot’d

• Small files stored in more efficient way– Blocks broken into fragments of 512 bytes– Block map associated with each cylinder

group records the space available in a cylinder group at the fragment level

– Is a block is available? look at aligned fragments

Utilization Cot’d

• Fragments of adjoining blocks cannot be used as a full block, even if they are large enough.

• If no block with enough aligned fragments is available at file creation, a full size block is split yielding the necessary fragments and a single unused fragment.

Utilization

• One of three conditions for file growth allocation– Enough space in allocated block data written to

space– Files contain no fragments– Files contain fragments

• Problem: file growth one fragment at a time many data copies

• Soln: User programs write one block at a time

Utilization

• Capacity– Problem: as unallocated space 0, the

throughput falls to 50%– System should keep ~10% unallocated space– Soln: at a threshold, only administrator can

write new blocks

Parameterization of HW

• Why?– Old FS had no information about physical

characteristics of storage device– Blocks allocated optimally

• Processor speed• HW support for mass transfers• Characteristics of mass storage devices (#

platters, physical data layouts, etc)

Parameterization of HW

• Physical characteristics of each disk:– number of blocks per track– rate of disk spin

• Cylinder group summary information: – Cost of rotationally optimal blocks is not free– Soln: count of the available blocks in a

cylinder group at different rotational positions.

• FS can be parameterized to support min. processor disk operation schedule

Layout Policies

• Two parts to data layout policies:– top level -- global policies use FS-wide

summary information to make decisions regarding the placement of inodes and blocks

– Lower level -- local allocation routines use a locally optimal scheme to lay out data blocks.

• Global policies try to balance conflict:– localizing data that is concurrently accessed– spreading out unrelated files

Layout Policies Cot’d

– Two allocatable resources• Inodes• Blocks

– Layout policy tries to place all the inodes of files in a directory in the same cylinder group.

– Data blocks usually accessed together layout policy tries to place all data blocks for a file in the same cylinder group, preferably at rotationally optimal positions in the same cylinder.

Performance

• % of bandwidth in Table 2 measures: effectiveness of utilization of the disk by the file system.

• upper bound on the transfer rate from the disk:– number of bytes on a track x number of revolutions of

the disk per second. • Bandwidth is calculated by comparing the data

rates the file system is able to achieve as a percentage of the bound.

• Results: – the old FS uses 3−5% of the disk bandwidth– new FS uses up to 47% of the bandwidth.

Performance• Some stats:

Performance Continued

• Limits– Processors limit throughput– Memory to memory copying – 40% of I/O time– Block chaining would require driver rewrites– One block allocated at a time – 10% of

system writes

Enhancements

• File system changes and required downtime allowed for some new ideas– Long File Names– File Locking– Symbolic Links– Rename– Quotas

Summary

• File System Changes from Old System– Block size increased– Layout more efficient– Fragments used to reduce space waste

• Performance increased

• New items implemented into FS that had been requested by users

Summary Cot’d

• New FS organization– Utilization– Parameterization– New Layout Policies

• Performance Improvements

Key Ideas

• Old FS used fraction of the available data throughput

• New FS:– same data structures– Same FS semantics

• New FS has new functionality