NVM & Implications on Data Infratsructure

22
Non Volatile Memory & Implications on Data Infrastructure Robert Geiger and Shankar Hundekar, Ampool Inc.

Transcript of NVM & Implications on Data Infratsructure

Page 1: NVM & Implications on Data Infratsructure

Non Volatile Memory & Implications on Data Infrastructure

Robert Geiger and Shankar Hundekar, Ampool Inc.

Page 2: NVM & Implications on Data Infratsructure

Memory and Storage Directions

Robert Geiger, Chief Architect and VP Engineering, Ampool Inc.

Page 3: NVM & Implications on Data Infratsructure

PreviousStorageClassMemory(SCM)

• SSDand“keepthepoweron”solutions

• SATASSD;NANDFlash§ NOTE:NORflashisbyteaddressablebutmoreexpensive(duetolowerdensity)

• NVMe SSD;NANDFlash(veryfast;2GBytes/seconnewMaclaptop)

• BatterybackedDRAM(replacementanissue)

• NVDIMM-F(blockdevice;highcapacity,high latency)– PCIe

• NVDIMM-N(byteaddressableatornearDRAMspeeds)– PCIe

Page 4: NVM & Implications on Data Infratsructure

EmergingStorageClassMemory(SCM)

• ActuallyStorageClassByteAddressable Memory

• NearDRAMlatencyandthroughput

• Basedononeofseveraltypesofmemorytechnology• MRAM(magnetic)• ReRAM (resistive)• FRAM(ferroelectric),PCM(phasechange)• XPoint (Intel/Micronwon’ttellyouexactly)

• 3-DXPoint (Intel-Micron)- DDRorPCIe

Page 5: NVM & Implications on Data Infratsructure

SCMintheMemory/StorageHierarchy

Page 6: NVM & Implications on Data Infratsructure

SCMInterfaces

• Filesystembased(standard&memorymapped)• Filesystemsgetting reallyfast

§ TestanewMaclaptopw/SSD- youcando2GBytes/secsustainedread/write• Linux(fedoracore/RedHat)andMSaddingSCMsupport in‘16• Capacitywillstillpushbigdatatodiskfordatalake,colddata• Whatdoesthismeanforin-memoryplays?

§ Beingin-memory isnolongerenough§ Sophisticationmustincrease

• PMEM.IO=>DirectaccesstoSCM• Mostlikelyinterfaceforin-memoryfocusedsolutions

Page 7: NVM & Implications on Data Infratsructure

In-MemoryFuture

• Linebetweenmemoryandstorageisblurring

• StorageisjustmemorywithaspecificQualityofService(QoS)

• Managingmultiplelayers/typesofmemorywilldifferentiate

• Nolongerin-memoryvs.disk

• Nowfilesystemvs.managedbyteaddressableobjectstore

Page 8: NVM & Implications on Data Infratsructure

WhatMattersforApp,DB,andCompute

• Theflexibilityandeaseofuseoftheinterfacesprovided

• HowmemoryismanagedtransparentlytoprovideQoS

• Theserviceguaranteeabstractions(QoS choices)provided

• Howconflictsaremanagedandprevented

Page 9: NVM & Implications on Data Infratsructure

WhatitMeanstoBigDataandAnalytics

• Throughput,latency,andcapacity:whatQoS isneeded?

• Avoidingpipelineconflictswithmultiplesolutionsusingmemory

• In-memorysolutionscancontrolallocations

• Thiswillallowmultipleframeworkstoco-existbetter

• “Yarnformemory”

• Moretypesofmemory:evictionalgorithmsneedadjustment• Hot,cold=>hot,warm,lukewarm,cold

Page 10: NVM & Implications on Data Infratsructure

Implications on Data Infrastructure

Shankar Hundekar. Senior Software Engineer, Ampool Inc.

Page 11: NVM & Implications on Data Infratsructure

ApacheGeodeProject

*Sourceapachegeode

Arealtimein-memorydatamanagementplatform

Page 12: NVM & Implications on Data Infratsructure

ApacheGeodeArchitecture

!

"

Locator

Server

Server

Server

Server +###

##

$"

$$$$

$

$ $

%%%%

$

$ $

%

' …

Page 13: NVM & Implications on Data Infratsructure

PersistenceinApacheGeode…

*Sourceapachegeode

� High availability

� Low latency, high throughput

� Durability

� Minimal impact to in memory speed– Minimize IO operations– Minimize locking– Minimize deserialization

Page 14: NVM & Implications on Data Infratsructure

…isaShared-nothingarchitecture

*Sourceapachegeode

Server 3Server 2

Bucket 3*primary*

Bucket 1secondary

Bucket 3secondary

Bucket 2*primary*

Server 1

Bucket 2secondary

Bucket 1*primary*

Page 15: NVM & Implications on Data Infratsructure

ampool• PoweredbyApacheGeode• BroadsupportforBigDatasolutions

• Pluggable,distributedmemorylayer

Page 16: NVM & Implications on Data Infratsructure

In-MemoryObjectStoreoptimized forAnalyticsampool

PLUGGABLE distributedmemorylayer

%

Page 17: NVM & Implications on Data Infratsructure

In-MemoryObjectStoreoptimized forAnalyticsampool

%

MANAGEDFLEXIBILITY

✅‘Bestofbreed’engines

✅ExtensibleCore

✅Pluggablestores

Page 18: NVM & Implications on Data Infratsructure

ampool

In-MemoryDistributedSys

Low-latencyComms.

Key-ValueStore

FunctionExecution

+

HighThroughput

TableStore

NativeInterfacePluggableStoreManager

JavaAPI

MASH(CLIExt)

JavaAPI

BuildingonPROVEN In-memoryTechnology

Page 19: NVM & Implications on Data Infratsructure

InmemoryStorageManagerforAnalytics

StorageManager

SCM

DRAM

• LeveragesStoragecharacteristics• Basedontheusageofdata• TieredStoragearchitecture …

Page 20: NVM & Implications on Data Infratsructure

Aboutthepmem.io library

• OpensourcelibrarythatsimplifiesdevelopmentofappsusingSCM

• WorkswithexistingCPUhardware

• Focusesspecificallyonhowpersistentmemory isexposedtoserver-classapplications

• Applicationswillexplicitlymanagetheplacementofdataamongthethreetiers(volatilememory,persistentmemory,andstorage)

*Sourcepmem.io

Page 21: NVM & Implications on Data Infratsructure

Librariesprovidedbypmem

• libpmem provides lowlevelpersistentmemorysupport

• libpmemobj libraryprovidesatransactionalobjectstore,providingmemoryallocation,transactions,andgeneralfacilitiesforpersistentmemoryprogramming.Developersnewtopersistentmemoryprobablywanttostartwiththislibrary

• libpmemblk supportsarraysofpmem-residentblocks,allthesamesize,thatareatomicallyupdated

• libpmemlog providesapmem-residentlog file.Thisisusefulforprograms likedatabasesthatappendfrequently toalogfile

*Sourcepmem.io

Page 22: NVM & Implications on Data Infratsructure

Conclusion

• NewStorageClassMemoriesarechangingthein-memorylandscape

• Bothopportunitiesandchallengesforin-memorysolutions

• ApacheGeodeisaprovenin-memorytechnologyplatform

• Ampool isworkingtomeetthechallengesofthesechanges!