1

Ziggurat: A Tiered File System for Non-Volatile Main Memories and Disks

Shengan Zheng†, Morteza Hoseinzadeh§, Steven Swanson§

† Shanghai Jiao Tong University§ University of California, San Diego

2

Background

• Non-volatile main memory (NVMM)– Byte-addressability– Persistence– Direct access (DAX)

• NVMM file systems– PMFS, SCMFS, NOVA– EXT4-DAX, XFS-DAX– Capacity? DRAM + Flash NVDIMM

3D-XPoint NVDIMM

3

Motivation

Bandwidth

$/GB

1GB/s

100MB/s

10GB/s

0.01 0.1 1 10

Hard Disk Drive

SATA SSDNVMe SSD

Optane SSDNVMM

DRAM

4

Motivation

Bandwidth

$/GB

1GB/s

100MB/s

10GB/s

0.01 0.1 1 10

Hard Disk Drive

SATA SSDNVMe SSD

Optane SSDNVMM

DRAM

5

HDD

Tiered Storage System

SSD

• SSD for speed• HDD for capacity

6

HDD

Tiered Storage System

• NVMM for speed• Disks for capacity

SSD

NVMM

7

Ziggurat Overview

• Intelligent data placement policy– Send writes to the most suitable tier– High NVMM space utilization

• Accurate predictors– Predict the synchronicity of each file (synchronicity predictor)– Predict the size of future writes to each file (write size predictor)

• Efficient migration mechanism– Only migrate cold data in cold files– Migrate file data in groups

8

Outline

• Motivation• Data placement policy• Migration mechanism• Evaluation• Conclusion

9

DiskNVMM

NVMMNVMM

Synchronously-updated Asynchronously-updated

Large writesSmall writes

Data Placement Policy

• Although NVMM is the fastest tier in Ziggurat, file writes should not always go to NVMM.

Synchronicity predictor

Write size

predictor

Data Placement

10

Synchronicity Predictor

• Predict whether the future accesses are likely to be synchronouswrite(0,2);fsync();write(2,2);fsync();write(4,2);fsync();

write(0,2);write(2,2);write(4,2);fsync();

0 1 2 3

Data blocks written: 0 / 4

Synchronous4 5 6 7

0 1 2 3

Data blocks written: 0 / 4

Asynchronous4 5 6 7

File log

File data

Write entry offset, length

File log

File data Synchronous

11

Synchronicity Predictor

• Predict whether the future accesses are likely to be synchronouswrite(0,2);fsync();write(2,2);fsync();write(4,2);fsync();

write(0,2);write(2,2);write(4,2);fsync();

0 1 2 3

Data blocks written: 2 / 4

Synchronous4 5 6 7

0 1 2 3

Data blocks written: 0 / 4

Asynchronous4 5 6 7

File log

File data

Write entry offset, length

File log

File data Synchronous

0,2

12

Synchronicity Predictor

• Predict whether the future accesses are likely to be synchronouswrite(0,2);fsync();write(2,2);fsync();write(4,2);fsync();

write(0,2);write(2,2);write(4,2);fsync();

0 1 2 3

Data blocks written: 0 / 4

Synchronous4 5 6 7

0 1 2 3

Data blocks written: 0 / 4

Asynchronous4 5 6 7

File log

File data

Write entry offset, length

File log

File data Synchronous

0,2

13

Synchronicity Predictor

• Predict whether the future accesses are likely to be synchronouswrite(0,2);fsync();write(2,2);fsync();write(4,2);fsync();

write(0,2);write(2,2);write(4,2);fsync();

0 1 2 3

Data blocks written: 2 / 4

Synchronous4 5 6 7

0 1 2 3

Data blocks written: 0 / 4

Asynchronous4 5 6 7

File log

File data

Write entry offset, length

File log

File data Synchronous

0,2 2,2

14

Synchronicity Predictor

• Predict whether the future accesses are likely to be synchronouswrite(0,2);fsync();write(2,2);fsync();write(4,2);fsync();

write(0,2);write(2,2);write(4,2);fsync();

0 1 2 3

Data blocks written: 0 / 4

Synchronous4 5 6 7

0 1 2 3

Data blocks written: 0 / 4

Asynchronous4 5 6 7

File log

File data

Write entry offset, length

File log

File data Synchronous

0,2 2,2

15

Synchronicity Predictor

• Predict whether the future accesses are likely to be synchronouswrite(0,2);fsync();write(2,2);fsync();write(4,2);fsync();

write(0,2);write(2,2);write(4,2);fsync();

0 1 2 3

Data blocks written: 2 / 4

Synchronous4 5 6 7

0 1 2 3

Data blocks written: 0 / 4

Asynchronous4 5 6 7

File log

File data

Write entry offset, length

File log

File data Synchronous

0,2 2,2 4,2

16

Synchronicity Predictor

• Predict whether the future accesses are likely to be synchronouswrite(0,2);fsync();write(2,2);fsync();write(4,2);fsync();

write(0,2);write(2,2);write(4,2);fsync();

0 1 2 3

Data blocks written: 0 / 4

Synchronous4 5 6 7

0 1 2 3

Data blocks written: 0 / 4

Asynchronous4 5 6 7

File log

File data

Write entry offset, length

File log

File data Synchronous

0,2 2,2 4,2

17

Synchronicity Predictor

• Predict whether the future accesses are likely to be synchronouswrite(0,2);fsync();write(2,2);fsync();write(4,2);fsync();

write(0,2);write(2,2);write(4,2);fsync();

0 1 2 3

Data blocks written: 0 / 4

Synchronous4 5 6 7

0 1 2 3

Data blocks written: 2 / 4

Asynchronous4 5 6 7

File log

File data

Write entry offset, length

File log

File data Synchronous

0,2 2,2 4,2

0,2

18

Synchronicity Predictor

• Predict whether the future accesses are likely to be synchronouswrite(0,2);fsync();write(2,2);fsync();write(4,2);fsync();

write(0,2);write(2,2);write(4,2);fsync();

0 1 2 3

Data blocks written: 0 / 4

Synchronous4 5 6 7

0 1 2 3

Data blocks written: 4 / 4

Asynchronous4 5 6 7

File log

File data

Write entry offset, length

File log

File data Synchronous

0,2 2,2 4,2

0,2 2,2

19

Synchronicity Predictor

• Predict whether the future accesses are likely to be synchronouswrite(0,2);fsync();write(2,2);fsync();write(4,2);fsync();

write(0,2);write(2,2);write(4,2);fsync();

0 1 2 3

Data blocks written: 0 / 4

Synchronous4 5 6 7

0 1 2 3

Data blocks written: 6 / 4

Asynchronous4 5 6 7

File log

File data

Write entry offset, length

File log

File data Synchronous

0,2 2,2 4,2

0,2 2,2 4,2

20

Synchronicity Predictor

• Predict whether the future accesses are likely to be synchronouswrite(0,2);fsync();write(2,2);fsync();write(4,2);fsync();

write(0,2);write(2,2);write(4,2);fsync();

0 1 2 3

Data blocks written: 0 / 4

Synchronous4 5 6 7

0 1 2 3

Data blocks written: 0 / 4

Asynchronous4 5 6 7

File log

File data

Write entry offset, length

File log

File data

0,2 2,2 4,2

0,2 2,2 4,2

21

Write Size Predictor

• Predict whether the incoming writes are both large and stable

0 1 2 3 4 5 6 7

File log 0,4,3

File data

4,4,1 5,1,0 6,1,0

write(0,4);write(6,1);write(4,4);

offset, length, counterWrite entry

22

Write Size Predictor

• Predict whether the incoming writes are both large and stable

0 1 2 3 4 5 6 7

File log 0,4,3

File data

4,4,1 5,1,0 6,1,0

write(0,4);write(6,1);write(4,4);

offset, length, counterWrite entry

0,4,?

23

Write Size Predictor

• Predict whether the incoming writes are both large and stable

0 1 2 3 4 5 6 7

File log 0,4,3

File data

4,4,1 5,1,0 6,1,0

write(0,4);write(6,1);write(4,4);

offset, length, counterWrite entry

0,4,?

Predecessor found

Large

Stable

Length ≥ 4

24

Write Size Predictor

• Predict whether the incoming writes are both large and stable

0 1 2 3 4 5 6 7

File log 0,4,3

File data

4,4,1 5,1,0 6,1,0

write(0,4);write(6,1);write(4,4);

offset, length, counterWrite entry

0,4,?

Predecessor found

Large

Stable

Length ≥ 4

0,4,4

25

Write Size Predictor

• Predict whether the incoming writes are both large and stable

0 1 2 3 4 5 6 7

File log 0,4,3

File data

4,4,1 5,1,0 6,1,0

write(0,4);write(6,1);write(4,4);

offset, length, counterWrite entry

0,4,?

Predecessor found

Large

Stable

Length ≥ 4

0,4,4

6,1,?

26

Write Size Predictor

• Predict whether the incoming writes are both large and stable

0 1 2 3 4 5 6 7

File log 0,4,3

File data

4,4,1 5,1,0 6,1,0

write(0,4);write(6,1);write(4,4);

offset, length, counterWrite entry

0,4,?

Predecessor found

Large

Stable

Length ≥ 4

0,4,4

6,1,?

Predecessor found

Small

Stable

Length < 4

27

Write Size Predictor

• Predict whether the incoming writes are both large and stable

0 1 2 3 4 5 6 7

File log 0,4,3

File data

4,4,1 5,1,0 6,1,0

write(0,4);write(6,1);write(4,4);

offset, length, counterWrite entry

0,4,?

Predecessor found

Large

Stable

Length ≥ 4

0,4,4

6,1,?

Predecessor found

Small

Stable

Length < 4

6,1,0

28

Write Size Predictor

• Predict whether the incoming writes are both large and stable

0 1 2 3 4 5 6 7

File log 0,4,3

File data

4,4,1 5,1,0 6,1,0

write(0,4);write(6,1);write(4,4);

offset, length, counterWrite entry

0,4,?

Predecessor found

Large

Stable

Length ≥ 4

0,4,4

6,1,?

Predecessor found

Small

Stable

Length < 4

6,1,0

4,4,?

29

Write Size Predictor

• Predict whether the incoming writes are both large and stable

0 1 2 3 4 5 6 7

File log 0,4,3

File data

4,4,1 5,1,0 6,1,0

write(0,4);write(6,1);write(4,4);

offset, length, counterWrite entry

0,4,?

Predecessor found

Large

Stable

Length ≥ 4

0,4,4

6,1,?

Predecessor found

Small

Stable

Length < 4

6,1,0

4,4,?

Predecessor not found

Large

Unstable

Length ≥ 4

30

Write Size Predictor

• Predict whether the incoming writes are both large and stable

0 1 2 3 4 5 6 7

File log 0,4,3

File data

4,4,1 5,1,0 6,1,0

write(0,4);write(6,1);write(4,4);

offset, length, counterWrite entry

0,4,?

Predecessor found

Large

Stable

Length ≥ 4

0,4,4

6,1,?

Predecessor found

Small

Stable

Length < 4

6,1,0

4,4,?

Predecessor not found

Large

Unstable

Length ≥ 4

4,4,0

31

Cold Data Identification

• Average modification time (amtime)– Updated differentially

0 1 2 3 4 5 6 7

File log 0,2,2

File data

2,2,4 4,2,6

0 1 2 3 4 5 6 7

File log 0,2,2

File data

2,2,4 4,2,6 6,2,8

amtime = amtime = 2 ∗ 2 + 4 ∗ 2 + 6 ∗ 22 + 2 + 2 = 44 ∗ 6 + 8 ∗ 2

6 + 2 = 5

offset, length, mtimeWrite entry

32

Cold Data Identification

Cold File

File

File

File

CPU 0

Hot

File

File

File

File

CPU 1

• Among files– Cold lists sorted by amtime

amtime

0 1 2 3 4 5 6 7

File log 0,2,2

File data

2,2,4 4,2,6 6,2,8

amtime = 5

• Within each file– Cold data blocks relative to amtime

Cold Cold Hot Hot

offset, length, mtimeWrite entry

33

Migration Mechanism

Chm

od

Writ

e0-

8K

Head TailInode

Inodelog

NVMM

Disk

... Write 0-4K

File Page 1

File Page 2

File Page 3

File Page 1

File Page 4

Writ

e8-

16K

Pages

Page state Stale Live

Entry type Inode update Old write entry New write entry

• Basic migration– Migrate the coldest data to disk– Consistency is ensured

• Reverse migration– Migrate from disk to NVMM– Handle mmap– Read-dominated workloads

34

Migration Mechanism

Chm

od

Writ

e0-

8K

Head TailInode

Inodelog

NVMM

Disk

... Write 0-4K

File Page 1

File Page 2

File Page 3

File Page 1

File Page 4

Writ

e8-

16K

File Page 3’

File Page 4’

Step 1

Pages

Page state Stale Live

Entry type Inode update Old write entry New write entry

• Basic migration– Migrate the coldest data to disk– Consistency is ensured

• Reverse migration– Migrate from disk to NVMM– Handle mmap– Read-dominated workloads

35

Migration Mechanism

Chm

od

Writ

e0-

8K

Head TailInode

Inodelog

NVMM

Disk

... Write 0-4K

File Page 1

File Page 2

File Page 3

File Page 1

File Page 4

Writ

e8-

16K

File Page 3’

File Page 4’

Step 1

Step 2

Pages

Page state Stale Live

Entry type Inode update Old write entry New write entry

• Basic migration– Migrate the coldest data to disk– Consistency is ensured

• Reverse migration– Migrate from disk to NVMM– Handle mmap– Read-dominated workloads

36

Migration Mechanism

Chm

od

Writ

e0-

8K

Head TailInode

Inodelog

NVMM

Disk

... Write 0-4K Write 8-16K

File Page 1

File Page 2

File Page 3

File Page 1

File Page 4

Writ

e8-

16K

File Page 3’

File Page 4’

Step 1

Step 2

Step 3

Pages

Page state Stale Live

Entry type Inode update Old write entry New write entry

• Basic migration– Migrate the coldest data to disk– Consistency is ensured

• Reverse migration– Migrate from disk to NVMM– Handle mmap– Read-dominated workloads

37

Migration Mechanism

Chm

od

Writ

e0-

8K

Head TailInode

Inodelog

NVMM

Disk

... Write 0-4K Write 8-16K

File Page 1

File Page 2

File Page 3

File Page 1

File Page 4

Writ

e8-

16K

File Page 3’

File Page 4’

Step 1

Step 2

Step 3

Step 4

Pages

Page state Stale Live

Entry type Inode update Old write entry New write entry

• Basic migration– Migrate the coldest data to disk– Consistency is ensured

• Reverse migration– Migrate from disk to NVMM– Handle mmap– Read-dominated workloads

38

Migration Mechanism

Chm

od

Writ

e0-

8K

Head TailInode

Inodelog

NVMM

Disk

... Write 0-4K Write 8-16K

File Page 1

File Page 2

File Page 3

File Page 1

File Page 4

Writ

e8-

16K

File Page 3’

File Page 4’

Step 1

Step 2

Step 3

Step 4

PagesStep 5

Page state Stale Live

Entry type Inode update Old write entry New write entry

• Basic migration– Migrate the coldest data to disk– Consistency is ensured

• Reverse migration– Migrate from disk to NVMM– Handle mmap– Read-dominated workloads

39

Migration Mechanism

Chm

od

Writ

e0-

8K

Head TailInode

... Write 4-8K

File Page 1

File Page 2

File Page 3

File Page 2

File Page 4

Writ

e8-

16K

Writ

e12

-16K

File Page 4

...Inodelog

NVMM

Disk

Pages

Page state Stale Live

Entry type Inode update Old write entry New write entry

• Group migration– Coalesce adjacent write entries– Utilize the high sequential

bandwidth of disks• Benefits

– Improve migration efficiency– Accelerate future reads– Reduce log size– Moderate disk fragmentation

40

Migration Mechanism

Step 1

File Page 1’

File Page 2’

File Page 3’

File Page 4’

Chm

od

Writ

e0-

8K

Head TailInode

... Write 4-8K

File Page 1

File Page 2

File Page 3

File Page 2

File Page 4

Writ

e8-

16K

Writ

e12

-16K

File Page 4

...Inodelog

NVMM

Disk

Pages

Page state Stale Live

Entry type Inode update Old write entry New write entry

• Group migration– Coalesce adjacent write entries– Utilize the high sequential

bandwidth of disks• Benefits

– Improve migration efficiency– Accelerate future reads– Reduce log size– Moderate disk fragmentation

41

Migration Mechanism

Step 1

File Page 1’

File Page 2’

File Page 3’

File Page 4’

Chm

od

Writ

e0-

8K

Head TailInode

... Write 4-8K

File Page 1

File Page 2

File Page 3

File Page 2

File Page 4

Writ

e8-

16K

Writ

e12

-16K

File Page 4

...Inodelog

NVMM

Disk

Pages

Page state Stale Live

Entry type Inode update Old write entry New write entry

Step 2

• Group migration– Coalesce adjacent write entries– Utilize the high sequential

bandwidth of disks• Benefits

– Improve migration efficiency– Accelerate future reads– Reduce log size– Moderate disk fragmentation

42

Migration Mechanism

Step 1

File Page 1’

File Page 2’

File Page 3’

File Page 4’

Chm

od

Writ

e0-

8K

Head TailInode

... Write 4-8K

File Page 1

File Page 2

File Page 3

File Page 2

File Page 4

Writ

e8-

16K

Writ

e12

-16K

Writ

e0-

16K

File Page 4

Step 3

...Inodelog

NVMM

Disk

Pages

Page state Stale Live

Entry type Inode update Old write entry New write entry

Step 2

• Group migration– Coalesce adjacent write entries– Utilize the high sequential

bandwidth of disks• Benefits

– Improve migration efficiency– Accelerate future reads– Reduce log size– Moderate disk fragmentation

43

Migration Mechanism

Step 1

File Page 1’

File Page 2’

File Page 3’

File Page 4’

Chm

od

Writ

e0-

8K

Head TailInode

... Write 4-8K

File Page 1

File Page 2

File Page 3

File Page 2

File Page 4

Writ

e8-

16K

Step 4

Writ

e12

-16K

Writ

e0-

16K

File Page 4

Step 3

...Inodelog

NVMM

Disk

Pages

Page state Stale Live

Entry type Inode update Old write entry New write entry

Step 2

• Group migration– Coalesce adjacent write entries– Utilize the high sequential

bandwidth of disks• Benefits

– Improve migration efficiency– Accelerate future reads– Reduce log size– Moderate disk fragmentation

44

Migration Mechanism

Step 1

Step 5

File Page 1’

File Page 2’

File Page 3’

File Page 4’

Chm

od

Writ

e0-

8K

Head TailInode

... Write 4-8K

File Page 1

File Page 2

File Page 3

File Page 2

File Page 4

Writ

e8-

16K

Step 4

Writ

e12

-16K

Writ

e0-

16K

File Page 4

Step 3

...Inodelog

NVMM

Disk

Pages

Page state Stale Live

Entry type Inode update Old write entry New write entry

Step 2

• Group migration– Coalesce adjacent write entries– Utilize the high sequential

bandwidth of disks• Benefits

– Improve migration efficiency– Accelerate future reads– Reduce log size– Moderate disk fragmentation

45

File Operations

NVMM

DRAM

write(0,8);write(8,8);fsync();append(16,1);append(17,1);...append(23,1);......migrate();......read(16,8);mmap(0,8);

DISK

0 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23

46

File Operations

NVMM

DRAM

write(0,8);write(8,8);fsync();append(16,1);append(17,1);...append(23,1);......migrate();......read(16,8);mmap(0,8);

DISK

0 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23

47

File Operations

NVMM

DRAM

write(0,8);write(8,8);fsync();append(16,1);append(17,1);...append(23,1);......migrate();......read(16,8);mmap(0,8);

DISK

0 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23

48

File Operations

NVMM

DRAM

write(0,8);write(8,8);fsync();append(16,1);append(17,1);...append(23,1);......migrate();......read(16,8);mmap(0,8);

DISK

0 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23

49

File Operations

NVMM

DRAM

write(0,8);write(8,8);fsync();append(16,1);append(17,1);...append(23,1);......migrate();......read(16,8);mmap(0,8);

DISK

0 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23

50

File Operations

NVMM

DRAM

write(0,8);write(8,8);fsync();append(16,1);append(17,1);...append(23,1);......migrate();......read(16,8);mmap(0,8);

DISK

0 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23

51

File Operations

NVMM

DRAM

write(0,8);write(8,8);fsync();append(16,1);append(17,1);...append(23,1);......migrate();......read(16,8);mmap(0,8);

DISK

0 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23

52

File Operations

NVMM

DRAM

write(0,8);write(8,8);fsync();append(16,1);append(17,1);...append(23,1);......migrate();......read(16,8);mmap(0,8);

DISK

0 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23

53

File Operations

NVMM

DRAM

write(0,8);write(8,8);fsync();append(16,1);append(17,1);...append(23,1);......migrate();......read(16,8);mmap(0,8);

DISK

0 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23

54

File Operations

NVMM

DRAM

write(0,8);write(8,8);fsync();append(16,1);append(17,1);...append(23,1);......migrate();......read(16,8);mmap(0,8);

DISK

0 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23

55

File Operations

NVMM

DRAM

write(0,8);write(8,8);fsync();append(16,1);append(17,1);...append(23,1);......migrate();......read(16,8);mmap(0,8);

DISK

0 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23

56

File Operations

NVMM

DRAM

write(0,8);write(8,8);fsync();append(16,1);append(17,1);...append(23,1);......migrate();......read(16,8);mmap(0,8);

DISK

0 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23

57

File Operations

NVMM

DRAM

write(0,8);write(8,8);fsync();append(16,1);append(17,1);...append(23,1);......migrate();......read(16,8);mmap(0,8);

DISK

0 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23

58

Evaluation

• Experimental setup– Dual-socket Intel Xeon E5 server– 256 GB DRAM– 187 GB Intel DC P4800X Optane SSD– 400 GB Intel DC P3600 NVMe SSD– Ubuntu 16.04 LTS, Linux kernel 4.13.0

• Emulated NVMM– NUMA effect on DRAM– NUMA node 1: NVMM emulation– NUMA node 0: Processors and memory for applications

• Disk-based file systems– EXT4-DJ (data journaling)– XFS-ML (metadata logging)

• Ziggurat– Ziggurat-X (Optane/NVMe

SSD + X GB of NVMM)• NVMM-based file systems

– NOVA, EXT4-DAX, XFS-DAX

59

Filebench

• With large amount of NVMM, Ziggurat’s performance nearly matches that of NVMM-only file systems.

Fileserver Varmail

Within 3%

2.1X

2.6X

60

Rocksdb

• Ziggurat shows good performance for inserting file data with write-ahead logging (WAL).

Random insert Sequential insert

9.9X

42.4X5.7X

13.7X

61

SQLite

• Ziggurat maintains near-NVMM performance because the hot journal files are frequently updated in NVMM.

PERSIST WAL

1.6X

4.8X

1.4X

5.3X

62

Conclusion

• Ziggurat fully utilizes the strengths of NVMM and disks to offer high file performance for a wide range of access patterns.

• [Prediction] Ziggurat steers the incoming writes to the most suitable tier based on the prediction results.

• [Migration] Ziggurat coalesces adjacent data blocks and migrates them in large chunks to disks.

• [Evaluation] Ziggurat achieves up to 38.9X and 46.5X throughput improvement compared with EXT4 and XFS running on an SSD alone, respectively.

63

Thank you

Ziggurat: A Tiered File System for �Non-Volatile Main Memories and DisksBackgroundMotivationMotivationTiered Storage SystemTiered Storage SystemZiggurat OverviewOutlineData Placement PolicySynchronicity PredictorSynchronicity PredictorSynchronicity PredictorSynchronicity PredictorSynchronicity PredictorSynchronicity PredictorSynchronicity PredictorSynchronicity PredictorSynchronicity PredictorSynchronicity PredictorSynchronicity PredictorWrite Size PredictorWrite Size PredictorWrite Size PredictorWrite Size PredictorWrite Size PredictorWrite Size PredictorWrite Size PredictorWrite Size PredictorWrite Size PredictorWrite Size PredictorCold Data IdentificationCold Data IdentificationMigration MechanismMigration MechanismMigration MechanismMigration MechanismMigration MechanismMigration MechanismMigration MechanismMigration MechanismMigration MechanismMigration MechanismMigration MechanismMigration MechanismFile OperationsFile OperationsFile OperationsFile OperationsFile OperationsFile OperationsFile OperationsFile OperationsFile OperationsFile OperationsFile OperationsFile OperationsFile OperationsEvaluationFilebenchRocksdbSQLiteConclusion幻灯片编号 63