1. Research on Disks and Disk Scheduling
Brian Railing
Monday, November 3rd 2003
Fall 2003
Original lecture given by Steve Muckle on Monday, March 31st 2003
Additional Slides Taken from Eno Thereska’s July Systems Talk
Also from Terrence Wong’s Midsemester Thesis Presentation

2. Carnegie Mellon University
Outline
Freeblock Scheduling
Timing Accurate Storage Emulation (TASE)
Self-*
Carnegie Mellon University

3. Carnegie Mellon University
Freeblock Scheduling
Research going on right here at CMU
Something I was involved in this past summer
Who would like some free bandwidth while their disk is busy? 
Carnegie Mellon University

4. Carnegie Mellon University
Freeblock Scheduling
Interface:
fb_read(logical numbers, …)
callback_fn(…)
Extracting Bandwidth
Send requests to the disk in between normal requests without effecting the normal requests
Carnegie Mellon University

5. Carnegie Mellon University
Freeblock Scheduling
As in SPTF scheduling, we must know the EXACT state of the disk
We need to be able to predict how much rotational latency we have to work with
Enemies of freeblock scheduling: disk prefetching internal disk cache hits unexpected disk activity (recalibration, etc) disk-reordered requests
Carnegie Mellon University

6. About to read blue sector
Goal:
Show the inefficiency of current disk requests.
Conveyed Ideas:
Rotational latency is wasted time that can be used to service tasks
Background Information:
None.
Slide Background:
Kill text and arrows
Carnegie Mellon University

7. After reading blue sector
After BLUE read
Goal:
Show the inefficiency of current disk requests.
Conveyed Ideas:
Rotational latency is wasted time that can be used to service tasks
Background Information:
None.
Slide Background:
Kill text and arrows
Carnegie Mellon University

8. Red request scheduled next
After BLUE read
Goal:
Show the inefficiency of current disk requests.
Conveyed Ideas:
Rotational latency is wasted time that can be used to service tasks
Background Information:
None.
Slide Background:
Kill text and arrows
Carnegie Mellon University

9. Carnegie Mellon University
Seek to Red’s track
After BLUE read
Seek for RED
Goal:
Show the inefficiency of current disk requests.
Conveyed Ideas:
Rotational latency is wasted time that can be used to service tasks
Background Information:
None.
Slide Background:
Kill text and arrows
SEEK
Carnegie Mellon University

10. Wait for Red sector to reach head
After BLUE read
Seek for RED
Rotational latency
Goal:
Show the inefficiency of current disk requests.
Conveyed Ideas:
Rotational latency is wasted time that can be used to service tasks
Background Information:
None.
Slide Background:
Kill text and arrows
SEEK
ROTATE
Carnegie Mellon University

11. Carnegie Mellon University
Read Red sector
After BLUE read
Seek for RED
Rotational latency
After RED read
Goal:
Show the inefficiency of current disk requests.
Conveyed Ideas:
Rotational latency is wasted time that can be used to service tasks
Background Information:
None.
Slide Background:
Kill text and arrows
SEEK
ROTATE
Carnegie Mellon University

12. Traditional components
After BLUE read
Seek for RED
Rotational latency
After RED read
Goal:
Show the inefficiency of current disk requests.
Conveyed Ideas:
Rotational latency is wasted time that can be used to service tasks
Background Information:
None.
Slide Background:
Kill text and arrows
Note: Rot. Latency is an artifact of rotation
Seeks are needed to keeps disk head on tracks
Carnegie Mellon University

13. Carnegie Mellon University
Initial setup again
After BLUE read
Goal:
Show the inefficiency of current disk requests.
Conveyed Ideas:
Rotational latency is wasted time that can be used to service tasks
Background Information:
None.
Slide Background:
Kill text and arrows
Carnegie Mellon University

14. Carnegie Mellon University
Seek to Third track
After BLUE read
Seek to Third
Goal:
Show the inefficiency of current disk requests.
Conveyed Ideas:
Rotational latency is wasted time that can be used to service tasks
Background Information:
None.
Slide Background:
Kill text and arrows
SEEK
Carnegie Mellon University

15. Carnegie Mellon University
Free transfer
After BLUE read
Seek to Third
Free transfer
Goal:
Show the inefficiency of current disk requests.
Conveyed Ideas:
Rotational latency is wasted time that can be used to service tasks
Background Information:
None.
Slide Background:
Kill text and arrows
SEEK
FREE TRANSFER
Carnegie Mellon University

16. Carnegie Mellon University
Seek to Red’s track
After BLUE read
Seek to Third
Free transfer
Seek to RED
Goal:
Show the inefficiency of current disk requests.
Conveyed Ideas:
Rotational latency is wasted time that can be used to service tasks
Background Information:
None.
Slide Background:
Kill text and arrows
SEEK
FREE TRANSFER
SEEK
Carnegie Mellon University

17. Carnegie Mellon University
Read Red sector
After BLUE read
Seek to Third
Free transfer
Seek to RED
After RED read
Goal:
Show the inefficiency of current disk requests.
Conveyed Ideas:
Rotational latency is wasted time that can be used to service tasks
Background Information:
None.
Slide Background:
Kill text and arrows
SEEK
FREE TRANSFER
SEEK
Carnegie Mellon University

18. Carnegie Mellon University
Resulting components
After BLUE read
Seek to Third
Free transfer
Seek to RED
After RED read
Goal:
Show the inefficiency of current disk requests.
Conveyed Ideas:
Rotational latency is wasted time that can be used to service tasks
Background Information:
None.
Slide Background:
Kill text and arrows
Interesting, but can apps use free bw?
Carnegie Mellon University

19. Carnegie Mellon University
Freeblock Scheduling
Results include 3.1MB/sec of free bandwidth
This free bandwidth is best suited to applications with loose time constraints
Some sample applications: - backup applications - disk array scrubbing - cache cleaning (perhaps…)
Carnegie Mellon University

20. Carnegie Mellon University
TASE
Research I’m currently involved with
Timing accurate
Can get performance measurements
Evaluate hypothetical storage devices
Without building a prototype
In real systems
Carnegie Mellon University

21. Carnegie Mellon University
TASE
Storage Evaluation Techniques
Hand calculations
Simulation
Emulation
Prototypes
Real System
Carnegie Mellon University

22. Carnegie Mellon University
TASE
System to Test
Disk
Bus
Carnegie Mellon University

23. Carnegie Mellon University
TASE
System to Test
TASE
Bus
Carnegie Mellon University

24. Carnegie Mellon University
TASE
“If it walks like a duck and talks like a duck, it must be a duck.”
Emulated device needs to “be” a disk
Respond over bus to system being tested
Behave like a disk by storing requests
Carnegie Mellon University

25. Carnegie Mellon University
TASE
Everything needs to be in physical memory
This limits what we can test with the device
Possible Solutions
Use multiple machines as emulators
Compress data
Find data that doesn’t need to be stored
Carnegie Mellon University

26. Carnegie Mellon University
TASE
Two expectations of disks
Data is accessible
It is returned correctly
Do we have to meet these expectations?
Carnegie Mellon University

27. Carnegie Mellon University
Self-*
Storage Management
Currently: 1 admin per TB
Goal is to increase to 1 admin per 1PB
What is necessary to allow this increase?
Could wait for hardware improvements
Or we could do research
Carnegie Mellon University

28. Carnegie Mellon University
Self-*
Self-*
Petabyte scale
Self-organizing
Self-managing
Self-tuning
Self-configuring
Self-repairing
Commodity hardware (heterogeneous)
Carnegie Mellon University

29. Carnegie Mellon University
Related Reading
Freeblock Scheduling
TASE
Self-*
Carnegie Mellon University

30. Carnegie Mellon University
Conclusions
Much research into improving disk access
This is just a small part of current research
Part of idea behind doing the book report
Carnegie Mellon University