File System Performance
CSE451Andrew Whitaker
Ways to Improve Performance
Access the disk less Caching!
Be smarter about accessing the disk Turn small operations into large operations Turn scattered operations into sequential operations
Technique #1: Caching
Memory is MUCH faster than diskSo, cache whatever we can in memory
File buffers i-nodes Directory entries (name => i-node)
Caching reads is a no-brainerCaching writes is more interesting…
Caching Writes
Two options Synchronous: data is immediately written out to disk
AKA: write-through Asynchronous: disk writes are delayed
AKA: write-back
Programmer’s perspective: what does it mean when the “write” system call returns? With asynchronous writes, the data has not necessarily hit
the disk
Why Use Asynchronous Writes?
Allows us to batch-up multiple writes to the same block
Allows for better overlap of CPU and I/O CPU does not stall waiting for the disk
Allows the disk scheduler to make better decisions Application: write(a); write (b); write(c); Disk: write(b); write(a); write(c);
Most data updates in UNIX systems use asynchronous writes by default Programmer can override: fsync(fd);
Problems with Asynchronous Writes
File system state can be lost during a crash Missing blocks, missing files,
missing directories, storage leaks, etc.
For this reason, meta-data updates tend to be done synchronously File/directory creation or
deletion
Consistency Problems
Problems still arise, even with synchronous meta-data updates For example, file creation must modify an i-node and a directory entry
Initialize the i-node Record the <fileName, i-node> mapping in the directory
Disks do not support atomic operations
Dealing with Consistency Problems
Always keep the disk in a “safe” stateRun a recovery program (like fsck) on
startup
i-check: File Consistency
Is each block on exactly one list? Create a bit vector with as many entries as there are
blocks Follow the free list and each i-node block list When a block is encountered, examine its bit
If the bit was 0, set it to 1 If the bit was already 1
• if the block is both in a file and on the free list, remove it from the free list and cross your fingers
• if the block is in two files, call support! If there are any 0’s left at the end, put those blocks on
the free list
d-check: Directory Consistency
Do the directories form a tree? Cycles are bad!
Does the link count of each file (i-node) equal the number of directory links to it?
Technique #2: Better Data Layout
Recall basic file system structure: Meta-data: i-nodes, free block list Data: file data, directory data
Metadata Data
Note: i-nodes are far from the data blocks they describe
Cylinder groups Basic idea: group commonly accessed data and
meta-data together This reduces seeks
Details: Disk is partitioned into groups of cylinders Data blocks from a file are all placed in the same
cylinder group Files in same directory are placed in the same cylinder
group i-node for file placed in same cylinder group as file’s
data
Cylinder Group Analysis
+ Reduces or eliminates seeks for some common access patterns
- Does not address rotational delay- Performance is workload dependent- Performance degrades if cylinders become full
- Partial solution: pro-actively reserve space
Log Structured File System
Let’s assume all reads are cached An iffy assumption, but let’s suspend disbelief
Q: How can we turn all writes into large, sequential writes?
Insight: this is possible if the location of data on disk can change
A Convention File System
Files live at fixed locationSo, file system writes
must use seeksFor example:
Write to Christine.txt Write to Andrew.txt Write to Colin.txt
Veneta.txtJoel.txt
Colin.txt
Matt.txt
Andrew.txtNolan.txt
Bishop.txtChristine.txt
Log-structured File System
Use the disk as an append-only log All writes go at the end of the
logThe location of a file
changes over timeOld data is not over-written
Until the file system becomes full
Christine.txtAndrew.txtColin.txt
Loggrowth
Christine.txt
LFS Details
Everything gets written to the log File data, i-nodes, directories
LFS tries to buffer many small writes into large segments Typically 512k, 1MB
How Can This Possibly Work?
Q: If nothing lives at a fixed location, how do we find “the data”?
A: Add a layer of indirection: An i-node map Maps from i-node number to current location The map resides at a fixed location on disk
NOT in the log! The map is cached in memory for performance
What Happens When the Disk Gets Full?
Partial solution: disk is managed in segments, which are threaded on disk Basically, a linked-list
But, this re-introduces seeks!
Segment Cleaner
Goal: make scattered segments contiguous again
Approach: Read a segment Write live data to the end of the log Presto: The segment is now clean
This is very expensive Each live byte is read and written
LFS Analysis
For reads, LFS and a traditional FS are largely equivalent
LFS has better performance for small writes and meta-data operations
The LFS cleaner has a large impact on performance How important is this?
LFS in Practice
LFS is implemented, but not widely usedReasons?
Assumptions about read behavior were not valid Reads have not gone away
Performance improvements were not sufficient to offset increase complexity, higher variability
LFS comeback? See Jim Gray’s article