New(old) problems in HDDs: How to cope with them ?
Heon Y. Yeom Seoul National University

Disk I/O Performance

Read 302-th Record Write xxx to 10-th Record Read 777-th Record
Scheduling Device system File call System /Reordering Driver g

Too long time !

Disk Mechanism
I/O performance sucks. Why? Head switch, Platter rotation. ,
Disks require Mechanical Movement to access stored data. So, Reducing the mechanical delay is important. This begins by an exact modeling of physical disk layout layout. The Key challenge is that hard-disk itself is a black-box.
What has changed in 25 years ?
Smaller size Larger capacity About the same speed
seek time + rotational delay + transfer time

More intelligence

TCQ, NCQ

Are they good ?

What does OS expect from HDD ?
Linear array of fixed size blocks.
not really y (Cylinder, Head, Sector).
I want to read values, fast. I want to write values, f i l fast, without l ih loss.
=> disk scheduling algorithms
I want some performance g p guarantee !!

Disk layout information

Is hidden. Different from disk to disk. It is possible but,
Takes 2-3 days for todays disk.
If we have the information,
It is possible to predict exact timing.

Standards needed.

LTA CHS
Track Layout Gathering Example
Head Track T k Track Switch Time (A, A+k) A A+1 A+2 A+k-1 A+k A+1 A+k A+k+1 A+2k

Track Switch Time (A, x)

Possible Mapping : 2 platters

A A+1 A+k

Track Switch Time (A+k, x)
A+2k-1 A+k+1 A+1 A+k A+k+1 A+2k
Samsung Disk hd400ld layout

0 4797

4797 4797

4796 9593

14391 9594

19187 14390

19188 23985

23984 28781

57564 104331

4797 4954

62360 109284

62361 117598

67157 122551

67158 132460

71954 137413

109285

114238

122552

127505

127506

132459
Performance Evaluation (Samsung hd400ld)
Filebench Trace with 1,000 requests (Queue size : 10)
Cylinder distance C li d di

100 120

LBA-based Scheduling
Total cyl distance Average Wait Time (usec) Average Service Time (usec)

Request Number

CHS-based Scheduling
3,147,948 (32.7%) 9,532 (86.8 % ) 97,133 (87.1 % )

9,611,435 111,437

10,975

Is it enough ?

Disk Model Interface Capacity Cylinder
LBA-based Average Response Time CHS-based Average Response Time
Wd2500jb IDE 250GB 6 10,209 9,567(93.7%) 35.3% St330007lw SCSI 300GB 8 4,048 3,809(94.1%) 33.7% St314007lw SCSI 146GB 4 5,113 5,000(97.8%) 39.1%
Hd400ld IDE 400GB 6 10,975 9,532(86.8%) 32.7%

Seek Distance ( ) (%)

We can probably do better inside the disk p y
Here it comes Native Command Queueing(NCQ)

*2.5 rotations

* 1.5 rotations

From Wikipedia

Problem 1 - Redundant Scheduling 1.
Case 1. No synergy effect effect.
I/O Scheduler + NCQ I/O Scheduler(or NCQ) ( )
Case II. Sometimes, its worse !
I/O Scheduler + NCQ < I/O Scheduler(or S h d l ( NCQ)
Problem 1. Redundant Scheduling. g
Case I (anticipatory, ( p y deadline noop) Case II (anticipatory case) C ( ti i t )
Problem 2. When requests are reordered, some requests are starved ! t t d
Solution for starvation Control the tagging
Count the number of reordering events. If exceeds threshold, stop the NCQ mechanism. Modification to the device driver.
Request Starvation Detection Architecture

* Linux 2.6.22

S-ATA 2
* Add Request Starvation Detection

Conclusions

Revisit the old HDD performance problem. f
Utilizing the internal disk layout information. It is cumbersome Can the disk manufacturers help ?
Identified a new problem.
NCQ is great, but Its not panacea p
Some cooperation would be great.