Shignled disk

Shingled Disk

The Big Data StorageStorage: Magnetic disks

High storage densityCurrent: 400GB/in2 - 550GB/in2

30-50% increase per year.

It’s reaching its physical limit…“Superparamagnetic Limit”Predicted limit: around 1TB/in2

Review: Hard disk

Track

Conventionally Written Track

Non-overlap

Track width w (e.g. 25nm)

Guard gaps between tracks (e.g. g = 5nm).

Bottleneck is the writing track width.Current read heads can work on

much narrower track.But it is hard to write narrower track.

Shingled Disk: Overlap Tracks

Wilder Track written. (e.g. w = 70nm).

Shingled writing overlaps tracks.

The remaining residual track could be much narrower. (e.g. r = 10nm).

CharacteristicsHigher Density without significant hardware

change.2-3 times of the conventional disk density.

Support Random Read / Sequential WriteA single write will destroy the next k tracksTypically, k = 4~8

Can we do better than a “tape with random read support”?

Two High-Level StrategiesMask the operational difference of a Shingled

Disk.Drop-in replacement for current disks.Uses the standard block interface.

Specialized file system with no/little hardware mask.More flexibility in the data layout and block

management. Increased knowledge at file system layer.

Strategy One: Masking the Operational Difference

Synergy with SSD: Slow erasure of block in SSDSSD: Flash Translation Layer (FTL)Shingled Disk: Shingled Translation Layer (STL)Translate from Virtual Block Address to Logical Block

Address on disk.How to perform random write.

One extreme: Read-modify-write.Another extreme: Remap the physical location of

written data.Benefit

No change for user and system.“Drop-in” replacement for current system.

Strategy One: Masking the Operational Difference

DrawbackExperience with SSDs indicates the performance

will be hard to predict.Reverse engineering on SSD to achieve higher level

goals.Sophisticated STL could be expensive.Data stored in Continuous Virtual Block Address

could be far away on disk.Database table with frequent edits.Concurrent downloads of movies.Might use large NVRAM (as cache) to mitigate the

problem.

Virtual Block Address Translation

Need to quickly translate Virtual Block Address to Logical Block Address.

Translation Table could be very large.Capacity 2T, each entry 8 bytes.Block Size 4K: Translation Table 4GB.Block Size 512 bytes: Translation Table 32GB.

Some B+-Tree type structure.

Strategy Two: Specialized System

Simple Shingled Translation Layer.Random Update: Read-Modify-Write.TRIM Command: Tell hardware that overwriting

subsequent tracks is fine.Support to format some part of the disk unshingled.

More Sophisticated System SoftwareAvoid writing to the middle of the band.Conceptualize writing as appending to a log.Perform necessary data remapping and garbage

collection.

Design Issues for Shingled Write Disk

Band AbstractionStore the bulk of data in band.

A collection of b-contiguous tracks.A buffer of k tracks at the end of each band.Bands are not interfered with each other

More flexible.

Proportional Capacity Lossc = 1 – b/(b+k): proportional capacity loss

• k = 5 and want to control c < 0.1• b > 45• Each band have

67.5MB• Reasonable for

modern LFS.

Band Usage1. Only writes complete bands.

Each band contains a segment of Log structured File System. Assumes data is buffering in NVRAM.

2. Only appends to bands. Less efficient.

3. Circular Log inside each band Consume data from head. Append data to the tail. Require additional k track gap between head and tail.

4. Flexible band size Neighboring bands could be joined. Not suitable for a general purpose SWD: Just for

completeness.

Reserved Space for Random Update

Option 1: NVRAM. Option 2: Random Access Zone (RAZ)

Every track is followed by k unused tracks.Density of RAZ is lower than current disk.

How Large Random Access Zone Can we

Have?

• If we want to guarantee L = 2 times of the conventional disk.• k = 5: 3.75% of

total storage capacity.

Assume without RAZ, the capacity of shingled disk is 2.3 times of conventional disk.

Trade offs for Two OptionsReserved Space for Random Access

Option 1: NVRAMFasterMore Expensive: cost 10 times to RAZ

Option 2: Random Access Zone (RAZ)Use some part of the disk for Random Access Zone.Cheaper but slower.

Trade-offs would be interesting.

Usage of NVRAM1. Buffering data for writing bands.

Be careful about the limited number of write-erase cycles of flash memory.

2. Use NVRAM to store metadata. Metadata tends to have a higher amount of activity. In Write Anywhere File System, NVRAM could be used

to maintain the log of file system activities.

3. Store recently created objects. Temporal locality: a block/object created long long

ago is less likely to be updated. If data is first written to NVRAM, we can also have

better placement of data on disk.

Number of LogsHere Log-structured File System is assumed to

be used.

What’s the benefit to have more than a single log?1. Separation between metadata and data.

E.g. Access Time.

2. Allocate files for more efficient read access later. E.g. Downloads several movies at the same time.

If only one log, all the movie objects will be interspersed.

Inefficient for read.

Workload for General Purpose Personal Usage

Workloads EvaluationRate of Block Updates

If few blocks are updated frequently.Less need for Random Access Zone / NVRAM.Shingled Disk is more usable to replace conventional

disk.

Evaluated Workloads1. General purpose personal usage for 1 month

2. Specialized workload: Video Edit for 3 hours.

3. Specialized workload: Music Library Management. Negligible block update. Not surprised.

Some Points From Workload

1. Identifying hot blocks is important since the volume of hot blocks is small enough to be held in the Random Access Zone / NVRAM.

2. Larger block sizes reduces the accuracy of identifying hot blocks. But it’s not that significant.

3. File system distinguish between metadata from user data would be helpful.

Workload for Disk Arrays

Shingled Disk ArraysCan the shingled disk used in a server environment?

Probably part of the disk array. It could have writing originating from different sources.

Two Impacts in Workload Data Striping. Workload Interleaving.

Replay workloads against a simulated drive. Log-structured Writing Scheme to perform in-band

update.

Logical Arrangement of Blocks