1. Disks and Storage Systems
CSE451
Andrew Whitaker
How is the disk abstraction different than memory?

2. Big Picture
OS provides abstractions to allow physical HW resources to be shared
CPU sharing with threads
Memory sharing with virtual memory
Disk sharing with files

3. Physical disk structure
Disk components
platters
surfaces
tracks
sectors
cylinders
arm
heads
sector
track
surface
cylinder
platter
arm
head

4. Disk Interface (as seen from the OS)
Disk exposes a contiguous array of logical block numbers
Internally, the disk maps these to cylinder, head, sector

5. Disk Size and Cost Cost per gigabyte has plummeted
Now approximately $ .33 per Gigabyte

6. Disk trends Disk capacity, 1975-1989 Disk capacity since 1990
doubled every 3+ years
25% improvement each year
factor of 10 every decade
exponential, but far less rapid than processor performance
Disk capacity since 1990
doubling every 12 months
100% improvement each year
factor of 1000 every decade
10x as fast an improvement as processor performance!

7. Components of Disk Access Latency
Seek time: moving the disk arm to the correct cylinder
Rotation time: waiting for the sector to rotate under head
Depends on rotation rate of disk
Transfer time: transferring data from surface into disk controller, and from there into main memory
Of these, only transfer time is getting better

8. Implications of Disk Performance
Large transfers are faster (per byte) than small transfers
Contiguous transfers are faster than scattered transfers
Contiguous transfer rate is 500X random access rate

9. Disk Partitions
Each physical disk can be split into logical disks, called partitions
Partition 1
Partition 2
Partition 3
Partition table provides sizes, types
Located in the first disk sector
Master Boot Record

10. Disks and File Systems Each partition can contain a file system
Examples: NTFS, ext3, FFS
UNIX mount command shows file system  partition mapping
/dev/sda2 on / type ext3 (rw)
/dev/sda5 on /usr type ext3 (rw,nodev)
/dev/sda6 on /var type ext3 (rw,nosuid,nodev)
/dev/sda7 on /tmp type ext3 (rw,nosuid,nodev)
proc on /proc type proc (rw)

11. Accessing the Disk Disk access is provided by a block device driver
Provides a function to enqueue a request
Privileged User Apps
File Systems
Enqueue
request
Device Driver
Read/write sectors
Disk

12. What Happens on Enqueue Request?
OS maps from <partition, offset> to a logical block number
Make sure the offset is in bounds!
The request is added to a pending request queue
Common optimization: clustering
If the request is adjacent (on disk) to another request, merge them
Request scheduling
OS chooses an order for pending requests (see next slides)
Once the device is free, the device driver dispatches one or more pending requests

13. Disk Scheduling
Problem: Given a set of pending disk requests, choose an ordering
Goals (Quite possibly conflicting)
Maximize throughput
Minimize waiting
Provide fairness
Avoid starvation

14. Policy #1: First-come, First-serve
Disk requests are dispatched in the order they arrive
Performance of FCFS is highly dependent on the workload
Performs well if disk requests arrive in contiguous order
Performs poorly if requests arrivals are scattered across disk

15. Policy #2: Shortest-seek-time-first
Schedule the request that is closest to the current head position
Advantage: much better performance than FCFS
Disadvantage: bias, starvation
Requests in the middle cylinders get better service
Requests in outer cylinders can starve

16. Policy #3: SCAN Scheduling
Head continuously moves back and forth across the disk
Sometimes called “elevator algorithm”
Advantages: starvation-free, reasonable performance, easy to implement (sort)

17. Looking Ahead: File Systems
One proposal: use disk partitions for files
Use one disk partition per file
Is this a good idea? Why or why not?
foo.txt
bar.txt
zag.txt