Eidetic Systems

Eidetic SystemsDavid Devecsery, Michael Chow, Xianzheng Dou,

Jason Flinn, Peter ChenUniversity of Michigan

What is an Eidetic System?

Eidetic – Having “Perfect memory” or “Total Recall”

Eidetic System – A system which can recall and trace through the lineage of any past computation

2David Devecsery

Motivation - Heartbleed

3

• Was Heartbleed exploited?• What data was leaked?

David Devecsery


4

HeartbleedMessage

Leaked Data

• Was Heartbleed exploited? - Yes• What data was leaked?

David Devecsery


5

• Was Heartbleed exploited? - Yes• What data was leaked?

Leaked Database Rows

David Devecsery

HeartbleedMessage

Leaked Data

Motivation – Wrong Reference

6

Bad Citation

• How did I get the wrong citation?

David Devecsery


7


David Devecsery


8


David Devecsery


9

• How did I get the wrong citation?• What else did this affect?

David Devecsery

Motivation

10

• How did I get the wrong citation?• What else did this affect?

David Devecsery

Arnold

•First practical eidetic computer system• Efficiently records & recalls all user-space computation• Process register/memory state• Inter-process communication

• Handles lineage queries• What data was affected?• What states and outputs were affected?

• Targeted towards desktop/workstation use•Reasonable overheads• Record 4 years of data on $150 commodity HD• Under 8% performance overhead on most benchmarks

11David Devecsery

Overview

• Introduction•Motivation•How Arnold remembers all state•How Arnold supports lineage queries•Conclusion

12David Devecsery

Remembering State

•Requirements:• Store years of state on a single disk• Memory/register space within a process• Inter process communication• File state

• Recall any state in reasonable time•Solution:• Deterministic record & replay• “Process group” based replay• “Process graph” to track inter-process lineage

• Log compression

13David Devecsery

Recording Granularity

•What granularity is best to record our system?

14

Pipe

1

2Read 1

Pipe

1

2Read 1 Pipe

1

2Read 1

Pipe

1

2Read 1

Pipe

1

2Read 1

ExternalInputs

David Devecsery


• Whole system recordingLow space overhead× Costly to replay

15

Pipe

1

2Read 1

Pipe

1

2Read 1 Pipe

1

2Read 1

Pipe

1

2Read 1

Pipe

1

2Read 1

ExternalInputs

David Devecsery


•Process level recordingEfficient to replay×Uses extra disk space×No Inter-process tracking

16

Pipe

1

2Read 1

Pipe

1

2Read 1 Pipe

1

2Read 1

Pipe

1

2Read 1

Pipe

1

2Read 1

ExternalInputs

David Devecsery


•Process group recordingEfficient to replayReasonable disk space×No Inter-process tracking

17

Pipe

1

2Read 1

Pipe

1

2Read 1 Pipe

1

2Read 1

Pipe

1

2Read 1

Pipe

1

2Read 1

ExternalInputs

David Devecsery

Implementation – Process Graph

18

Record Log

Pipe

1

2Read 1

1

Pipe

1

2Read 1Pipe

1

2Read 1

2

IPC Read

David Devecsery

Implementation – Process Graph

19

Record Log

Pipe

1

2Read 1

1

Pipe

1

2Read 1Pipe

1

2Read 1

2

IPC ReadPipe

1

2Read 1Pipe

1

2Read 1

David Devecsery

Recording

•Process group recording + process graphEfficient to replayReasonable disk spaceInter-process tracking

20

Pipe

1

2Read 1

Pipe

1

2Read 1 Pipe

1

2Read 1

Pipe

1

2Read 1

Pipe

1

2Read 1

ExternalInputs

David Devecsery

Space Optimizations

21

Baselin

e

+Model-

Based Compres

sion

+Ded

uplicate

d File

Cache

+X Se

rver C

ompressio

n

+Sem

i-Dete

rministi

c Tim

e+G

zip

0

0.2

0.4

0.6

0.8

1

1.2

Log

Com

pres

sion

vs B

asel

ine

David Devecsery

Space Optimizations

22

Baselin

e

+Model-

Based Compres

sion

+Ded

uplicate

d File

Cache

+X Se

rver C

ompressio

n

+Sem

i-Dete

rministi

c Tim

e+G

zip0

0.2

0.4

0.6

0.8

1

1.2

Log

Com

pres

sion

vs B

asel

ine

411:1 Ratio

David Devecsery

Space Optimizations

23

Baselin

e

+Model-

Based Compres

sion

+Ded

uplicate

d File

Cache

+X Se

rver C

ompressio

n

+Sem

i-Dete

rministi

c Tim

e+G

zip

Only Gzip

0

0.2

0.4

0.6

0.8

1

1.2

Log

Com

pres

sion

vs B

asel

ine

411:1 Ratio

6:1 Ratio

David Devecsery

Space Optimizations

24

Baselin

e

+Model-

Based Compres

sion

+Ded

uplicate

d File

Cache

+X Se

rver C

ompressio

n

+Sem

i-Dete

rministi

c Tim

e+G

zip

Only Gzip

0

0.2

0.4

0.6

0.8

1

1.2

Log

Com

pres

sion

vs B

asel

ine

411:1 Ratio

6:1 Ratio

David Devecsery

4 years of data on a $150 4TB commodity HD

Model-Based Compression

• Formulate a model of a typical execution • Only record deviations from that model

ret_val = sys_read (fd, buffer, count);

• Idea: Partial determinism• Encourage the program to conform to the model

25

usually equal

David Devecsery

Semi-Deterministic Time

• Frequent time queries are non-deterministic• Use partially deterministic clock• Real time clock & deterministic clock• Bound deviation

26

if (deterministic_clock – real_time_clock < threshold) {adjust deterministic_clockrecord deviation

}return deterministic_clock

David Devecsery

Performance Evaluation

27

kern

el copy

cvs c

heckout

make

latex

apac

hege

dit

facebook

spread

sheet

0.9

0.95

1

1.05

1.1

1.15

1.2 Baseline Arnold

Nor

mal

ized

Runti

me

David Devecsery

Overview

• Introduction•Motivation•How Arnold remembers all state•How Arnold supports lineage queries•Conclusion

28David Devecsery

Querying Lineage

•Two types of queries:•Reverse: Where did this data come from?•Forward: What did this data affect?

•How does Arnold support these queries?•User specifies initial state•Trace the lineage of the computation• Intra-process tracking• Inter-process tracking

29David Devecsery

Intra-Process Lineage

• Use taint tracking for intra-process causality• Run retroactively, on recorded execution• Parallelizable

• Arnold supports several notions of causality:

30

Copy Only Data Flow Data+Index Flow

May miss relations Misses few relationsRecall

Strong input/output relation

Weak input/output Relation

Precision

Control Flow

David Devecsery


31

Which linkage tool should Arnold use?

Inputs

Program

David Devecsery


32

Data Flow Data+IndexFlow




Precision

Copy

David Devecsery


33





Precision

Copy

David Devecsery


34





Precision

Copy

David Devecsery


35





Precision

Copy

David Devecsery


36





Precision

Copy

David Devecsery


37

Data Flow




Precision

Arnold selects themost precise tool withat least one result

David Devecsery

Inter-Process Lineage

• Two notions of inter-process linkage• Process graph• Tracks lineage through inter-process communication• Precise • Captures group to group communication

• Human linkage• Handles relations between user inputs and outputs• Infers linkages based on data content and time• Imprecise – may have false negatives and false positives• Can capture linkages the process graph can miss

38David Devecsery

Evaluation – Wrong Reference

39

Data + IndexDataCopyCopyData

• Few false positives (font files, latex sty files, libc.so, libXt.so)• No false negatives

Record Time Replay Time Replay + Pin Time

Query Time

96.1s 2.2s 70.0s 209.5s

HumanLinkage

David Devecsery

Evaluation – Heartbleed

40

• No false positives or negatives

Data + IndexData + Index Data + Index

Record Time Replay Time Replay + Pin Time

Query Time

230.3s 0.4s 139.5s 235.1s

David Devecsery

Conclusion

•Eidetic Systems are powerful tools• Complete vision into past computation• Answer powerful queries about state’s lineage

•Arnold – First practical Eidetic System• Low runtime overhead• 4 years of computation on a commodity HD• Supports powerful lineage queries

•Code is releasedhttps://github.com/endplay/omniplay

41David Devecsery

Documents

Eidetic Systems