early lessons from energy-saving storage systems research

Early Lessons from Energy-Saving Storage Systems Research

Sara Alspaugh and Arka Bhattacharya

Saving Energy At the Block Device Level

SSD Opportunities and Challenges

Exploring the Design Space

Takeaways

Redundancy: reconstruct request from parity group on active disks and cache

(Examples: RIMAC, Diverted Access, PARAID, EERAID)

Caching and Layout: keep less frequently accessed data on spun down disks, eject blocks to active disks

(Examples: MAID, PA-LRU, PB-LRU, Hibernator, EED, PDC)

Multispeed Disks: adjust disk speed according to workload demands and performance requirements

(Examples: Hibernator, DPRM)

Device Idle Seq. Read Seq. Write Ran. Read Ran. Write

Corsair F60 (Consumer SSD)

0.70 W 0.80 W 228Mb/sec

0.99W 195Mb/sec 0.75 W 229 Mb/sec 0.98 W 194 Mb/sec

Intel X25-E Extreme (Enterprise SSD)

0.60 W 1.51 W200Mb/sec

2.70 W158Mb/sec

1.51W201 Mb/sec

2.71 W186 Mb/sec

Seagate Momentus XT (Hybrid SSD)

1.63 W 3.46W104Mb/sec

3.03W90Mb/sec

3.41 W 96 Mb/sec

2.97W88 Mb/sec

Western Digital 1600BEVS (HDD)

2.26 W 2.27W59Mb/sec

3.34W43Mb/sec

3.58W36 Mb/sec

3.17W36 Mb/sec

Scheme Block Device / RAID Level

File System Level

Fixed Thresh-hold

Pred-ictive Erasure Codes (RAID5)

Mirror-ing (RAID1)

Write Logging

Access Freq.-Based Layout

Solid State Devices

Multi-speed Disks

Hybrid / Tiered

DIV-ACC X X X X

EERAID1 X X X X

EERAID5 X X X X X

RIMAC X X X X

PARAID X X X X

PDC X X X

PA-LRU X X X

PB-LRU X X X X

HIBERN X X X X X X

DPRM X X X X

WOL X X X X X X

MAID X X X X

SSD-RAID X X X X X X

EED X X X X X X

SIERRA X X X X

RABBIT X X X X X

Effects of Workload Characteristics

Choosing the Best Scheme

Left: Idle periods longer than 10 seconds represent opportunities to spin down disks to save power.

Right: Most workloads display a high degree of skew. In this MSR build server trace, 75% of the requests go to 35% of the blocks.

Increased skew results in decreased latency, but write-dominated workloads have lower latencies than their read-dominated counterparts – an artifact of DiskSim.

When caching-based techniques are used especially with SSDs, power and performance are correlated.

Comparison of 95th percentile latencies for varying cache replacement algorithms and number of SSDs.

Most promising: hybrid of SSDs and traditional disks, place most frequently accessed data on SSDs (does not preclude use of redundancy-based techniques)

Cache- and layout-based strategies that reduce power consumption also improve performance; redundancy-based strategies do not have this effect

Assumptions about the workload play an important role

SSDs not a panacea: cost, black box flash translation layer

Future WorkOptimizations: prefetching, zero-copy

Using file system and above level knowledge (super challenge: co-located storage and computation)

Considerations of cost and capacity requirements