1. An Introduction to Disk Drive Modeling Chris Ruemmler & John Wilkes Hewlett-Packard Laboratories Presented by Hang Zhao

2. The earliest hard disks … First Hard Disk (1956): IBM's RAMAC, capacity is 5 MB, stored on 50 24" disks First Hard Disk (1956): IBM's RAMAC, capacity is 5 MB, stored on 50 24" disks First Air Bearing Heads (1962): IBM's model 1301 increases both areal density and throughput by about 1000% First Air Bearing Heads (1962): IBM's model 1301 increases both areal density and throughput by about 1000% First Removable Disk Drive (1965): IBM's model 2310 First Removable Disk Drive (1965): IBM's model 2310 …… ……

3. A brief history of hard disk drive Engineers over the last few decades have done at improving them in every respect: reliability, capacity, speed, power usage, and more. Engineers over the last few decades have done at improving them in every respect: reliability, capacity, speed, power usage, and more.

4. Evolution of areal density of hard disk platters The areal density of hard disk platters continues to increase at an amazing rate even exceeding some of the optimistic predictions of a few years ago. Modern disks are now packing as much as 20 GB of data onto a single 3.5" platter! The areal density of hard disk platters continues to increase at an amazing rate even exceeding some of the optimistic predictions of a few years ago. Modern disks are now packing as much as 20 GB of data onto a single 3.5" platter!

5. Why do we need to model disk drive behavior? CPU technology is advancing rapidly; while the overall system behavior is restrict to disk system performance. CPU technology is advancing rapidly; while the overall system behavior is restrict to disk system performance. The behavior of disk drive itself is a dominant factor in overall I/O performance. The behavior of disk drive itself is a dominant factor in overall I/O performance. Existing hard disk models have limitations … Existing hard disk models have limitations …

6. Characteristics of modern disk drives Mechanism Mechanism recording component recording component positioning component positioning component Controller Controller microprocessor microprocessor buffer memory buffer memory interface to SCSI bus interface to SCSI bus

7. Disk Drive Terminology Several platters, with information recorded magnetically on both surfaces (usually) Several platters, with information recorded magnetically on both surfaces (usually) Actuator moves head (end of arm,1/surface) over track (seek), select surface, wait for sector rotate under head, then read or write Actuator moves head (end of arm,1/surface) over track (seek), select surface, wait for sector rotate under head, then read or write Cylinder: all tracks under heads Cylinder: all tracks under heads Bits recorded in tracks, which in turn divided into sectors Bits recorded in tracks, which in turn divided into sectors Platter Outer Track Inner Track Sector Actuator Head Arm Platter

8. The recording components Modern disks range in size from 1.3 to 8 inches in diameter: Modern disks range in size from 1.3 to 8 inches in diameter: Smaller disks VS larger disks Smaller disks VS larger disks less surface area/storage less surface area/storage consume less power consume less power spin faster spin faster smaller seek distance smaller seek distance

9. The increased storage density The incremental trends result from better linear recording density: a measure of how tightly the bits are packed within a length of track. (50,000 BPI 1994; 524,000 BPI 2000 ) better linear recording density: a measure of how tightly the bits are packed within a length of track. (50,000 BPI 1994; 524,000 BPI 2000 ) packing separate tracks more closely together (20,000 TPI; 67,300 TPI around 2000 ) packing separate tracks more closely together (20,000 TPI; 67,300 TPI around 2000 )

10. Platters and disk rotation Platters rotates in lockstep on a central spindle at rates varying from 3,600 to 7,200 rpm Platters rotates in lockstep on a central spindle at rates varying from 3,600 to 7,200 rpm Higher spin rate increases transfer rates and shortens rotation latencies; on the other hand, power consumption increases Higher spin rate increases transfer rates and shortens rotation latencies; on the other hand, power consumption increases Each platter surface is associated with a disk head for writing and reading operating under a single read-write channel Each platter surface is associated with a disk head for writing and reading operating under a single read-write channel

11. The positioning components Seeking: speed of head movement, limited by the power available for pivot and the arm s stiffness. Seek time is composed of: Seeking: speed of head movement, limited by the power available for pivot and the arm s stiffness. Seek time is composed of: speedup speedup coast: for long seeks where arms move at max υ coast: for long seeks where arms move at max υ slowdown slowdown settle: dominant factor of very short seeks settle: dominant factor of very short seeks

12. The demerit of average seek time Average seek times are commonly used as a figure of merit for disk drives, but they can be misleading. Average seek times are commonly used as a figure of merit for disk drives, but they can be misleading. Independent seeks are rare in practice. Independent seeks are rare in practice. Shorter seeks are much more common. Shorter seeks are much more common. The one-third-stroke calculation is only applicable for completely independent seeks. The one-third-stroke calculation is only applicable for completely independent seeks. N-1 weighted seek time calculation provides better approximation. N-1 weighted seek time calculation provides better approximation. Seek-time-versus-distance profile matters for modeling!

13. The track following system Fine-tuning the head position at the end of a seek and keeping the head on the desired track. Fine-tuning the head position at the end of a seek and keeping the head on the desired track. Performing a head switch from one surface to the next in the same cylinder. Performing a head switch from one surface to the next in the same cylinder. Aggressive and optimistic approach applied to head settling before a read operation. Aggressive and optimistic approach applied to head settling before a read operation.

14. Data layout in SCSI disk Disk appears to client as a linear vector of addressable blocks, which are mapped to physical sectors on the disk. Disk appears to client as a linear vector of addressable blocks, which are mapped to physical sectors on the disk. Zoning: adjacent cylinders are grouped into zones Zoning: adjacent cylinders are grouped into zones Track skewing: logical sector zero on each track is skewed for fast sequential access across track and cylinder boundaries. Track skewing: logical sector zero on each track is skewed for fast sequential access across track and cylinder boundaries. Sparing: map flawed sectors to other locations Sparing: map flawed sectors to other locations

15. The disk controller Mediates access to the mechanism Mediates access to the mechanism Runs the track-following system Runs the track-following system Transfers data between disk drive and its client Transfers data between disk drive and its client Manages the embedded cache Manages the embedded cache

16. Bus Interfaces

17. Caching requests Read ahead Read ahead Write caching Write caching Command queuing Command queuing

18. Modeling disk drives Disk drive cannot be modeled analytically with any accuracy due to its nonlinear, state- dependent behavior. Disk drive cannot be modeled analytically with any accuracy due to its nonlinear, state- dependent behavior. Limitations for current modeling strategies: Limitations for current modeling strategies: Seek times modeled as a linear function of seek distance Seek times modeled as a linear function of seek distance Rotational latency follows uniform distribution Rotational latency follows uniform distribution Media transfer time ignored or as fixed constant Media transfer time ignored or as fixed constant Bus contention often ignored Bus contention often ignored

19. The simulator and traces Event-based simulator in C++ Event-based simulator in C++ Disk drive is modeled as two tasks and some additional control structure Disk drive is modeled as two tasks and some additional control structure Representative samples from a longer trace series of HP-UX was selected Representative samples from a longer trace series of HP-UX was selected Two HP disk drives Two HP disk drives HP C2200A for non-caching disk drive HP C2200A for non-caching disk drive HP 97560 for caching disk drive HP 97560 for caching disk drive

20. Simulation model structure

21. Evaluation Metric: the root mean square of the horizontal distance between real drive curve and model curve Metric: the root mean square of the horizontal distance between real drive curve and model curve The demerit was presented in both absolute term and relative term The demerit was presented in both absolute term and relative term

22. Evaluation cont. Non-linear seek time profile is added to the 3 rd model in Figure c. Non-linear seek time profile is added to the 3 rd model in Figure c. The cost of head and track switching was also included The cost of head and track switching was also included Rotational latency and spare – sector placement were added to the final model in Figure d. Rotational latency and spare – sector placement were added to the final model in Figure d.

23. Evaluation cont.

24. Summary of disk drive model Factors not included in the final model: Factors not included in the final model: Soft-error reentries Individual spared sectors or tracks

25. Impact of proposed disk drive model 269 citations found with most recent reference Including Models of Black-Box Storage Arrays- Terence Kelly Ira (2004) 269 citations found with most recent reference Including Models of Black-Box Storage Arrays- Terence Kelly Ira (2004) A A Stochastic Disk I/O Simulation Technique, Winter Simulation Conference, 1997 Modeling hard disk power consumption by Princeton, published in FAST 03 Modeling hard disk power consumption by Princeton, published in FAST 03