RemoteFlashTuning

FlashStorageDisaggrega/onAnaKlimovic1,ChristosKozyrakis1,EnoThereska2,BinuJohn3andSanjeevKumar3

1StanfordUniversity,2ConfluentInc.&ImperialCollegeLondon,3FacebookInc.

Mo2va2on

•  FlashisoNenoverprovisionedindatacentersduetoimbalancedresourcerequirementsbetweenapplica/onsandover/me

•  ExposingFlashoverahighbandwidthnetwork(disaggrega(ngFlash)allowsustocustomizeresourcesperappatlowcost

DisaggregatedFlashArchitecture

ResultsSummary

Performance-CostAnalysis

FutureWork

•  StartwithiSCSI,astandardprotocolthattransportsblockstoragecommandsoverTCP/IP

•  IncreaseremoteFlashserverthroughputwithsystemsandnetworkop/miza/ons,achieving:

•  1.5×IOPSformicrobenchmark•  1.4×throughputfortargetapplica/on

✔ScaleFlashindependentlyfromCPU&RAM ✔ ShareFlashtoincreaseresourceu/liza/on ✔ Lower$/GBwhenbuyFlashinlargearray

•  Remotevs.localFlashdecreasesend-to-endthroughputby20%onaverage(and10%atthetail)forakey-valuestoreworkload

•  InherenttaillatencyoverheadofiSCSIis260µs,whiletheapplica/on’staillatencySLAis5-10ms

•  RemoteaccesstosharedFlashincreasestailreadlatencyduetomorewriteinterference

•  Op/mizenetworkprocessingstacktoreduceCPUoverheadofremoteFlashdataplane

• QoSmechanismsforsharingremoteFlash

•  Controlplaneforalloca/ngFlashtoapplica/ons

•  Disaggrega/ngFlashiscost-effec/vewhenFlash&CPUdemandsscaledifferently

Resourcesavingsofremotevs.localFlash(asapprequirementsscale)

DeploymoreFlashthanCPU

DeploymoreCPU,shareremoteFlash

I/O-intensivemicrobenchmark(4kBrandomreads)

StorageCapacityScalingFactor

Compu

teIntensity

ScalingFactor 40%

30%

20%

10%

0%

-10%

246810

10

8

6

4

2

End-to-EndPerformance

•  ModelrealFacebookapplica/onsthatuseFlashforembeddedkey-valuestoredatabase

Applica2onPerformancewithdatastoreaccessingremotevs.localFlash