MiniCrypt: Reconciling Compression and Encryption in Big ...wzheng/eurosys_2017_talk.pdf ·...

MiniCrypt: Reconciling Compression and Encryption

in Big Data Stores

Wenting Zheng, Frank Li, Raluca Ada Popa, Ion Stoica, Rachit Agarwal

client cloud provider

Sensitive data

A lot of sensitive data is collected

Existing systems

Existing systems

Systems Compression Encryption

Existing systems


Cassandra [LM10], MongoDB, Succinct [AKS15]

Existing systems


Cassandra [LM10], MongoDB, Succinct [AKS15]

CryptDB [PRZB11], Plutus [KRSWF03], Sirius [GSMB03],

etc.

How to combine compression and encryption?

MiniCrypt

Provides both compression and encryption for key-value stores

MiniCrypt

Provides both compression and encryption for key-value stores

MiniCrypt

Works on top of existing key-value stores without modification

System setup

System setup

Client 1

client

Client 2

Client 3

System setup

Client 1


Client 2

Client 3

Server 2

Server 1

System setup

Client 1

MiniCrypt


Client 2

MiniCrypt

Client 3

MiniCrypt

Server 2

Server 1

Server 2

Server 1

Threat model

cloud provider

Client 1

MiniCrypt

Client 2

MiniCrypt

Client 3

MiniCrypt

client

Encryption before compression?


Ticking away the moments that make up a dull day Fritter and waste the hours in an off-hand way Kicking around on a piece of ground in your home town …



01010101101010101111100001010100000011001010101001010101010010101011010101011111000010101000000110010101010010101010100101010110 0000110010101010 …



01010101101010101111100001010100000011001010101001010101010010101011010101011111000010101000000110010101010010101010100101010110 0000110010101010 …

01010101101010101111100001010100000011001010101001010101010010101011010101011111000010101000000110010101010010101010100101010110 0000110010101010 …



01010101101010101111100001010100000011001010101001010101010010101011010101011111000010101000000110010101010010101010100101010110 0000110010101010 …

ZIP01010101101010101111100001010100000011001010101001010101010010101011010101011111000010101000000110010101010010101010100101010110 0000110010101010 …



01010101101010101111100001010100000011001010101001010101010010101011010101011111000010101000000110010101010010101010100101010110 0000110010101010 …

ZIP

Pseudorandom data is not compressible

01010101101010101111100001010100000011001010101001010101010010101011010101011111000010101000000110010101010010101010100101010110 0000110010101010 …

Compression before encryption?



ZIP



010101011010101011111000010101000000110010101010010101010100101010110101010111110000101010000001 …

ZIP



010101011010101011111000010101000000110010101010010101010100101010110101010111110000101010000001 …

ZIP

010101011010101011111000010101000000110010101010010101010100101010110101010111110000101010000001 …



010101011010101011111000010101000000110010101010010101010100101010110101010111110000101010000001 …

ZIP

Compression and encryption

010101011010101011111000010101000000110010101010010101010100101010110101010111110000101010000001 …

Strawman #1: compress & encrypt entire database


Client 1

MiniCrypt

Client 2

MiniCrypt

Client 3

MiniCrypt


Client 1

MiniCrypt

Client 2

MiniCrypt

Client 3

MiniCrypt

12809 Amanda D. Edwards 40 Diabetes29489 Robert R. McGowan 56 Diabetes13744 Kimberly R. Seay 51 Cancer18740 Dennis G. Bates 32 Diabetes98329 Ronald S. Ogden 53 Cancer32591 Donna R. Bridges 26 Diabetes


Client 1

MiniCrypt

Client 2

MiniCrypt

Client 3

MiniCrypt


ZIP


Client 1

MiniCrypt

Client 2

MiniCrypt

Client 3

MiniCrypt


ZIP

Problem: high network bandwidth

Strawman #2: compress & encrypt single row


Client 1

MiniCrypt

Client 2

MiniCrypt

Client 3

MiniCrypt


29489 Robert R. McGowan 56 Diabetes

13744 Kimberly R. Seay 51 Cancer

18740 Dennis G. Bates 32 Diabetes

98329 Ronald S. Ogden 53 Cancer

32591 Donna R. Bridges 26 Diabetes

Client 1

MiniCrypt

Client 2

MiniCrypt

Client 3

MiniCrypt

12809 Amanda D. Edwards 40 DiabetesZIP

ZIP

ZIP

ZIP

ZIP

ZIP


29489 Robert R. McGowan 56 Diabetes

13744 Kimberly R. Seay 51 Cancer

18740 Dennis G. Bates 32 Diabetes

98329 Ronald S. Ogden 53 Cancer

32591 Donna R. Bridges 26 Diabetes

Client 1

MiniCrypt

Client 2

MiniCrypt

Client 3

MiniCrypt

Problem: small compression ratio

12809 Amanda D. Edwards 40 DiabetesZIP

ZIP

ZIP

ZIP

ZIP

ZIP

Observation

Observation

1.6x

Observation

1.6x4.6x

Observation

1.6x4.6x

5.1x

Observation

Small # of key-value pairs gives high compression

1.6x4.6x

5.1x

Approach

Approach

1. Combine small number of key-value pairs into packs

Approach

1. Combine small number of key-value pairs into packs

2. Compress and encrypt each pack

Challenges

Challenges

1. Server cannot decrypt packs

Challenges

1. Server cannot decrypt packs

2. Multi-key transactions are expensive and not available for some key-value stores

Problem 1 Updates to different keys within a pack can overwrite each other

1013

1521

Client A

MiniCrypt

Client B

MiniCrypt

13: ID1

15: ID2

50: ID3

Index

1013

ID1:

Problem 2 Additional index is difficult to keep consistent

ID2: 1521

Client A

MiniCrypt

MiniCrypt’s techniques

• lookup algorithm

• update algorithm

• splitting algorithm

• epoch-based appends

• merge algorithm






• merge algorithm

using only






• merge algorithm

using only

• single-key atomic compare-and-swap






• merge algorithm

using only

• single-key atomic compare-and-swap

• sorted index on the primary key

API• get(key)

• put(key, value)

• delete(key)

• range(low_key, high_key)

Packingkey value

Sort by primary key

471411343328302627251510

Packingkey value

Sort by primary keysort

471411343328302627251510

Packingkey value

Sort by primary keysort

47

1411

3433

2830

2627

2515

10

Packing

101114

152526

272830

333447

key value

Partition into packs

Packing

101114

152526

272830

333447

key value

Packing

101114

152526

272830

333447

key valueNew key =

smallest key in the pack

Packing

101114

152526

272830

333447

key valueNew key =

smallest key in the pack

New value = a pack of key-value pairs

Packing

101114

152526

272830

333447

key value

10:

15:

27:

33:

New key = smallest key in the pack

New value = a pack of key-value pairs

Packing

101114

152526

272830

333447

key value

10:

15:

27:

33:

Compress & encrypt

each pack

Packing

10:

15:

27:

33:

101114

152526

272830

333447

key valueCompress &

encrypt each pack

10:

15:

27:

33:

Compress & encrypt

each pack

101114

152526

272830

333447

key value

Packing

ZIP

ZIP

ZIP

ZIP

Pack ID

10:

15:

27:

33:

Compress & encrypt

each pack

101114

152526

272830

333447

key value

Packing

ZIP

ZIP

ZIP

ZIP

Client

MiniCrypt

Single-key get

get(key): select pack with highest pack ID from all pack ID’s that are at most key

10:

15:

27:

33:

ZIP

ZIP

ZIP

ZIP

Client

MiniCrypt

Single-key get


10:

15:

27:

33:query = get(28)

ZIP

ZIP

ZIP

ZIP

Client

MiniCrypt

Single-key get


10:

15:

27:

33:query = get(28)

28ZIP

ZIP

ZIP

ZIP

Client

MiniCrypt

Single-key get


10:

15:

27:

33:query = get(28)

28

ZIP

ZIP

ZIP

ZIP

Client

MiniCrypt

Single-key get


10:

15:

27:

33:query = get(28)

ZIP

ZIP

ZIP

ZIP

Client

MiniCrypt

Single-key get


10:

15:

27:

33:query = get(28)

ZIP

ZIP

ZIP

Single-key put

10:

15:

27:

33:put(key, value): find pack using get()

Client

MiniCrypt

ZIP

ZIP

ZIP

ZIP

Single-key put

10:

15:

27:


Client

MiniCrypt

ZIP

ZIP

ZIP

ZIP

query = put(28, 5)

Single-key put

10:

15:

27:


Client

MiniCrypt

28ZIP

ZIP

ZIP

ZIP

query = put(28, 5)

Single-key put

10:

15:

27:


Client

MiniCrypt28

ZIP

ZIP

ZIP

ZIP

query = put(28, 5)

Single-key put

10:

15:

27:


Client

MiniCrypt

ZIP

ZIP

ZIP

ZIP

query = put(28, 5)

Single-key put

10:

15:

27:


Client

MiniCrypt

ZIP

ZIP

ZIP

query = put(28, 5)

Single-key put

10:

15:

27:


Client

MiniCrypt

ZIP

ZIP

ZIP

modify pack

query = put(28, 5)

Single-key put

10:

15:

27:


Client

MiniCrypt

ZIP

ZIP

ZIP

ZIP

modify pack

query = put(28, 5)

Single-key put

10:

15:

27:


Client

MiniCrypt

ZIP

ZIP

ZIP

ZIP

modify packwrite back using compare-and-swap

query = put(28, 5)

Problem: writes are expensive


• Solution: append-only mode



• Inserted keys are roughly monotonically increasing




• Key-value pairs can be modified within some bounded time




• Key-value pairs can be modified within some bounded time

• Applications: streaming data, time-series data

Approach

Approach• Server keeps track of epoch

Approach• Server keeps track of epoch• Each insert assigned an epoch

• Each new record inserted normally

Approach• Server keeps track of epoch• Each insert assigned an epoch

• Each new record inserted normally• Merge epochs into packs in the background

• merge deterministically

Append mode10:

15:

27:

33:

ZIP

ZIP

ZIP

ZIP

Client

MiniCrypt

Append modeCurrent epoch: 4 10:

15:

27:

33:

ZIP

ZIP

ZIP

ZIP

Client

MiniCrypt

Append mode

time

Current epoch: 4 10:

15:

27:

33:

ZIP

ZIP

ZIP

ZIP

Client

MiniCrypt

Append mode

time

Current epoch:

504

4 10:

15:

27:

33:

ZIP

ZIP

ZIP

ZIP

Client

MiniCrypt

Append mode

time

Current epoch:

504534

4 10:

15:

27:

33:

ZIP

ZIP

ZIP

ZIP

Client

MiniCrypt

Append mode

time

Current epoch:

504534614

4 10:

15:

27:

33:

ZIP

ZIP

ZIP

ZIP

Client

MiniCrypt

Append mode

time

Current epoch:

504534614

5 10:

15:

27:

33:

ZIP

ZIP

ZIP

ZIP

Client

MiniCrypt

Append mode

time

Current epoch:

504534614

625

5 10:

15:

27:

33:

ZIP

ZIP

ZIP

ZIP

Client

MiniCrypt

Append mode

time

Current epoch:

504534614

625635

5 10:

15:

27:

33:

ZIP

ZIP

ZIP

ZIP

Client

MiniCrypt

Append mode

time

Current epoch:

504534614

625635655

5 10:

15:

27:

33:

ZIP

ZIP

ZIP

ZIP

Client

MiniCrypt

5

Append mode

time

Current epoch:

504534614

625635655

10:

15:

27:

33:

ZIP

ZIP

ZIP

ZIP

Client

MiniCrypt

5

Append mode

time

Current epoch:

625635655

10:

15:

27:

33:

ZIP

ZIP

ZIP

ZIP

Client

MiniCrypt

504534614

5

Append mode

time

Current epoch:

625635655

10:

15:

27:

33:

ZIP

ZIP

ZIP

ZIP

Client

MiniCrypt

ZIP

5

Append mode

time

Current epoch:

625635655

10:

15:

27:

33:

ZIP

ZIP

ZIP

ZIP

Client

MiniCrypt

ZIP

50:

Evaluation

Evaluation• MiniCrypt built on unmodified Cassandra


• Comparison with


• Comparison with

• Baseline: each row compressed and encrypted


• Comparison with


• Vanilla: no encryption


• Comparison with



• YCSB benchmark


• Comparison with



• YCSB benchmark

• Anonymized Conviva data (video analytics)

Evaluation setup

Evaluation setup• Amazon EC2 cluster of 3 instances


• 15 GB of memory


• 15 GB of memory

• 8 cores


• 15 GB of memory

• 8 cores

• Zlib compression


• 15 GB of memory

• 8 cores

• Zlib compression

• AES-256 CBC for encryption

YCSB 100% single-record read

SSD


SSD

9.2x


SSD Disk

9.2x


SSD Disk

9.2x 100x

SSD

YCSB 100% range query

SSD


5x

SSD Disk


5x

SSD Disk


5x 38x

YCSB 100% single-record write

Default

YCSB 100% single-record write

Default Append-only

Conclusion

MiniCrypt shows that we can achieve compression and encryption using systems techniques

Conclusion

MiniCrypt shows that we can achieve compression and encryption using systems techniques

Thank you!

MiniCrypt: Reconciling Compression and Encryption in Big ...wzheng/eurosys_2017_talk.pdf ·...

Documents

Transcript of MiniCrypt: Reconciling Compression and Encryption in Big ...wzheng/eurosys_2017_talk.pdf ·...