Towards Higher Disk Utilization: Extracting Free Bandwidth From Busy Disk Driver

1
Towards Higher Disk UtilizationExtracting Free
Bandwidth From Busy Disk Driver

• Christopher R. Lumb, Jiri Schindler,
• Gregory R. Ganger, David F. Nagle
• Carnegie Mellon University
• Erik Riedel
• Hewlett-Packard Labs

2
Content

• Introduction
• Free Bandwidth
• Availability of Free Bandwidth
• Freeblock Scheduling Decisions
• Free cleaning of LFS segments
• Free data mining on OLTP systems
• Related Work
• Conclusions

3
Introduction

• Disk head usage for several modern disk

4
Free Bandwidth

• Disk media access time

Tacess Tseek Trotate Ttransfer
5
Free Bandwidth

• Considerations of Freeblock scheduling
• How much rotational latency will occur before
  the next foreground media transfer
• Accurate decision is based on overall positioning
  overheads
• Seek Time Rotational Time
• Using Free Bandwidth
• Low priority
• Large sets of desired block
• No particular order of access
• Small working memory footprints

6
Availability of Free Bandwidth

• Impact of disk characteristics
• Impact of workload characteristics

7
Availability of Free Bandwidth

• Impact of scheduling algorithm

-First-Come-First-Served(FCFS) -Circular-LOOK(C-LO
OK) -Shortest-Seek-Time-First(SSTF) -Shortest-Posi
tioning-Time-First(SPTF)
8
Availability of Free Bandwidth

• SPTF vs. SSTF

9
Freeblock Scheduling Decisions

• Freeblock scheduling
• Identifying freebandwidth and matching them to
  pending freeblock request
• Maintains separate queue
• Decision step
• for each track on the disk, how many desired
  blocks could be accessed in this opportunity?
• Computing extra seek time
• Determining whitch disk blocks will pass under
  the head during remaining rotational latency time

10
Free cleaning of LFS segments

• cleaning operation
• Experimental Setup
• LDD(log-structured logical disk)
• Segment 128 Blocks ( each 4KB-Block)
• Run under Linux 2.2.14 with DiskSim
• DiskSim disk simulator
• Configured to model a modified Quantum Atlas 10K
• Merge LDD with DiskSim

dead
active
active
active
active
active
11
Free cleaning of LFS segments

• Experiments
• Using the Postmark v.1.11 benchmark
• Transaction
• read/write(11)
• creation/deletion(11)
• 25000 transaction
• File size 5-8KB file
• Create 100 subdirectories
• Results
• FREEBLOCK
• Show slow divergence form IDEAL

12
Free data mining on OLTP systems

• Goal
• With freeblock scheduling, significant mining
  bandwidth can be extracted from the original
  system without affecting the original transaction
  processing activity
• Experimental Setup
• DiskSim simulator
• Configured to model the Quantum Atlas 10K
• Synthetic foreground work load
• Approximation of OLTP workload characteristics
• With request per 30 milliseconds think-times
• MPL(multiprogramming level) activate request at
  any given point
• Read/write request (21)
• each request size is n4KB (mean is 8KB)
• Background data mining workload using free
  bandwidth (full scan 4KB)

13
Free data mining on OLTP systems

• Result(1)
• Low OLTP loads result in low data mining
  throughout
• As foreground request is increase, freeblock
  requests are more plentiful
• When the Freeblock request is not in start or
  destination track throughput is decreased

14
Free data mining on OLTP systems

• Result(2)
• The efficiency of freeblock scheduling drops
  steadily as the set of still-desired background
  blocks shrinks
• The extra seek-time increase
• Unused rotational latencies also increase

15
Free data mining on OLTP systems

• Result(3)
• One solution is to increase the priority of the
  last few freeblock request
• Alternate solution is using statistical nature of
  many data mining queries ( only subset of total
  data is used)
• Abort freeblock request when enough of data set
  has been mined

16
Related Work

• Trend
• Reduce mechanical positioning overhead to
  amortize these overheads over large media
  transfer
• All these approaches increase disk head
  utilization for foreground workloads
• Related work to extraction and use of free
  bandwidth
• Dynamic set and disk-directed I/O interface
• Characteristics of background workload easily
  utilize free bandwidth
• Idle time detection algorithms
• Freeblock scheduling complements exploitation of
  idle time
• Ability to make progress during busy periods
• Ability to make progress during no impact to
  foreground disk access times
• Piggybacking
• A free prewriting mechanism

17
Related Work

• Related work to extraction and use of free
  bandwidth
• Addressing the use of free bandwidth and related
  needs such as CPU, bus resources
• Eagar writing
• Remapping new versions of disk block to free
  locations very near the disk head

18
Conclusions

• By serving background requests in the context of
  mechanical positioning for normal foreground
  requests, 2050 of a disks potential media
  bandwidth can be obtained with no impact on the
  original request
• Future work
• To refine and realize freebock scheduling in
  practice.
• whether freeblock scheduling can be implemented
  outside of modern disk driver, given their
  high-level interface and complex firmware
  algorithms
• More advanced freeblock scheduling algorithms
• Deal with request fragmentation, starvation and
  priority mixes