Upload
others
View
2
Download
0
Embed Size (px)
Citation preview
HandheldDeviceArchitectures:AreWeDoingEnough?
ManuAwasthiAshokaUniversity
HandheldDevices
AndroidVersions
https://www.counterpointresearch.com/can-android-o-de-fragment-android/
MemoryUsage
https://eitik.com/17-android-browser-tested-for-memory-usage-in-2018/
AFewTrends
5
MobileArchitectureResearch● Mobile computing research: 1% of research papers published eachfocus on mobile computing.
● Lack of tools
V.J.Reddi,H.Yoon,andA.Knies,“Twobilliondevicesandcounting,”IEEEMicro,vol.38,no.1,pp.6–21,January/February2018.
Whatisneeded?
• ToolandSimulators
• Benchmarks
BBench(Michigan)
AsimBench(ICT,China)
CurrentStatus
AndroidEmulator
AndroidOpenSourceProject(AOSP)
AndroidVirtualDevice(AVD)
AndroidEmulatorDeviceBehavior
HowdoweleveragetheexistingecosystemtostudymemorybehaviorofAndroidapps?
META:ToolDesign
MobiCom2018
TracerModule- ModificationsinQEMU
ModifiedTranslationPath
RawTraces
META:ToolDesign :CacheModule
CacheSimulationModule
L1andL2CachehitratesavailableafterrunningacalculatoronarangeofandroidversionsfromAndroid4(Kitkat)toAndroid7(Nougat).
META:ToolDesign:MemoryModule
NVMainIntegration
● NVMain :cycle-levelmainmemorysimulator
● CansimulateDRAM,emergingnon-volatilememoriesatthearchitecturallevel.○ Phasechangememory,STT-RAM
PotentialUse-cases
● TraceGeneration○ Thetracescanalsobeusedtoanalyzeinstructiondistributionprofile.○ Creationofsyntheticinputstomodelsbasedonrealinstructionprofiles
● CacheHierarchyModeling○ Acustom, N-levelcachehierarchy
● DRAM,Non-volatile,HybridMemorySimulation○ NVMain canmodelmosttechnologies
TrendsinHandheldDevices
18
https://thehackernews.com/2015/09/6gb-ram-smartphone.html
http://www.es.ele.tue.nl/~kgoossens/Chandrasekar14PHD.pdfhttps://www.pwc.com/gx/en/technology/mobile-innovation/assets/pwc-dram-memory.pdf
HandheldsandSmartphones
19HexagonDSP:AnArchitectureOptimizedforMobileMultimediaandCommunications,IEEEMicro,Vol34,2014
MainMemoryinHandhelds
2020
MC
Core1 Core2
Core3Core4
Fabric Fabric
GPU
Camera
Display
Audio
Sensors
Modem
DRAM DRAM
HandheldApplications
21
Iffrequentlyaccesseddatacanbeconcentratedtothefastestregionsofahybridmemoryhierarchy,memory
system’senergyconsumptioncanbereducedsignificantly,without anysignificantlossinperformanceanduser
experience.
HybridMemoryArchitecturesforHandhelds
TotalPhysicalAddressSpace0GB 4GB
SubAddressSpace SubAddressSpace SubAddressSpace
DRAMMemTech2NVM
HybridMainMemoryinHandhelds
2323
MC
Core1 Core2
Core3Core4
Fabric Fabric
GPU
Camera
Display
Audio
Sensors
Modem
DRAM NVM
MC
DATE2018
HybridMainMemoryinHandhelds
24
MC
Core1 Core2
Core3Core4
Fabric Fabric
GPU
Camera
Display
Audio
Sensors
Modem
DRAM
NVM
MC MC MC
NVM
DRAM
DATE2018
Results
25
Results– 4Controllers
26
KeyTakeways
• Researchintohandhelddevicesarchitecturesisimportant,moresointheeraofwearables• Memorysub-systemisbecomingincreasinglyimportant,eveninhandheld• Needtools,benchmarkstocarryresearchforward• META– onestepinthatdirection
• NVMswilleventuallybeintegratedintomemoryhierarchy• Mechanismstoprovideaccesstohighcapacity,lowlatencymemoriesmightrequireintelligentdatamanagement• H/W– S/Wco-designisbetterthanoneortheother.
27
Acknowledgements
VarunGohil,Sneha Ved (IITGandhinagar)
NisargParikh(LDCollegeofEngineering)