1 / 20

Scheduling Algorithms in Modern Disk Drives

Scheduling Algorithms in Modern Disk Drives. B. Worthington, G. Ganger, Y. Patt. Presented by: Chiu Tan. Introduction. Why study scheduling algorithms ? Latency gap between disk and memory is high. Good algorithms can narrow this gap.

miron
Télécharger la présentation

Scheduling Algorithms in Modern Disk Drives

An Image/Link below is provided (as is) to download presentation Download Policy: Content on the Website is provided to you AS IS for your information and personal use and may not be sold / licensed / shared on other websites without getting consent from its author. Content is provided to you AS IS for your information and personal use only. Download presentation by click this link. While downloading, if for some reason you are not able to download a presentation, the publisher may have deleted the file from their server. During download, if you can't get a presentation, the file might be deleted by the publisher.

E N D

Presentation Transcript


  1. Scheduling Algorithms in Modern Disk Drives B. Worthington, G. Ganger, Y. Patt Presented by: Chiu Tan

  2. Introduction • Why study scheduling algorithms? • Latency gap between disk and memory is high. • Good algorithms can narrow this gap. • Disk traffic is bursty, resulting in long queues of requests. Algorithms can reschedule requests with respect to disk state to improve performance.

  3. Outline • Disk basics • Disk Evaluation • Common Algorithms • Simulation Issues • Findings • Conclusion

  4. Disk Basics Sector (512 bytes + error correction) Track (One of many on a surface) Cylinder(Set of tracks, one from each surface)

  5. Disk Basics (II) • How does a disk service a request? • Read/Write performed by a disk arm. • Disk arm has to first find the correct cylinder. This is known as the seek time. • The disk is spinning. The arm waits until the correct sector reaches the read/write head of the disk arm. This is the rotational latency. • Generally, multiple surfaces, but only 1 read/wrote head is active.

  6. Disk Evaluation • Disk performance primarily measured using access time. • Access time comprises of seek time and rotational latency. • Seek time: Time needed to move disk head to correct cylinder. • Rotational latency: Time needed for disk to rotate to correct sector.

  7. Common Algorithms • To improve response time, we can reduce seek time or reduce rotational latency. • Simplest algorithm: FCFS • Seek Time: SSTF. SCAN, LOOK, CLOOK • Seek + Rotational Latency: SPTF • To illustrate: 5,9,18,3,12,1,13,6,7

  8. FCFS: First come first serve 5, 9, 18, 3, 12, 1, 13, 6, 7 Total: 63

  9. SSTF: Shortest Seek Time 5, 9, 18, 3, 12, 1, 13, 6, 7 Total: 30

  10. SCAN 5, 9, 18, 3, 12, 1, 13, 6, 7 Total: 23

  11. LOOK 5, 9, 18, 3, 12, 1, 13, 6, 7 Total: 21

  12. CLOOK 5, 9, 18, 3, 12, 1, 13, 6, 7 Total: 33

  13. SPTF: Shortest Positioning • Reducing seek delay needs only relative seek distances. Easy to estimate. • SPTF is like SSTF, but need to minimize seek time and rotational latency. • Need actual LBN-to-PBN number, accurate position of read/write head.

  14. Simulation Issues • Synthetic workload and generated traces used. • Synthetic workloads are easier to generate, but unrealistic. • Traces are more realistic, but adjustments needed. • Different storage capacity. • Different service rates.

  15. Simulation Issues (II) • 2 metrics are used: average response time and squared coefficient of variance • A balance between the two is most desirable. Faster performance with lesser starvation.

  16. Findings: Scheduling via LBN • For synthetic workloads, CLOOK average response time is 5% worst than LOOK, SSTF, VSCAN. • But CLOOK outperformed the rest in terms of starvation resistance. • For traces, CLOOK on average outperforms the rest, and on starvation resistance as well.

  17. Insights: Scheduling via LBN Why the difference between the two? • Synthetic workloads are random, traces are not. • Random read/writes fail to take advantage of prefetching cache. • Algorithms that preserve read sequentiality can take advantage of perfetching cache to perform better!

  18. Findings: via Full Knowledge • With full knowledge, SPTF considered. Aging component is added to prevent starvation. • For synthetic workloads, all SPTFs perform better than CLOOK. Appropriate aging results in SPFT with CLOOK starvation performance. • For traces, clear wins for SPTFs over CLOOK for some traces, ambiguous for others.

  19. Insights: via Full Knowledge • SPTF-based algorithms perform well compared to CLOOK, but clearer wins occur when on-board cache is exploited! • Assign positioning time to zero if found inside cache. ASPCTF and SPCTF. • Good choice of aging weights have potential to have superior performance and starvation resistance.

  20. Conclusion • CLOOK performs best under real world situations under limited knowledge. • SPTF-based algorithms result in better performance, but results not that much better than CLOOK. • Careful aging and exploitation of cache needed for superior performance. • Algorithms have to exploit perfetching cache to improve performance regardless.

More Related