varemu: an emulation testbed for variability-aware software€¦ · variability expedition varemu:...

CODES+ISSS 2013 Embedded Systems Week

!

VarEMU: An Emulation Testbed for Variability-Aware SoftwareLucas Wanner, Salma Elmalaki, Liangzhen Lai, Puneet Gupta, and Mani [email protected]

Copyright (c) 2013

Variability ExpeditionNESLNetworked & EmbeddedSystems Laboratory NanoCAD

mailto:[email protected]

Variability Expedition VarEMU: An Emulation Testbed for Variability-Aware Software CODES+ISSS 2013

The Variability Problem

2

Nominally identical parts, different characteristicsFrequency Errors Power

Varia

bilit

y (%

)

0

250

500

750

1000

Year2009 2012 2015 2018 2021 2024

Total PowerStatic (Sleep) Power


Variability-Aware Software

3

OS

App

Software

App

Sense and Expose


Variability-Aware Software

4

Quality

OS

App

Software

App

Activation Control

HW/SW Parameters

App Code

...

Sense and ExposeProblem: hardware is not instrumented for

variability sensing


Motivation

5

Sense and ExposeHow to Variations to Software?

Observ

ability

Perform

ance

Contro

l

Cost

Instrumented HardwareHW-In-The Loop SimGate/RTL-Level Sim

VarEMU

✔

✔

✔ ✔ ✔

✔ ✔ ✔ ✔


Architecture

6

QEMUVirtual Machine Monitor Runs unmodified SW images Cross-architecture support Binary translation (fast)


Architecture

7

VarEMUQEMU Core Power + Aging Model Fault Framework Energy + Time Accounting

Virtual Machine

Query energy Control faults Suffer faults

User + Software Monitor

Configure, change, query

Variability Expedition VarEMU: An Emulation Testbed for Variability-Aware Software CODES+ISSS 2013 8

VarEMU

Virtual Machine


Instruction disassembly & translation

Cycle & Time Accounting


Cycle and Time Accounting

Translation time

add r1, r2, r3

Execution time

Translated ops x y z

cycle counting

# cycles class

fault status

Special case: WFI (standby instruction): Start sleep interval until next interrupt


VarEMU

Virtual Machine




Power ModelAging Model


Aging and Power Model

11

Voltage!

Temperature

Power = f (current state)

➾

Active + Sleep Power and Aging

Model

➾➾

!

Parameters for each (class of) instructions

Frequency

change due to aging & user

input

technology + design

dependent


Power Model

12

Slee

p Po

wer

(µW

)

0

100

200

300

400

500

600

Temperature (°C)

0 10 20 30 40 50 60

Best Case Worst Case Nominal Case

normalized to match 45nm Cortex-M3 test chip


Aging Model

13

∂Vt (

V)

0

0.01

0.02

0.03

0.04

Time (s)

0 7500000 15000000 22500000 30000000

0.8V (r) 1V (r) 1.2V (r)0.8V (m) 1V (m) 1.2V (m)

60°C, referece (r) vs. model (m)


VarEMU

Virtual Machine





Energy Accounting

Energy = Time × Power


VarEMU

Virtual Machine





Energy Accounting

Fault Model


Faults

Translation time

add r1, r2, r3

Execution time


cycle counting

fault status


Faults

Translation time

add r1, r2, r3

Execution time


cycle counting

fault status

pre

post

replace


Fault Module: translation Time

18

Instruction marked as faulty?

No: generate standard code

!!!

Yes: generate two code paths

!!

mul rd, rn, rs mul rd, rn, rsfault module


Translated instruction !!!!!!!!

Fault Module: execution time

19

Tests

Standard (correct) version

Faulty version




20

Tests


Faulty version

Input: architectural state, instruction info

Privileged mode?Faults enabled in the VM?User-defined conditions

Chose between correct and faulty version




21

Tests


Faulty version

Example: MUL + stuck at one

env-‐>regs[rd] =

env-‐>regs[rm] *

env-‐>regs[rs]

| 0x01;


VarEMU

Virtual Machine





Energy Accounting

Fault Model

Virtual Hardware

Device


VarEMU

Virtual Machine


Virtual Hardware

Device


VM-VarEMU interaction

24

time [class]energy [class]cycles [class]

total timetotal energytotal cyclessleep time

sleep energyfault status

activesleep

• VM interacts with VarEMU through memory mapped registers

• Commands• read command creates a checkpoint for all

VarEMU registers• faults command propagates value to fault

module• zero: faults disabled• not zero: implementation-dependent,

typically means faults may now occur• Multiple registers for time, energy, cycles• Linux driver handles interaction through syscall

interface and provides per-process energy accounting


VarEMU Linux support

25

process

kernel

read registers save error status

read registers restore error status

VarEMU status is part of process

context


App example

26

int main() {

}

vemu_regs regs; ...

printf("active energy: %d\n", regs.total_act_energy);

syscall(VAR_SYSCALL_WRITE, FAULT_ST, &fault_status);

x = y * z;

int fault_status = 1;

fault_status = 0;

syscall(VAR_SYSCALL_WRITE, FAULT_ST, &fault_status);

syscall(VAR_SYSCALL_READ, READ_PROC, &regs);


VarEMU

Virtual Machine




Power ModelAging ModelI/O

Energy Accounting

Fault Model

Virtual Hardware

Device


VarEMU

Virtual Machine


I/O


Dynamic monitoring and control

• Socket interface (QEMU Machine Protocol)

• VarEMU status queries (energy, error status, ...)

• VarEMU power model parameter updates

• Example: change parameter zero (temperature) of class zero (all) to 60 (°C)

29

varemu-‐model-‐param p="0" v="60" c="0"


Evaluation: time accounting accuracy

30

Accu

racy

(%)

95%

96%

97%

98%

99%

100%

Time

0 50 100 150 200

empty loopmatrix multiplicationauto correlation

Hardware timing vs. VarEMU


Evaluation: overhead

31


Case study: approximate arithmetic

• Approximate adder using partial sums [Kahng and Kang, DAC’12]

• Fault implementation with variable accuracy: 25% accuracy means that 25% of the partial sums are corrected

• Clock period reduced by 25%, with 6% recovery overhead

32

99.8% of edges detected

18% faster, no increase in energy

97% of pixels within 5% of “correct” version


• Higher voltage is needed to support same frequency as the device ages• Non-adaptive voltage, margined for the worst-case• Adaptive voltage, dynamically tailored to match aging process

Case Study: Dynamic Reliability Management

33

11% energy savings


• Variability-Aware Real-Time OS [Martin, et. al., submitted to TECS]

• Learn power characteristics for each instance, tailor task activations and quality to achieve desired lifetime

• Application robustness to intermittent and permanent faults [Sharma et. at., ICCD’13]

• Fault model using VarEMU + ModelSim

• Code transformations (e.g. operand swap) to increase robustness to faults

Research and Education using VarEMU

34

B/B B/N B/W N/B N/N N/W W/B W/N W/W0

2

4

6

8

10

12

Instance and Temperature Profile (Inst/Temp)

Velo

city

Est

imate

Err

or

(RM

SE

)

Assumed Worst Case

With Instance Modeling

• Graduate-level courses at UCLA

• Battery models

• Fault experiments


Final remarks

• VarEMU: a tool to evaluate variability-aware software• Variability-aware aging and power models• Energy, cycle accounting• Fault framework• Linux + embedded software stack

• Per-process accounting• Fine-grained control over faults

• Limitations• Cycle counting: simple model doesn’t take pipeline hazards, branch

prediction misses, cache, etc. into account• Fault framework: only architecturally visible states can be altered

• External tools may be used to simulate more complex faults• Availability:

• VarEMU is open software. github.com/nesl/varemu• NSF Variability Expeditions: variability.org

35

http://github.com/nesl/varemu

http://variability.org

CODES+ISSS 2013 Embedded Systems Week

!

VarEMU: An Emulation Testbed for Variability-Aware SoftwareLucas Wanner, Salma Elmalaki, Liangzhen Lai, Puneet Gupta, and Mani [email protected]

Copyright (c) 2013

Variability ExpeditionNESLNetworked & EmbeddedSystems Laboratory NanoCAD

mailto:[email protected]

varemu: an emulation testbed for variability-aware software€¦ · variability expedition varemu:...

Documents