Submit Search
Upload
Page Fault Support for Network Controllers
•
0 likes
•
75 views
TakuyaFukuoka2
Follow
Reading Circle
Read less
Read more
Software
Report
Share
Report
Share
1 of 33
Download now
Download to read offline
Recommended
LITE Kernel RDMA Support for Datacenter Applications
LITE Kernel RDMA Support for Datacenter Applications
TakuyaFukuoka2
A Survey on Performance Analytical Tools for Partitioned Global Address Space
A Survey on Performance Analytical Tools for Partitioned Global Address Space
TakuyaFukuoka2
Loom: flexible and efficient NIC packet scheduling
Loom: flexible and efficient NIC packet scheduling
TakuyaFukuoka2
Microkernels in the Era of Data-Centric Computing
Microkernels in the Era of Data-Centric Computing
Martin Děcký
Hardware/Software Co-Design for Efficient Microkernel Execution
Hardware/Software Co-Design for Efficient Microkernel Execution
Martin Děcký
Formal Verification of Functional Code
Formal Verification of Functional Code
Martin Děcký
Software Preservation: challenges and opportunities for reproductibility (Sci...
Software Preservation: challenges and opportunities for reproductibility (Sci...
Roberto Di Cosmo
IPC in Microkernel Systems, Capabilities
IPC in Microkernel Systems, Capabilities
Martin Děcký
Recommended
LITE Kernel RDMA Support for Datacenter Applications
LITE Kernel RDMA Support for Datacenter Applications
TakuyaFukuoka2
A Survey on Performance Analytical Tools for Partitioned Global Address Space
A Survey on Performance Analytical Tools for Partitioned Global Address Space
TakuyaFukuoka2
Loom: flexible and efficient NIC packet scheduling
Loom: flexible and efficient NIC packet scheduling
TakuyaFukuoka2
Microkernels in the Era of Data-Centric Computing
Microkernels in the Era of Data-Centric Computing
Martin Děcký
Hardware/Software Co-Design for Efficient Microkernel Execution
Hardware/Software Co-Design for Efficient Microkernel Execution
Martin Děcký
Formal Verification of Functional Code
Formal Verification of Functional Code
Martin Děcký
Software Preservation: challenges and opportunities for reproductibility (Sci...
Software Preservation: challenges and opportunities for reproductibility (Sci...
Roberto Di Cosmo
IPC in Microkernel Systems, Capabilities
IPC in Microkernel Systems, Capabilities
Martin Děcký
Artificial intelligence in IoT-to-core network operations and management
Artificial intelligence in IoT-to-core network operations and management
ADVA
ARIN 34 IPv6 IAB/IETF Activities Report
ARIN 34 IPv6 IAB/IETF Activities Report
ARIN
The Impact on Security due to the Vulnerabilities Existing in the network a S...
The Impact on Security due to the Vulnerabilities Existing in the network a S...
IJAEMSJORNAL
Asad Javaid Network Administrator CV
Asad Javaid Network Administrator CV
Muhammad Asad Javaid
The Evolution of IP Storage and Its Impact on the Network
The Evolution of IP Storage and Its Impact on the Network
EMC
Software Defined Networking
Software Defined Networking
Sonam Dargay
Asigra Product Marketing Strategy
Asigra Product Marketing Strategy
Jas Mann
NFV Tutorial
NFV Tutorial
Rashid Mijumbi
NFV Tutorial
NFV Tutorial
Rashid Mijumbi
Evolving Virtual Networking with IO Visor [OpenStack Summit Austin | April 2016]
Evolving Virtual Networking with IO Visor [OpenStack Summit Austin | April 2016]
IO Visor Project
Evolving Virtual Networking with IO Visor
Evolving Virtual Networking with IO Visor
Larry Lang
IT Vulnerability & Tools Watch 2011
IT Vulnerability & Tools Watch 2011
WASecurity
NFV evolution towards 5G
NFV evolution towards 5G
Marie-Paule Odini
Global IoT Cloud Services Survey-Aug-20160527
Global IoT Cloud Services Survey-Aug-20160527
August Lin
Metaverse - The 'Killer App' for 5G, 6G and Beyond
Metaverse - The 'Killer App' for 5G, 6G and Beyond
Anand Bhojan
The Role of Machine Learning in Fluid Network Control and Data Planes.pdf
The Role of Machine Learning in Fluid Network Control and Data Planes.pdf
Förderverein Technische Fakultät
14A81A05B1
14A81A05B1
Chaitanya Ram
IRJET- Survey on SDN based Network Intrusion Detection System using Machi...
IRJET- Survey on SDN based Network Intrusion Detection System using Machi...
IRJET Journal
Understanding NFV and the MANO stack
Understanding NFV and the MANO stack
Affan Syed
Akraino and Edge Computing
Akraino and Edge Computing
Liz Warner
業務で ISUCON することになった話.pdf
業務で ISUCON することになった話.pdf
TakuyaFukuoka2
Using RDMA Efficiently for Key-Value Services
Using RDMA Efficiently for Key-Value Services
TakuyaFukuoka2
More Related Content
Similar to Page Fault Support for Network Controllers
Artificial intelligence in IoT-to-core network operations and management
Artificial intelligence in IoT-to-core network operations and management
ADVA
ARIN 34 IPv6 IAB/IETF Activities Report
ARIN 34 IPv6 IAB/IETF Activities Report
ARIN
The Impact on Security due to the Vulnerabilities Existing in the network a S...
The Impact on Security due to the Vulnerabilities Existing in the network a S...
IJAEMSJORNAL
Asad Javaid Network Administrator CV
Asad Javaid Network Administrator CV
Muhammad Asad Javaid
The Evolution of IP Storage and Its Impact on the Network
The Evolution of IP Storage and Its Impact on the Network
EMC
Software Defined Networking
Software Defined Networking
Sonam Dargay
Asigra Product Marketing Strategy
Asigra Product Marketing Strategy
Jas Mann
NFV Tutorial
NFV Tutorial
Rashid Mijumbi
NFV Tutorial
NFV Tutorial
Rashid Mijumbi
Evolving Virtual Networking with IO Visor [OpenStack Summit Austin | April 2016]
Evolving Virtual Networking with IO Visor [OpenStack Summit Austin | April 2016]
IO Visor Project
Evolving Virtual Networking with IO Visor
Evolving Virtual Networking with IO Visor
Larry Lang
IT Vulnerability & Tools Watch 2011
IT Vulnerability & Tools Watch 2011
WASecurity
NFV evolution towards 5G
NFV evolution towards 5G
Marie-Paule Odini
Global IoT Cloud Services Survey-Aug-20160527
Global IoT Cloud Services Survey-Aug-20160527
August Lin
Metaverse - The 'Killer App' for 5G, 6G and Beyond
Metaverse - The 'Killer App' for 5G, 6G and Beyond
Anand Bhojan
The Role of Machine Learning in Fluid Network Control and Data Planes.pdf
The Role of Machine Learning in Fluid Network Control and Data Planes.pdf
Förderverein Technische Fakultät
14A81A05B1
14A81A05B1
Chaitanya Ram
IRJET- Survey on SDN based Network Intrusion Detection System using Machi...
IRJET- Survey on SDN based Network Intrusion Detection System using Machi...
IRJET Journal
Understanding NFV and the MANO stack
Understanding NFV and the MANO stack
Affan Syed
Akraino and Edge Computing
Akraino and Edge Computing
Liz Warner
Similar to Page Fault Support for Network Controllers
(20)
Artificial intelligence in IoT-to-core network operations and management
Artificial intelligence in IoT-to-core network operations and management
ARIN 34 IPv6 IAB/IETF Activities Report
ARIN 34 IPv6 IAB/IETF Activities Report
The Impact on Security due to the Vulnerabilities Existing in the network a S...
The Impact on Security due to the Vulnerabilities Existing in the network a S...
Asad Javaid Network Administrator CV
Asad Javaid Network Administrator CV
The Evolution of IP Storage and Its Impact on the Network
The Evolution of IP Storage and Its Impact on the Network
Software Defined Networking
Software Defined Networking
Asigra Product Marketing Strategy
Asigra Product Marketing Strategy
NFV Tutorial
NFV Tutorial
NFV Tutorial
NFV Tutorial
Evolving Virtual Networking with IO Visor [OpenStack Summit Austin | April 2016]
Evolving Virtual Networking with IO Visor [OpenStack Summit Austin | April 2016]
Evolving Virtual Networking with IO Visor
Evolving Virtual Networking with IO Visor
IT Vulnerability & Tools Watch 2011
IT Vulnerability & Tools Watch 2011
NFV evolution towards 5G
NFV evolution towards 5G
Global IoT Cloud Services Survey-Aug-20160527
Global IoT Cloud Services Survey-Aug-20160527
Metaverse - The 'Killer App' for 5G, 6G and Beyond
Metaverse - The 'Killer App' for 5G, 6G and Beyond
The Role of Machine Learning in Fluid Network Control and Data Planes.pdf
The Role of Machine Learning in Fluid Network Control and Data Planes.pdf
14A81A05B1
14A81A05B1
IRJET- Survey on SDN based Network Intrusion Detection System using Machi...
IRJET- Survey on SDN based Network Intrusion Detection System using Machi...
Understanding NFV and the MANO stack
Understanding NFV and the MANO stack
Akraino and Edge Computing
Akraino and Edge Computing
More from TakuyaFukuoka2
業務で ISUCON することになった話.pdf
業務で ISUCON することになった話.pdf
TakuyaFukuoka2
Using RDMA Efficiently for Key-Value Services
Using RDMA Efficiently for Key-Value Services
TakuyaFukuoka2
Hyperbolic Caching: Flexible Caching for Web Applications
Hyperbolic Caching: Flexible Caching for Web Applications
TakuyaFukuoka2
Latency-Tolerant Software Distributed Shared Memory
Latency-Tolerant Software Distributed Shared Memory
TakuyaFukuoka2
Tardis: Time Traveling Coherence Algorithm for Distributed Shared Memory
Tardis: Time Traveling Coherence Algorithm for Distributed Shared Memory
TakuyaFukuoka2
Hybrid MPI: Efficient Message Passing for Multi-core Systems
Hybrid MPI: Efficient Message Passing for Multi-core Systems
TakuyaFukuoka2
Overlapping Communication and Computation by Using a Hybrid MPI/SMPSs Approach
Overlapping Communication and Computation by Using a Hybrid MPI/SMPSs Approach
TakuyaFukuoka2
More from TakuyaFukuoka2
(7)
業務で ISUCON することになった話.pdf
業務で ISUCON することになった話.pdf
Using RDMA Efficiently for Key-Value Services
Using RDMA Efficiently for Key-Value Services
Hyperbolic Caching: Flexible Caching for Web Applications
Hyperbolic Caching: Flexible Caching for Web Applications
Latency-Tolerant Software Distributed Shared Memory
Latency-Tolerant Software Distributed Shared Memory
Tardis: Time Traveling Coherence Algorithm for Distributed Shared Memory
Tardis: Time Traveling Coherence Algorithm for Distributed Shared Memory
Hybrid MPI: Efficient Message Passing for Multi-core Systems
Hybrid MPI: Efficient Message Passing for Multi-core Systems
Overlapping Communication and Computation by Using a Hybrid MPI/SMPSs Approach
Overlapping Communication and Computation by Using a Hybrid MPI/SMPSs Approach
Recently uploaded
Unlocking the Future of AI Agents with Large Language Models
Unlocking the Future of AI Agents with Large Language Models
aagamshah0812
Hand gesture recognition PROJECT PPT.pptx
Hand gesture recognition PROJECT PPT.pptx
bodapatigopi8531
Salesforce Certified Field Service Consultant
Salesforce Certified Field Service Consultant
AxelRicardoTrocheRiq
How To Troubleshoot Collaboration Apps for the Modern Connected Worker
How To Troubleshoot Collaboration Apps for the Modern Connected Worker
ThousandEyes
Building a General PDE Solving Framework with Symbolic-Numeric Scientific Mac...
Building a General PDE Solving Framework with Symbolic-Numeric Scientific Mac...
stazi3110
Diamond Application Development Crafting Solutions with Precision
Diamond Application Development Crafting Solutions with Precision
SolGuruz
The Ultimate Test Automation Guide_ Best Practices and Tips.pdf
The Ultimate Test Automation Guide_ Best Practices and Tips.pdf
kalichargn70th171
Professional Resume Template for Software Developers
Professional Resume Template for Software Developers
Vinodh Ram
Test Automation Strategy for Frontend and Backend
Test Automation Strategy for Frontend and Backend
Arshad QA
DNT_Corporate presentation know about us
DNT_Corporate presentation know about us
Dynamic Netsoft
Short Story: Unveiling the Reasoning Abilities of Large Language Models by Ke...
Short Story: Unveiling the Reasoning Abilities of Large Language Models by Ke...
kellynguyen01
How To Use Server-Side Rendering with Nuxt.js
How To Use Server-Side Rendering with Nuxt.js
Andolasoft Inc
(Genuine) Escort Service Lucknow | Starting ₹,5K To @25k with A/C 🧑🏽❤️🧑🏻 89...
(Genuine) Escort Service Lucknow | Starting ₹,5K To @25k with A/C 🧑🏽❤️🧑🏻 89...
gurkirankumar98700
The Essentials of Digital Experience Monitoring_ A Comprehensive Guide.pdf
The Essentials of Digital Experience Monitoring_ A Comprehensive Guide.pdf
kalichargn70th171
Try MyIntelliAccount Cloud Accounting Software As A Service Solution Risk Fre...
Try MyIntelliAccount Cloud Accounting Software As A Service Solution Risk Fre...
MyIntelliSource, Inc.
Adobe Marketo Engage Deep Dives: Using Webhooks to Transfer Data
Adobe Marketo Engage Deep Dives: Using Webhooks to Transfer Data
BradBedford3
Unveiling the Tech Salsa of LAMs with Janus in Real-Time Applications
Unveiling the Tech Salsa of LAMs with Janus in Real-Time Applications
Alberto González Trastoy
Optimizing AI for immediate response in Smart CCTV
Optimizing AI for immediate response in Smart CCTV
shikhaohhpro
Cloud Management Software Platforms: OpenStack
Cloud Management Software Platforms: OpenStack
VICTOR MAESTRE RAMIREZ
CALL ON ➥8923113531 🔝Call Girls Kakori Lucknow best sexual service Online ☂️
CALL ON ➥8923113531 🔝Call Girls Kakori Lucknow best sexual service Online ☂️
anilsa9823
Recently uploaded
(20)
Unlocking the Future of AI Agents with Large Language Models
Unlocking the Future of AI Agents with Large Language Models
Hand gesture recognition PROJECT PPT.pptx
Hand gesture recognition PROJECT PPT.pptx
Salesforce Certified Field Service Consultant
Salesforce Certified Field Service Consultant
How To Troubleshoot Collaboration Apps for the Modern Connected Worker
How To Troubleshoot Collaboration Apps for the Modern Connected Worker
Building a General PDE Solving Framework with Symbolic-Numeric Scientific Mac...
Building a General PDE Solving Framework with Symbolic-Numeric Scientific Mac...
Diamond Application Development Crafting Solutions with Precision
Diamond Application Development Crafting Solutions with Precision
The Ultimate Test Automation Guide_ Best Practices and Tips.pdf
The Ultimate Test Automation Guide_ Best Practices and Tips.pdf
Professional Resume Template for Software Developers
Professional Resume Template for Software Developers
Test Automation Strategy for Frontend and Backend
Test Automation Strategy for Frontend and Backend
DNT_Corporate presentation know about us
DNT_Corporate presentation know about us
Short Story: Unveiling the Reasoning Abilities of Large Language Models by Ke...
Short Story: Unveiling the Reasoning Abilities of Large Language Models by Ke...
How To Use Server-Side Rendering with Nuxt.js
How To Use Server-Side Rendering with Nuxt.js
(Genuine) Escort Service Lucknow | Starting ₹,5K To @25k with A/C 🧑🏽❤️🧑🏻 89...
(Genuine) Escort Service Lucknow | Starting ₹,5K To @25k with A/C 🧑🏽❤️🧑🏻 89...
The Essentials of Digital Experience Monitoring_ A Comprehensive Guide.pdf
The Essentials of Digital Experience Monitoring_ A Comprehensive Guide.pdf
Try MyIntelliAccount Cloud Accounting Software As A Service Solution Risk Fre...
Try MyIntelliAccount Cloud Accounting Software As A Service Solution Risk Fre...
Adobe Marketo Engage Deep Dives: Using Webhooks to Transfer Data
Adobe Marketo Engage Deep Dives: Using Webhooks to Transfer Data
Unveiling the Tech Salsa of LAMs with Janus in Real-Time Applications
Unveiling the Tech Salsa of LAMs with Janus in Real-Time Applications
Optimizing AI for immediate response in Smart CCTV
Optimizing AI for immediate response in Smart CCTV
Cloud Management Software Platforms: OpenStack
Cloud Management Software Platforms: OpenStack
CALL ON ➥8923113531 🔝Call Girls Kakori Lucknow best sexual service Online ☂️
CALL ON ➥8923113531 🔝Call Girls Kakori Lucknow best sexual service Online ☂️
Page Fault Support for Network Controllers
1.
PageFaultSupportforNetwork Controllers ReadingCircle M2TakuyaFukuoka May27,2020 1
2.
AboutThisPaper Author:IlyaLesokhin,HaggaiEran,ShacharRaindel,GuyShapiro, SagiGrimberg,LiranLiss,MuliBen‑Yehuda,NadavAmitandDan Tsafrir ThecollaborationofTechnisonIsraelInstituteofTechnology, MellanoxTechnologiesandVMwareResearch PublishedinASPLOS'17 Proposedesignandimplementationofnetworkpagefault(NPF) supportforInfiniBandandEthernetNICs NPFsupportofInfiniBandisalreadyavailableforusason‑demand paging(ODP)[1] [1]https://community.mellanox.com/s/article/understanding‑on‑demand‑paging‑‑odp‑x 2
3.
BriefSummary DirectI/Odoesnotmakeuseoftwobenefitsofvirtualmemory Simplifiedprogrammingmodel Canonicalmemoryoptimization Networkpagefaults(NPFs)supportcansolvetheproblem HowtohandlereceiveNPFs(rNPFs)isachallengingtask ThispaperproposesdesignandimplementationofNPFssupport forInfiniBandandEthernetNICwithrNPFshandling Theevaluationshowsproposedmethodsachievememory efficienceandgoodperformance 3
4.
ThreeBenefitsofVertualMemory 4
5.
TheRiseofDirectNetworkI/O OS‑providedabstractionandisolationmightdegradeperformance CustomizedIOchannelssoftwarecanbe4‑20xmore performantthanawell‑tunedPOSIX‑compliantprocess Overheadscanbeavoidedifhardwaresupportabstractionand isolationdirectly 5
6.
ExamplesofDirectNetworkI/O RemoteDMA(RDMA) Single‑RootI/OVirautalization(SRIOV) DataPlaneDevlopmentKit(DPDK) https://ru.mellanox.com/products/adapter‑ethernet‑SW/RDMA‑RoCE‑Ethernet‑Network‑Efficiency https://access.redhat.com/documentation/ja‑ jp/red_hat_enterprise_linux/6/html/virtualization_host_configuration_and_guest_installation_guide/chap‑ virtualization_host_configuration_and_guest_installation_guide‑sr_iov http://www.mashfords.com/2018/08/21/dpdk‑data‑plane‑development‑kit‑for‑linux‑vms‑now‑generally‑available/ 6
7.
TheProblemofDirectNetworkNIC NICscannotcopewithpagefaultswhichDMAdevicesexperience Thereforelosingtwovirtualmemorybenefits 7
8.
HowtoMakeUseofVirtualMemory BenefitswithDMA Twoexistingmethods Staticallypinningtheentireaddressspacetophysicalmemory Dynamicallypin/unpinningDMAtargetbuffers Simplifiedprogramming model CanonicalMemory Optimization Static Yes No Dynamic
No Yes NPF (Proposed) Yes Yes 8
9.
ExistingDMAPageFaultsSupport Recently,PCIestandardprovidessimilarfunctionalityofNPFs ButtheproblemofreceiveNPFs(rNPFs)causedbyincoming externaldataisoutsidethescope ThispaperproposesimplementationwithsupportofrNPFs 9
10.
Implementation ▶InfiniBandPageFaultSupport Basicidea:relyonRCprotocolofInfiniBand EthernetPageFaultSupport Basicidea:preparebackupring 10
11.
InfiniBandImplementationofNPFs ImplementtheNPFsbymodifyingthedriverandfirmwareof InfiniBand NPFflowisbasicallyasfollows 1.NICconsultstheIOMMUpagetablesandcheckifNPFisneeded 2.Thefarmware[1]detectsthisfaultandraisesanNPFinterrupt 3.Thedrives'sNPFinterrputhandlerqueriestheOS 4.ThedriverupdatestheIOMMUpagetableandinformsthe farmware [1]farmwareissoftwareinNIC 11
12.
OverheadofNPFsinInfiniBand MellanoxConnect‑IB56Gb/sInfiniBandNIC Send4KBand4MBmessageswithnodiskaccess Theoverheadisdominatedbymodifiedhardware(firmware) 12
13.
HowtoHandlerNPFsinInfiniBand Makeuseofreliableconnection(RC)protocolofInfiniBandwhich assuresthemessagedelivery RNRNACKandretransmission Whenthereceiverisnotready,itsendareceiver‑not‑ready(RNR) negativeacknowledgmentpacket(NACK)toaskthesendertowait foraspecifiedtime Retransmitpacketsincaseofpacket‑drops 13
14.
Implementation InfiniBandPageFaultSupport Basicidea:relyonRCprotocolofInfiniBand ▶EthernetPageFaultSupport Basicidea:preparebackupring 14
15.
EthernetImplementationofNPFs TCPalsoassuresthepacketdelivery.DoesithandlerNPFsinthe samewayasInfiniBand? TheanswerisNo Thatisbecausepacketdropsareconsideredtoasignof congestionandcauseTCPtoreducethetransmissionrate Thecommunicationmayeventuallyhalt 15
16.
HowmuchSufferfromThisProblem Comparethethroughputofmemcached(key‑valuestore) Dropping(whichrelyonTCPretransmission)suffersforabout60 secondsduringwhichitsthroughputisalmostzero 16
17.
Solution:PrepareBackupRing Ifapagefaultisencountered,thepacketiswrittentoasmall, pinnedbackupringownedbytheIOprovider(OSorhypervisor) 17
18.
Evaluation MemoryUtilization Key‑valuestore Storagesystem OverheadofPinningandCopying MPImicrobenchmark HPCworkloadapplication ProgrammingComplexity What‑IfAnalysis MorefrequentrNPFs 18
19.
ExperimentalEnvironment Ethernet TwoDellPowerEdgeR210IIRackServers 8GB1333MHzmemory One4‑coreIntelXeonE3‑1220CPUat3.10GHzrunningUbuntu13.10witha Linux3.11.4 NPF‑supporting12GbpsprototypeNIC InfiniBand EightHPProLiantDL380pGen8servers 128GBmemory One12‑coredualsocketXeonE5‑2697v2CPUat2.7GHzrunningRedHat7.0 withLinux3.10 NPF‑supporting56GbpsConnect‑IBNICs 19
20.
MemoryUtilizationofKVS Key‑valuestorewithEthernet OnememcachedinstanceinoneVM Memslapbenchmark(90%get,10%set,1KBvaluesbydefault) Itispossibletorun4VMswithNPFswhileupto2VMswith pinning 20
21.
MemoryUtilizationofKVS WithNPFs,youcanrunfourVMsinproductivemanner 21
22.
MemoryUtilizationofKVSwithDynamic Reallocation Changethesizeofworkingsetafter50seconds InoneVM,100MBto900MBwhileinotherVM,900MBto 100MB 22
23.
MemoryutilizationofKVSwithDynamic Reallocation Withpinning,oneofthetwoinstancesalwayssuffer Staticallyallocatedmemory(500MB)istoosmalltoholdits workingset(900MB) OverallperformanceofNPFsisbetterthanpinning 23
24.
MemoryUtilizationofStorageSystem AMachineemploystgt,aniSCSItargetimplementationthat supportsiSCSIextensionsforRDMA(iSER) Evaluatebandwidthbymeasuringtherandom512KBread bandwidthusingfiobenchmark Comparetwosettings:pinningandNPFs 24
25.
MemoryUtilizationofStorageSystem Inthepinnedconfiguration,thestaticmemoryallocationleavesa smalleramountofmemorytothepagecache,whichcausesmore cachemisses WhatIthinkaboutisthatthebandwidthdiffersdependingonwhen themeasurementisconducted 25
26.
OverheadofPinningandCopying(1) MicrobenchmarksinIntelMPIbenchmarkssuits ComparebetweenthreeconfigurationsasMPIcommunication backend copying,pinning(astate‑of‑the‑artheuristicpin‑downcache) andNPFs 26
27.
OverheadofPinningandCopying(2) BeffbenchmarkwhichmodelsrealHPCworkloads NPFprovidessamebenefitaspinningwithouttheneedtopin downmemory 27
28.
ProgrammingComplexity PortingthetgtstoragedemonwithNPFrequireonly40LOC(lines ofcodes) IntheMPImiddlewarelibrary,NPFenableseliminationof thousandsofLOC 28
29.
What‑ifAnalysis SyntheticsituationwithmorefrequentrNPFs MeasurebandwidthofsendoperationswithrNPFsatavariable specifiedfrequency BackupringsignificantlyimprovesperformancewithEthernet IthinktheabsolutevalueoffrequencyshowsrNPFsdegrades performancegreatly 29
30.
Conclusion ThispaperproposedNPFssupportforInfiniBandandEthernetNIC InordertocopewithrNPFs InfiniBandsupportreliesonRCprotocolwithRNRNACKand retransmission Ethernetsupportreliesonbackupring TheevaluationsshowthattheirNPFssupport Improvestheutilizationofmemory Increasesoverallperformance Simplifiestheprogrammodel 30
31.
MyThoughts Thispaperisagoodexampleofwell‑writtenpapersinsystemsoftware Clearlyexplainsmotivationandrequirementwithmanycitations fromawideperspective Describebothoverallpicturesandthetechnicaldetailsseparetely intheimplementationpart Thegoodpointisthatitexplainsnotonlytheimplementation itselfbutalsowhyoneway(orsimplesolution)doesnotwork Plentifulevaluationassumingmanydifferentsituations 31
32.
FromthePerspectiveofmyResearch Provideproofthatproposedmethodworksinordinaryconditions butnotinextremeconditions TheymainlyevaluatesituationswithrarerNPFssuchasget‑ intensiveKVSandstorage Atleast,what‑ifanalysisreportsfrequentrNPFcandegrade performancegreatly ProvidenoevaluationoflatencyofrNPFs IwanttoseetheoverheadofrNPFsincludingretransmission Iwanttoknowmoreimformationaboutwhatmethodswereused withtheevalutionofRDMA Isuspectmostaresendoperations Howaboutone‑sidedoperationssuchasRDMA READ/WRITE? 32
33.
ThingsmentionedaboutRDMAOperation ButinsomecasesRCdoesnotpermitRNRNACKsforRDMA. Whenaninitiatorofaremotereadrequestencountersapagefault, RCprovidesnowayforittoasktherespondertostop.Theonly waytogetthesendertoretransmitisbyaskingittorewind,after therNPFisresolved.Untilthen,wemustdropallincoming packets. Iamnotsurewhat"insomecases"specificallymeans OnethingIgotfromthesesentencesisthattheimplementationof NPFsforRDMAoperationsismoredifficult ItisnowonderthattheusageofRDMAoperationsinOn‑demand Pagingcausesunexpectedoverhead 33
Download now