Anticipatory Disk Scheduling SOSP 2001

1
Anticipatory Disk SchedulingSOSP 2001

• Sitaram Iyer Peter Druschel
• Rice University

2
Operating system
App
App
I/O
FS
VM
Disk Scheduler
Disk Driver
disk subsystem
3
Disk scheduler

• Reorders available disk requests for
• performance by seek optimization,
• proportional resource allocation, etc.
• With multiple outstanding requests available, any
  disk scheduling policy can take appropriate
  decisions.

4
With enough requests
issued by process A
issued by process B
time
seek
location on disk
Throughput 21 MB/s (IBM Deskstar disk)
5
With synchronous I/O
schedule
Throughput 5 MB/s
Next
6
Deceptive idleness

• Process A is about to issue next request.
• but
• Scheduler hastily assumes that process A has no
  further requests!

7
Proportional scheduler

• Allocate disk service in say 12 ratio

• Deceptive idleness causes 11 allocation

Next
8
Anticipatory scheduling

• Key idea Sometimes wait for process whose
  request was last serviced.
• Keeps disk idle for short intervals.
• But with informed decisions, this
• Substantially improves throughput
• Achieves desired proportions

9
Cost-benefit analysis

• Balance expected benefits of waiting
• against cost of keeping disk idle.
• Tradeoffs sensitive to scheduling policy
• e.g., 1. seek optimizing scheduler
• 2. proportional scheduler

10
Statistics

• For each process, measure
• 1. Expected median and 95percentile thinktime
• 2. Expected positioning time

Number of requests
Thinktime
Median
95percentile
last
next
11
Cost-benefit analysisfor seek optimizing
scheduler

• best best available request chosen by
  scheduler
• next expected forthcoming request from
  process whose request was last serviced

12
Proportional scheduler

• Costs and benefits are different.
• e.g., proportional scheduler
• Wait for process whose request was last serviced,
• 1. if it has received less than its allocation,
  and
• 2. if it has thinktime below a threshold (e.g.,
  3ms)
• Waiting_duration next.95percentile_thinktim
  e

13
Prefetch

• Overlaps computation with I/O.
• Side-effect
• avoids deceptive idleness!
• Application-driven
• Kernel-driven

14
Experiments

• FreeBSD-4.3 patch kernel module
• (1500 lines of C code)
• 7200 rpm IDE disk (IBM Deskstar)
• Also in the paper
• 15000 rpm SCSI disk (Seagate Cheetah)

15
Microbenchmark
25
Original
20
Anticipatory
15
Throughput (MB/s)
10
5
0
Sequential
Alternate
Random within file
16
Real workloads

• Whats the impact on real applications and
  benchmarks?
• Andrew benchmark
• Apache web server
• (large working set)
• Database benchmark

• Disk-intensive
• Prefetching enabled

17
Andrew filesystem benchmark
2 (or more) concurrent clients
Lower is better

• Overall 8 performance improvement

18
Apache web server

• CS.Berkeley trace
• Large working set
• 48 web clients

19
Database benchmark

• MySQL DB
• Two clients
• One or two databases on same disk

20
Intelligent adversary
no prefetch
20
21
Proportional scheduler
Database benchmark two databases, select queries
22
Conclusion

• Anticipatory scheduling
• eliminates deceptive idleness,
• achieves significant performance improvements on
  real applications,
• adheres to desired proportions,
• and is easy to implement!

23
Anticipatory Disk Scheduling

• Sitaram Iyer Peter Druschel
• http//www.cs.rice.edu/ssiyer/r/antsched/

Next Memory Management...
24
Memory Mgmt Superpages

• TLB coverage potential benefits
• Fragmentation, contiguity mgmt
• Transparent, practical solution
• achieves and sustains improvements
• 30-60 in standard benchmarks
• low pathological case overheads
• with Juan Navarro, Peter Druschel.
• submitted to OSDI 2002.

25
Memory Mgmt via Resizing

• Many modern apps are resizable
• Shrink selectively under pressure
• Lookahead window based solution
• Automatic memory configuration
• High memory utilization
• System robustness in low memory
• in preparation

26
Peer-to-peer web caching
(Separate talk!) with Antony Rowstron,
Peter Druschel. PODC 2002.
27
(No Transcript)
28
GnuLD
Backup
Concurrent 68 execution time reduction