we are living in a new virtualized world sorav bansal iit delhi feb 26, 2011

We are living in aNew Virtualized World

Sorav BansalIIT Delhi

Feb 26, 2011

Old Virtualized World

IBM Mainframes (circa 1960)

IBM Mainframe

VMM

OS OS OS

App AppAppAppApp App

New Virtualized World

“Cloud-OS”

OS OS OS

App AppAppAppApp App

“Cloud-OS”(stuff that you have heard many times before… uh yawn…)

• Infrastructure Layer (slave) + Management layer (master)

• Divide hardware into resource pools• Unit of abstraction = VM• Efficient• Effective Isolation• Dynamic• Fault-Tolerant

“Cloud-OS”(more exciting stuff)

• Dynamic Performance Optimizations– Compiler Optimizations– OS-level Optimizations

• Providing Determinism– Efficient Para-virtual Record/Replay

• Improving Reliability– Micro-Replays

“Cloud-OS”(more exciting stuff)

• Security– VMM-level security checks

• Efficient Thin Clients– Remote Desktopping using VM Record/Replay

Performance Optimizations

• Dynamic Binary Translation (Compiler Optimizations)– Translation Blocks– Direct Jump Chaining– Peephole Optimizations– Trace Optimizations– Exception rollbacks– Interrupt delays

Performance Optimizations

• Dynamic Binary Translation (OS-level Optimizations)– Eliminate traps from system calls– Better TLB/cache locality by using dedicated OS

cores

Traditional Picture

OS

Hardware

Application1 Application

2

Virtualized Picture

OS

Application1 Application

2

Optimizing VMM

bubsort

emptyloop

fibo_iter

hanoi1

hanoi2

hanoi3

printf

-15

-10

-5

0

5

10

15

-0.8%

0.3% 0.3%

12.9%

-3.1%

9.1%

Ove

rhea

d (P

erce

ntag

e of

Nati

ve)

Lower is Better

Some Initial Results

Providing Determinism: Record/Replay

• Uniprocessor– Non-determinism is quite low. Can be efficiently

recorded.

• Multiprocessor– Non-determinism high due to shared memory.– Recording overhead scales poorly with multiple

processors– Assuming we can patch the guest in some way, can

we improve this situation?

Micro-Replays

Snapshot

Recording non-determinism

Hit a Bug(e.g., assertion failure)

Execution timeline

Replay

Choose a rollback point.Also guess bug-inducingnon-deterministic choice

Potentially bug-freeexecution

Tolerating Non-deterministic Bugsusing Record/Replay

debit = 0;credit = total;

void transfer(void) { for (i = 0; i < 1000; i++) { debit--; credit++; assert(debit + credit == total); }}

for (t = 0; t < max_threads; t++) { thread_create(transfer);}

shared vars

unprotectedcriticalsection

VMM records an execution

On assert failure, the VMM interposes and rolls back the execution a few milliseconds

VMM guesses the non-deterministic choices that could have caused the failure (e.g., instruction at timer interrupt)

VMM replays the execution avoiding the previous non-deterministic choices

In this example, VMM infers thecritical section after a few runs and avoids interrupting it

Number of Replays Required?

360 720 1080 1440 1800 2160 2520 2880 3240 36000

5

10

15

20

25

30

35

Critical Section Size (bytes)

Num

ber o

f mic

ro-r

epla

ys re

quire

d

Technical Report:Micro-Replays: Improving Reliability in Presence of Non-deterministic Software Bugshttp://www.cse.iitd.ac.in/~sbansal/pubs/micro_replays.pdf

Security ExampleA Simple Scheme to Prevent Stack-Overflows

call

ret

…push ra, shadow…

…ra popra1pop shadowif (ra != ra1) error…

Remote Desktop Using Streaming VM Record/Replay

Typical Remote Desktop

Remote Desktop Using Streaming VM Record/Replay

Record

Replay

Remote Desktop using Streaming VM Rec/Rep

Bandwidth Comparison

Cumulative Data Transfer as function of time

Steady-state Bandwidth ComparisonRa

te (M

iB/s

)

Steady State Bandwidth Requirement

Conclusions

• We are living in a new virtualized world– Many implications in different application areas

Backup Slides

Translation Blocks

• Divide code into “translation blocks”– A translation block ends if• Reach a control-flow instruction• Or, MAX_INSNS instructions have been translated

A Simple Scheme

Original code

fragment

BinaryTranslator

x:Translated

code fragment

tx:

Use a Cache

Original code

fragment

BinaryTranslator

x:Translated

code fragment

tx:

TranslationCache

Lookup using xsavefound

not-found

Direct Jump Chaining

a

b c

d

Ta

Tb Tc

Td

lookup(b) lookup(c)

lookup(d) lookup(d)

Indirect Jumps

a

bf

call

ret

Ta

Tf

Tb

lookup(retaddr)

push bjmp Tf

pop retaddr

tmp JTABLE[retaddr & MASK]if (tmp.src == retaddr) goto tmp.dst

bubsort

emptyloop

fibo_iter

hanoi1

hanoi2

hanoi3

printf

-15

-10

-5

0

5

10

15

-0.8%

0.3% 0.3%

12.9%

-3.1%

9.1%

Ove

rhea

d (P

erce

ntag

e of

Nati

ve)

Lower is Better

euclid fibo_rec erastothenes0

50

100

150

200

250

300

350

400

156.7%

114.3%

37.4%

36x 46x 11x

defaultno jumptable

Ove

rhea

d (P

erce

ntag

e of

Nati

ve)

Lower is Better

fibo_rec

printf

1.1x45x

710x

no chainingno jumptabledefault

Series1

-9%

-6%

58%no chainingno jumptabledefault

printf

Overheads

logarithmic scale

1 3 5 7 9 11 13 15 17 19 21 23-50

0

50

100

150

200

250

bubsort

1 3 5 7 9 11 13 15 17 19 21 230

50

100

150

200

emptyloop

1 3 5 7 9 11 13 15 17 19 21 230

400

800

1200

1600

2000

euclid

1 3 5 7 9 11 13 15 17 19 21 230

50

100

150

200

250

fibo_iter

1 3 5 7 9 11 13 15 17 19 21 230

100200300400500

fibo_rec

1 3 5 7 9 11 13 15 17 19 21 230

50100150200250300

hanoi1

1 3 5 7 9 11 13 15 17 19 21 23-50

50

150

250

350

hanoi2

1 3 5 7 9 11 13 15 17 19 21 230

50100150200250300350

hanoi3

1 3 5 7 9 11 13 15 17 19 21 23-100

0

100

200

300

400

500

printf

Effect of Maximum Size of Translation Block

Max Size of Translation Block

Ove

rhea

d

Effect of Translation Cache Size

16 17 18 20 22 24 32 64 96 128-505

1015202530

bubsort

16 17 18 20 22 24 32 64 96 1280

10

20

30

40

emptyloop

16 17 18 20 22 24 32 64 96 1280

50100150200250300

euclid

16 17 18 20 22 24 32 64 96 128012345678

fibo_iter

16 17 18 20 22 24 32 64 96 12805

1015202530354045

hanoi1

16 17 18 20 22 24 32 64 96 128

-4-202468

1012

hanoi2

16 17 18 20 22 24 32 64 96 1280

5

10

15

20

25

30

hanoi316 17 18 20 22 24 32 64 96 128-50

050

100150200250300

printf

16 17 18 20 22 24 32 64 96 1280

20

40

60

80

erastothenes

Ove

rhea

d

Number of 4k pages in Translation Cache

clockrandom

Optimizations

• Peephole Optimizations• Trace Optimizations• Cross-layer optimizations

An Example

ld M, r1ld M, r0

ld M, r0mov r0, r1

Interrupts

ld M, r1ld M, r0

ld M, r0mov r0, r1

Delay Interrupt delivery till end ofcurrent translation

Precise Exceptions

ret ld (sp),t0add $4, sp…jmp t0 Page fault

sub $4, sprestore t0

rollback code

page fault handler