$ which
• NSDI 2023
• Boston, MA, USA, April 17-19, 2023
• https://www.usenix.org/conference/nsdi23/technical-sessions
• '23: 96/560 papers, acceptance rate: 17%
• '22: 78/396 papers, acceptance rate: 19.7%
• o
ffl
ineͷΈʂʢڈॳͷhybrid։࠵ʣ
• dual-trackܧଓ
Slide 4
Slide 4 text
Awards
• Best Paper
• LeakyScatter: A Frequency-Agile Directional Backscatter Network Above 100 GHz
• CausalSim: A Causal Framework for Unbiased Trace-Driven Simulation
• DOTE: Rethinking (Predictive) WAN Tra
ffi
c Engineering
• Community Award
• Building Flexible, Low-Cost Wireless Access Networks With Magma
Slide 5
Slide 5 text
NSDI ’23 Technical Sessions
• 2023/04/17
• RDMA
• Learning with GPUs
• RPC and Remote Memory
• Congestion Control
• Distributed Systems
• Wireless
• Cloud
• Internet-Scale Network
• 2023/05/19
• Programming the Network
• Alternative Networks
• Performance
• Serverless and Network Functions
• Real Networks
• Cellular
• Testing Physical Layer
• 2023/04/18
• Synthesis and Formal Methods
• Data Centers
• Systems for Learning
• Privacy and Security
• Video
• Data
• Making Systems Learn
• IoT Networks
23 tracks, 96sessions
Slide 6
Slide 6 text
ࢀߟɿNSDI ’22 Technical Sessions
• 2022/04/04
• Cluster Resource Management
• Transport Layer - Part 1
• Video Streaming
• Programmable Switches - Part 1
• Security and Privacy
• Network Troubleshooting and
Debugging
• Operational Track - Part 1
• Wireless - Part 1
• 2022/04/06
• Operational Track - Part 2
• Edge IoT Applications
• Cloud Scale Services
• ISPs and CDNs
• Cloud Scale Resource Management
• Data Center Network Infrastructure
• Multi-tenancy
• Software Switching and Beyond
• 2022/04/05
• Reliable Distributed Systems
• Raising the Bar for Programmable
Hardware
• Testing and Veri
fi
cation
• Programmable Switches - Part 2
• Sketch-based Telemetry
• Transport Layer - Part 2
• Troubleshooting
• Wireless - Part 2
24 tracks, 78sessions
Slide 7
Slide 7 text
ࢀߟɿNSDI '19 Technical Sessions
• 2019/02/26
• Host Networking
• Distributed Systems
• Modern Network Hardware
• Analytics
• Data Center Network Architecture
• 2019/02/28
• Network Characterization
• Privacy and Security
• Network Modeling
• Wireless Applications
• 2019/02/27
• Wireless Technologies
• Operating Systems
• Monitoring and Diagnosis
• Improving Machine Learning
• Network Functions
• Wireless Applications
15 tracks, 50sessions
·ͱΊΔํ
• ҙ
• ʢࢲͷʣڵຯ͕͋ͬͨͷ͚ͩհ
• ʢࢲͷʣཧղͰ͖ͨͷ͚ͩհ
• ͪΌΜͱઆ໌͢Δͷ͕͍͠ͷͨͪ: NIC queue, Distributed system, AI/
DL, Semantics, Veri
fi
cation, Compiler, Wireless, Edge/IoT
• ͭ·Γɺ͍ͭͷʢࢲͷʣج४
Slide 10
Slide 10 text
Ξϯέʔτ
• հͨ͠ͷͷͳ͔ͰɺڵຯΛͻ͔ΕΔͷΛ3ͭબΜͰ͍ͩ͘͞ɻ
Slide 11
Slide 11 text
Day 1
Slide 12
Slide 12 text
RDMA
Slide 13
Slide 13 text
SRNIC: A Scalable Architecture for RDMA NICs
Hong Kong University of Science and Technology, ByteDance, Unaf
fi
liated
• scalable RDMA NICΞʔΩςΫνϟ: SRNICͰεέʔϥϏϦςΟվળ
• FPGAͰϓϩτλΠϓ࣮
• QPs (Q Pairs)͕10kͰ҆ఆ
• PFC free
Slide 14
Slide 14 text
Hostping: Diagnosing Intra-host Network Bottlenecks in RDMA Servers
BUPT, Purple Mountain Laboratories, ByteDance Inc.
• GPU w/ RDMAͰ100G~ʹͳΔͱϗετNW͕ϘτϧωοΫ
• Hostping: RNICͱϗετEPͰϧʔϓόοΫςετͰԆͱଳҬΛஅɾੳ
• طଘҎ֎ʹ৽ͨʹ6ͭϘτϧωοΫΛൃݟ
Intra-host
Inter-host
(Miss con
fi
g.)
Slide 15
Slide 15 text
Understanding RDMA Microarchitecture Resources for
Performance Isolation
Duke University, Microsoft, Shanghai Jiao Tong University
• RDMAΛVM͝ͱʹੑೳisolation͍ͨ͠
• RNICੑೳͰ͖ΔϚΠΫϩΞʔΩςΫνϟݱঢ়ଘࡏͤͣɻ
• NVIDIA, Chelsio, Intelʹڞ༗ࡁΈɻ
Slide 16
Slide 16 text
Empowering Azure Storage with RDMA
Microsoft
• AzureϦʔδϣϯͰRDMAετϨʔδΛαϙʔτ࢝͠Ίͨ
• RDMAΛVM (HV), Storage྆ํͰ༗ޮԽɻregionDCؒͰ͏
• NICͰDCQCN, sK-RDMAϓϩτίϧɺNWͰPFC/SONiC/SAI
• RDMA over commodity Ethernet v2Λ͍ɺطଘΠϯϑϥΛ͏
• 70%RDMAτϥϑΟοΫ
Slide 17
Slide 17 text
Learning with GPUs
Slide 18
Slide 18 text
Zeus: Understanding and Optimizing GPU Energy Consumption of DNN Training
University of Michigan
• ֶश͕ྃ࣌ؒओ؟ɺΤωϧΪʔޮฦ͠
• ΤωϧΪʔফඅྔͱτϨʔχϯά࣌ؒͷτϨʔ
υΦϑΛ໌Β͔ʹͨ͠
Slide 19
Slide 19 text
RPC and Remote Memory
Slide 20
Slide 20 text
Remote Procedure Call as a Managed System Service
DukeUniversity, University of Washington, Shanghai Jiao Tong University
• RPCΛ֤ΞϓϦͰ࣮͢ΔͷඇޮͳͷͰɺαʔϏεԽʢσʔϞϯԽʁʣͨ͠
• mRPC: αΠυΧʔൺֱͰ2.5ഒɻॊೈੑ૿͢
Slide 21
Slide 21 text
Congestion Control
Slide 22
Slide 22 text
Bolt: Sub-RTT Congestion Control for Ultra-Low Latency
Stanford University, Google LLC
• 200G, 400G࣌ͷ੍ޚɻBDPʹऩ·Βͳ͍
• SRCʢαϒRTT੍ޚʣͰૣ͘ʹؾͮ͘ɺProactive Ramp UpͰϑϩʔিಥΛ༧ݟͯ͠ػΛૉૣ͘
༗͢Δ
• Swift, HPCCൺͰ99%ileͷͪ࣌ؒΛ88%ॖɺFCTΛ3ഒվળ
Slide 23
Slide 23 text
Understanding the impact of host networking elements on traf
fi
c bursts
Johns Hopkins University, Meta
• eBPFͰτϥϑΟοΫॲཧͷՄࢹԽ
• όʔετɺ੍ޚɺqdisc, sched. NIC-sched. HW-o
ffl
oad, protocol
• [ns]͔Β[s]Φʔμʔ·ͰݟΕΔ
Slide 24
Slide 24 text
Distributed Systems
Slide 25
Slide 25 text
DiSh: Dynamic Shell-Script Distribution
MIT, University of Pennsylvania, Purdue University, Brown University
• DISH:
• γΣϧεΫϦϓτͰࢄίϯϐϡʔςΟϯά͠Α͏ͥʂ
• BashϕʔεͰɺࣗಈฒྻγεςϜར༻(PASH)ɺHDFS/
Hadoop Streamingར༻
Slide 26
Slide 26 text
Wireless
• Skip
Slide 27
Slide 27 text
Cloud
Slide 28
Slide 28 text
SkyPilot: An Intercloud Broker for Sky Computing
University of California, Berkeley, UC Berkeley and ICSI
• Sky of Computing = Inter cloud broker
• ϫʔΫϩʔυ͝ͱʹҧ͏public cloudΛ͍͚Δ͜ͱͰɺίετϝϦοτʢ࣌ؒɺՁ֨ʣΛग़͢
• cf: https://misreading.chat/2023/04/25/112-skypilot-an-intercloud-broker-for-sky-computing/
Slide 29
Slide 29 text
Invisinets: Removing Networking from Cloud Networks
UC Berkeley, Google, Microsoft
• ΫϥυωοτϫʔΫར༻͢Δͷେม͗͢Δ
• ςφϯτNWΛநԽͨ͠APIͷఏڙ
• PRDO: Publicly Routable but Default O
ff
• routingग़དྷΔ͕ɺσϑΥϧτdeny
• શΤϯυϙΠϯτʹIPv6༩
• ෳࡶ͞ͷ90%ΛݮͰ͖ͨ
• Cf: https://misreading.chat/2023/05/18/114-invisinets-removing-networking-from-cloud-networks/
Slide 30
Slide 30 text
Internet-Scale Networks
Slide 31
Slide 31 text
xBGP: Faster Innovation in Routing Protocols
ICTEAM, UCLouvain, I
IJ
/Arrcus, Inc, NSG, ETH Zürich
• BGPͷػೳՃ͍ɺ͕ɺૣ͍͍ͨ͘
• ϕϯμʔχϡʔτϥϧͳAPIͱBGP࣮ͷ֦ு෦ΛeBPFͰఆٛɾ࣮
• FRR/BIRDͰ࣮
• Use case 7ͭհ: withdrawࣦഊ࣌ʹTSͰϧʔτഁغػೳɻϧʔτબํ๏ͷࢹͱڞ༗ɻൖ࣌ؒͷଌఆɻetc...
• Cf: https://blog.apnic.net/2021/01/27/xbgp-toward-a-fully-extensible-bgp/
873k route@IPv4
120k route@IPv6
Slide 32
Slide 32 text
Ҏ߱ޙͰ
Slide 33
Slide 33 text
Day 2
Slide 34
Slide 34 text
Synthesis and Formal Methods
• Skip
Slide 35
Slide 35 text
Data Centers
Slide 36
Slide 36 text
Flattened Clos: Designing High-performance Deadlock-free
Expander Data Center Networks Using Graph Contraction
Shanghai Jiao Tong University, Chinese Academy of Sciences
• FC: Flattened ClosߏͷఏҊ
• ToRΛཧతʹkݸʹ͚ɺྡԾ
Up-down pathΛ࡞Γɺ
fl
attenedͤ͞Δ
• CBD-free routing
Slide 37
Slide 37 text
Systems for Learning
Slide 38
Slide 38 text
TOPOOPT: Co-optimizing Network Topology and Parallelization
Strategy for Distributed Training Jobs
Massachusetts Institute of Technology, Meta, CMU, Telescent
• TOPOOPTτϙϩδͰ100G RDMAΛͬͯDNNֶश
• Direct connect NW w/ ޫεΠον + ύονύωϧ + NPAR
• Fat-TreeൺͰ3ഒ͘ɺ҆Ձ@12node
ֶशதͷ௨৴ύλʔϯ
Slide 39
Slide 39 text
Privacy and Security
• Skip
Slide 40
Slide 40 text
Video
• Skip
Slide 41
Slide 41 text
Data
• Skip
Slide 42
Slide 42 text
Making Systems Learn
• Skip
Slide 43
Slide 43 text
IoT Networks
• Skip
Slide 44
Slide 44 text
Day 3
Slide 45
Slide 45 text
Programming the Network
Slide 46
Slide 46 text
A High-Speed Stateful Packet Processing Approach for Tbps
Programmable Switches
KTH Royal Institute of Technology, Roma Tre University, UCLouvain
• RDMAసૹ࣌ɺstateNFʹɾసૹ͢Δ
• ͜ΕΛP4ͰΔ
• 300GbpsΛୡ
Slide 47
Slide 47 text
ExoPlane: An Operating System for On-Rack Switch Resource
Augmentation
Microsoft, University of Texas at Austin, Carnegie Mellon University
• In-network computing on Rack
• ToR (P4)ͱSmartNICΛͬͯɺINCΛ࣮ݱɻಛʹstateཧΛ࿈ಈͯ͠Δ
Slide 48
Slide 48 text
RingLeader: Ef
fi
ciently Of
fl
oading Intra-Server Orchestration to NICs
Google, UT Austin
• αʔόΦʔέετϨʔγϣϯʢsked.?ʣΛNIC assisted CPU sked.ͱ͢Δ
• FPGAͰ࣮͠ɺtail-latency, throughput, CPU༻Λվળ
Slide 49
Slide 49 text
Alternative Networks
• Skip
Slide 50
Slide 50 text
Performance
Slide 51
Slide 51 text
Skyplane: Optimizing Transfer Cost and Throughput Using Cloud-
Aware Overlays
University of California, Berkeley
• Inter cloudͰόϧΫσʔλసૹγεςϜ
• Ұ൪Ձ֨ޮ͕ྑ͍ํ๏Λݟ͚ͭΔʢSkyplane plannerʣ
• ઢܗܭը๏Ͱղ͘
• Ϋϥυ: ࠷େ4.6ഒ
• Ϋϥυؒ: ࠷େ5.0ഒ
Disaggregating Stateful Network Functions
Microsoft and AMD Pensando
• ൚༻ARMίΞͱASICʢߴstateful match/actionʣ
Λ༻͍ͯɺॲཧΛϗετ͔ΒΓ͠ɺNFΛࢄԽ
• 12NICϚγϯΛ࣮͠ɺNFੑೳ͕10ഒ্
• Azureͷ࣮ӡ༻݁Ռͷհ
Slide 55
Slide 55 text
Real Networks
• Skip
Slide 56
Slide 56 text
DOTE: Rethinking (Predictive) WAN Traf
fi
c Engineering
Hebrew University of Jerusalem, Microsoft Research, Technion
• Best paper !
• DOTE: աڈͷσʔλͷΈΛͬͯDL͠ɺWAN TE͢Δ
• Direct Optimization for Tra
ffi
c Engineering
• धཁ༧ଌʢNot IPFIXͰࡉ͔͘ੳ or Not demand-basedʣͰͳ͘࠷దԽ
• ֬࠷దԽ + ࣮ੈքରԠͷͨΊʹML/DL͏
• ܭࢉ࣌ؒૣ͘ɺ݁Ռྑ͍
• τϥϑΟοΫมԽোݎ࿚ੑྑ͍
Slide 57
Slide 57 text
Dashlet: Taming Swipe Uncertainty for Robust Short Video
Streaming
Princeton University
• εϫΠϓͷλΠϛϯάʹಛԽͨ͠ϏσΦετϦʔϛϯάख๏վળ
• videoϨίϝϯυͱ࿈ܞͨ͠όοϑΝϦϯάɺϏοτϨʔτվળͷ࣮
• ϏσΦ্࣭Λ֬ೝ
Slide 58
Slide 58 text
Cellular
• Skip
Slide 59
Slide 59 text
Testing
Slide 60
Slide 60 text
Norma: Towards Practical Network Load Testing
Nanjing University, Alibaba Group
• pktgenͰग़དྷͯͳ͍͜ͱ
• εςʔτϑϧ/ϦΞϧͳτϥϑΟοΫ
• TbpsͳଳҬͱϨʔτ੍ޚ
• Norma: Programmable SW ASIC (To
fi
no w/ P4 1kߦ*)Ͱ࡞ͬͨ
• 3TbpsͷTCP, 1TbpsͷHTTPτϥϑΟοΫΛੜ
+ SWجຊػೳͰ8kߦ