Hybrid SMR Drives Fenggang Wu , Bingzhe Li, Zhichao Cao, Baoquan - - PowerPoint PPT Presentation

▶
hybrid smr drives
SMART_READER_LITE
LIVE PREVIEW

Hybrid SMR Drives Fenggang Wu , Bingzhe Li, Zhichao Cao, Baoquan - - PowerPoint PPT Presentation

Mar 16, 2023 276 likes •627 views

ZoneAlloy : Elastic ZoneAlloy Data and Space Management for Hybrid SMR Drives Fenggang Wu , Bingzhe Li, Zhichao Cao, Baoquan Zhang Ming-Hong Yang, Hao Wen, David H.C. Du University of Minnesota, Twin Cities Jul. 8, 2019. HotStorage19 C

slide-1
SLIDE 1

Center for Research in

Intelligent Storage

ZoneAlloy: Elastic Data and Space Management for Hybrid SMR Drives

Fenggang Wu, Bingzhe Li, Zhichao Cao, Baoquan Zhang Ming-Hong Yang, Hao Wen, David H.C. Du University of Minnesota, Twin Cities

  • Jul. 8, 2019. HotStorage’19

ZoneAlloy

slide-2
SLIDE 2

Center for Research in

Intelligent Storage

Shingled Magnetic Recording (SMR)

Platter Read/Write Head Tracks Traditional non-overlapping track design Rotational Disk Shingled tracks SMR Technology Shingled Magnetic Recording: + enables higher data density by overlapping data tracks.

  • requires careful data handling when updating old blocks.
slide-3
SLIDE 3

Center for Research in

Intelligent Storage

  • Drive Managed

– Black box/drop-in solution: the drive handles all out-of-order write

  • perations.
  • Host Managed

– White box/application modification needed: the drive reports zone layout information; out-of-order writes will be rejected.

  • Host Aware

– Grey box: the drive reports zone layout information; out-of-order writes will still be handled internally. – Applications can use HA-SMR drive as is, and also have the opportunity for zone-layout aware optimizations.

T10 SMR Drive Models

slide-4
SLIDE 4

Center for Research in

Intelligent Storage

Example: Seagate HA SMR Sample Drives

  • Model: ST8000AS0022-1WL, prototype firmware revision ZN03.
  • Small Conventional Zone 64GB/8TB ~= 1%
  • Most disk space is sequential write preferred zone
  • Media Cache hidden from the user

write pointer (wp) wp wp

Conventional Zones Write Pointer Zones Outer track Inner track

slide-5
SLIDE 5

Center for Research in

Intelligent Storage

  • Motivation: To meet the challenge of using SMR drives in large-

scale storage systems.

Motivation & Goals

SMR Layout Awareness

  • Exploring which level to be SMR zone layout aware

to support different applications (FS, DB, etc.)

SMR Drive Aggregation

  • How to reduce the design complexity for multiple

SMR drive applications.

slide-6
SLIDE 6

Center for Research in

Intelligent Storage

  • Avoid non-seq: Convert all workload to sequential

– Always perform sequential write to SMR drive. – Achieving near-HDD performance.

  • Accept non-seq: Know performance characteristic

and tweak workload accordingly.

– when to avoid non-sequential write; when to let it go. – reduce management overhead.

  • Which layer to be SMR-Aware?
  • Fully Aware or Partial Aware?

– Hide/expose SMR information further up

SMR Layout Awareness

FS APP S S S S Multiple SMR Drives RAID LVM

slide-7
SLIDE 7

Center for Research in

Intelligent Storage

  • Abstract multiple physical

SMR drives into logical

  • ne(s).
  • Preserve the I/O

characteristic

– How much we can preserve? – Parametrized SMR drives

  • Reduce design complexity,

again.

SMR Drive Aggregation

FS APP One Logical SMR Driv RAID LVM FS APP S S S S Multiple SMR Drives RAID LVM Aggregation

slide-8
SLIDE 8

Center for Research in

Intelligent Storage

  • Understanding single SMR drive I/O performance for different

workload

– Defines “what to be aware of”. – Inspires “how to profile a logical aggregated SMR drive”.

  • Indirection Buffer Design

– One HA-SMR Awareness design. – Preliminary result shows its effectiveness.

  • First version of prototype for aggregating SMR Drives (libvir)

– User level implementation based on libzbc

Progress So Far

slide-9
SLIDE 9

Center for Research in

Intelligent Storage

  • Test goal focuses on unique features of HA-SMR:

– Open zone issue – Non-sequential zone issue – Media cache cleaning efficiency

  • Test Setup

– Replay micro-benchmarking traces to HA-SMR drives

Host Aware SMR (HA-SMR) Drive Testing

SMR

Micro-benchmarking Trace

fio libzbc

slide-10
SLIDE 10

Center for Research in

Intelligent Storage

CMR: Conventional Magnetic Recording SMR: Shingled Magnetic Recording

SMR: (+) more data density; (-) update overhead CMR: (-) less data density; (+) no update overhead.

Update: all overlapping tracks needs to be rewritten, causing update overhead.

How about a combination of the two?

slide-11
SLIDE 11

Center for Research in

Intelligent Storage

Emerging Hybrid SMR Drives

11 Disk Platters

CMR SMR

Hybrid SMR Drive

Online Conversion

  • Hybrid SMR (H-SMR): mix of CMR and

SMR; can be converted on line by H-SMR API.

  • Benefit: utilize both IOPS and Capacity.

Flexible and reconfigurable.

Objective: How do we efficiently manage the data and space for such Hybrid SMR drives?

slide-12
SLIDE 12

Center for Research in

Intelligent Storage

Outline

  • Introduction
  • Design Goals
  • Background and Challenges
  • Design and Evaluation
  • Summary
slide-13
SLIDE 13

Center for Research in

Intelligent Storage

Design Goal

  • Handle growing utilization.
  • Reduce SMR update overhead.
  • Adapt to dynamic workload.

13

slide-14
SLIDE 14

Center for Research in

Intelligent Storage

Disk Physical Space CMR (allocated)

allocation direction

Disk Physical Space CMR . SMR

CMR/SMR boundary

CMR (unallocated)

Phase I: CMR-only Phase II: CMR + SMR

Two-Phase Allocation

  • uter

tracks inner tracks

  • uter

tracks inner tracks

Intuition: use CMR first! then convert CMR to SMR when capacity is not enough.

Disk Physical Space CMR

Initial State: Empty Disk

  • uter

tracks inner tracks

slide-15
SLIDE 15

Center for Research in

Intelligent Storage

15

The devil is in the detail! Buckle up. Challenge ahead!

slide-16
SLIDE 16

Center for Research in

Intelligent Storage

16

  • uter track

inner track CMR Partition SMR Partition CMR Space SMR Space to be converted

Background: Format Conversion

Logical Space Physical Space

slide-17
SLIDE 17

Center for Research in

Intelligent Storage

17

  • uter track

inner track CMR Partition SMR Partition CMR Space SMR Space converted

Background: Format Conversion

Logical Space Physical Space

slide-18
SLIDE 18

Center for Research in

Intelligent Storage

18

  • uter track

inner track CMR Partition SMR Partition CMR Space SMR Space converted

Conversion After

  • uter track

inner track CMR Partition SMR Partition CMR Space SMR Space

Conversion

to be converted

Before

  • ffline from CMR
  • nline in SMR

Background: Format Conversion

live data migration new capacity created

Logical Space Physical Space Logical Space Physical Space

slide-19
SLIDE 19

Center for Research in

Intelligent Storage

19

  • uter track

inner track CMR Partition SMR Partition CMR Space SMR Space converted

Conversion After

  • uter track

inner track CMR Partition SMR Partition CMR Space SMR Space

Conversion

to be converted

Before

Format Conversion

live data migration new capacity created Challenge:

  • Address space / mapping
  • SMR update overhead
  • Live data migration cost

Disk Physical Space CMR (allocated)

allocation direction

Disk Physical Space CMR . SMR

CMR/SMR boundary

CMR (unallocated)

  • uter

tracks inner tracks

  • uter

tracks inner tracks

slide-20
SLIDE 20

Center for Research in

Intelligent Storage

Challenges and Solutions

Challenge:

  • Address space / mapping
  • SMR update overhead
  • Live data migration cost

Solution:

  • Elastic Address Space with Zone-level Mapping
  • H-Buffer and Zone-Swap to reduce SMR update
  • verhead
  • Quantized Migration to mitigate live data

migration cost in the format conversion

slide-21
SLIDE 21

Center for Research in

Intelligent Storage

Challenges and Solutions

Challenge:

  • Address space / mapping
  • SMR update overhead
  • Live data migration cost

Solution:

  • Elastic Address Space with Zone-level Mapping
  • H-Buffer and Zone-Swap to reduce SMR update
  • verhead
  • Quantized Migration to mitigate live data

migration cost in the format conversion

slide-22
SLIDE 22

Center for Research in

Intelligent Storage

DESIGN I: ADDRESS SPACE / MAPPING

Elastic Address Space with Zone-level Mapping

slide-23
SLIDE 23

Center for Research in

Intelligent Storage

23

  • uter track

inner track CMR Partition SMR Partition CMR Space SMR Space to be converted Logical Space Physical Space User Space

...

extendable size

zone mapping user zone logical zone

Elastic Address Space read/write/extend size

slide-24
SLIDE 24

Center for Research in

Intelligent Storage

24

  • uter track

inner track CMR Partition SMR Partition CMR Space SMR Space converted Logical Space Physical Space User Space

...

extendable size new space migrated live data mapping updated Elastic Address Space read/write/extend size

slide-25
SLIDE 25

Center for Research in

Intelligent Storage

DESIGN II: REDUCING SMR UPDATE OVERHEAD

H-Buffer and Zone-Swap

slide-26
SLIDE 26

Center for Research in

Intelligent Storage

H-Buffer: Overview

  • uter track

inner track CMR Space SMR Space

Logical Space Physical Space User Space extendable size

CMR Partition SMR Partition

Elastic Address Space

...

Updates Map

slide-27
SLIDE 27

Center for Research in

Intelligent Storage

H-Buffer: Overview

  • uter track

inner track CMR Space SMR Space

Logical Space Physical Space User Space

...

extendable size H-Buffer

CMR Partition SMR Partition H-Buffer: Host-controlled Buffer Basic Idea: using some reserved CMR space to buffer SMR updates and migrate to SMR zones later.

Updates Elastic Address Space Redirect Migrate Map

slide-28
SLIDE 28

Center for Research in

Intelligent Storage

H-Buffer Management: Alternatives

  • Block-based (e.g. LRU)
  • Log-based

– In-place FIFO – Loop-back Log (with hot/cold classification)

28

Problem: Random I/O in redirecting/cleaning

SMR Zones H-Buffer evict loop back

data blocks

Loop-back Log HEAD grow TAIL clean Idea: Re-queue the hot data blocks to the log head without evicting to SMR zones.

slide-29
SLIDE 29

Center for Research in

Intelligent Storage

Zone-Swap: Overview

CMR Zones SMR Zones H-Buffer Basic Idea: swap hot zones (heavily updated ones) from SMR to CMR to reduce SMR update overhead.

cold zone hot zone

Zone-Swap

Co-design with H-Buffer:

  • Swapping happens when H-Buffer evicts.
  • H-Buffer eviction choice also depends on Zone-Swap decision.
slide-30
SLIDE 30

Center for Research in

Intelligent Storage

H-Buffer and Zone-Swap Evaluation

2 4 6 8 10 12 prn_1 proj_0 I/O Latency (ms) inplace fifo h-buffer (loop-back) zone-alloy (h-buffer + zone-swap)

More result: in the poster session.

slide-31
SLIDE 31

Center for Research in

Intelligent Storage

Background

Summary: ZoneAlloy

  • - Data management for Hybrid-SMR

Disk Platters

CMR SMR

Hybrid SMR Drive

Online Conversion

Solutions/Contributions Problem

Data and Space Management in Hybrid SMR drives

Elastic Address Space/ Zone- level Mapping Quantized Migration H-Buffer Zone-Swapping Two-Phase Allocation

slide-32
SLIDE 32

Center for Research in

Intelligent Storage

I/O stack change & API (ask for feedback)

32

Question: Which layer(s) of the I/O stack should do the heavy-lifting?

small/mid-size business large-size business storage vendor storage service black box white box buy rent build + open source Applications Block, FS, Object Storage OLTP, OLAP ML/AI Batching/Stream Traditional API Applications Richer API Storage OLTP, OLAP ML/AI Batching/Stream

H-SMR Aware

  • Appl. Aware

Hint Convert Migrate H-SMR

slide-33
SLIDE 33

Center for Research in

Intelligent Storage

Thank You! Questions?

Fenggang Wu wuxx0835@umn.edu

small/mid-size business large-size business storage vendor storage service black box white box buy rent build + open source