Paper 2: Douglis, F., Kaashoek, F., Marsh, B., Caceres, R., Kai Li, Tauber, J.A

Storage alternatives for mobile computers

Proceedings of the First USENIX Symposium on Operating Systems Design and Implementation (OSDI), Monterey, CA, USA, 14-17 Nov. 1994. p. 25-37

Presentation slides

Goal

• Measure power consumption & performance of 3 storage devices
  • Magnetic Disk
  • Flash Disk (FDisk) Emulator (block interface)
  • Flash Card (FCard) - using Microsoft Flash File System (MFFS) - (byte interface)
• Analyze techniques to improve performance.

Motivation

• Storage devices consume a large fraction of power (upto 40%) in mobile devices.

Tradeoffs between the devices

• Magnetic Disks: (+) Good performance (especially for large writes); (+) Cheap: $1-5/MB; (-) Consumes lot of power; (-) Disk spin-up involves a huge latency.
• Flash Disks: (+) Fast reads; (+) lower power consumption than magnetic disks; (-) Large write latency: should erase previously written data; (-) High cost: $30-50/MB; (-) Cannot be used reliably after a certain number of erasures.

Raw Hardware Measurements

• They measured the raw bandwidth & latency of the 3 devices by performing read/write of files.
• Results:
  • Magnetic Disk is best for throughput.
  • Large writes are bad on Flash Card due to MFFS anomaly.
  • Flash card is better than flash-disk for small files - due to byte interface.
  • Flash card is worse than flash-disk for large files.

Statistics from Manufacturers

• Disk spin-up takes 1s
• Flash card erase 1.6s
• Disk uses an order of magnitude more than either flash devices which use the same amount of power

They next performed trace-driven simulations to study the trade-off between different optimizations.

Traces

• They used different machine's workload to get the traces.
• Recorded: filename, read/write, size, offset, time taken.

Simulations

• They assumed buffer cache & write-through policy.
• What they computed:
  • power consumption in each mode (reading, writing, sleeping, idle)
  • time taken for operations
  • time to switch between modes

Simulation Results

1. Basic raw data
   • Fcard write worse than magnetic-disk write
   • Fdisk better than Fcard for writes (although datasheet indicates the opposite)
2. Storage utilization on Fcard
   • As storage utilization goes up from 40% to 95%, (a) write response becomes worse by 30% (b) power consumption increases by 150%.
3. Decoupled Erasure/Writes on Fdisk
   • Erase bandwidth is only 150Kb/s. Write bandwidth on pre-erased area is 400Kb/s.
   • Next generation of Fdisk products will have the asynchronous cleaning feature.
   • Asynchronous cleaning has minimal impact on power consumption.
   • Write performance could be improved by 56-61%.
4. DRAM caching for Flash devices
   • DRAM performs better than Flash but is volatile & more importantly, consumes more power.
   • What is the trade-off between adding more DRAM chips versus adding more Flash memory?
   • Results: DRAM has almost no effect on performance (response time) but increases power utilization (hence, no benefit in adding DRAM).
5. NVRAM caching
   • What is the impact of NVRAM on write performance (gain due to deferred spin-up)?
   • Simulation by using NVRAM between disk & buffer cache
   • Results: Energy consumption goes down slightly, Response time improves dramatically
   • Conclusion: 32Kb NVRAM is a good thing to have