Upgrade to Pro
— share decks privately, control downloads, hide ads and more …
Speaker Deck
Features
Speaker Deck
PRO
Sign in
Sign up for free
Search
Search
クラウドデータセンターネットワークの “いま”と “これから”
Search
Masayuki Kobayashi
June 12, 2023
Technology
8
2.3k
クラウドデータセンターネットワークの “いま”と “これから”
Masayuki Kobayashi
June 12, 2023
Tweet
Share
More Decks by Masayuki Kobayashi
See All by Masayuki Kobayashi
GPUネットワーク設計・運用 基礎勉強会 Lossless Ethernet – PFC/ECN編
markunet
6
560
HPCネットワーク基礎(RDMA/Infiniband/RoCE編)
markunet
37
4.6k
Clos Network Topology 再入門
markunet
22
7.2k
RIFT A new routing protocol for IP fabrics
markunet
0
500
大規模DCのネットワークデザイン
markunet
3
1.3k
閉域網接続の技術入門
markunet
1
310
Other Decks in Technology
See All in Technology
小さなチーム 大きな仕事 - 個人開発でAIをフル活用する
himaratsu
0
140
実践アプリケーション設計 ③ドメイン駆動設計
recruitengineers
PRO
10
2.5k
浸透しなさいRFC 5322&7208
hinono
0
130
Understanding Go GC #coefl_go_jp
bengo4com
0
1.1k
「守る」から「進化させる」セキュリティへ ~AWS re:Inforce 2025参加報告~ / AWS re:Inforce 2025 Participation Report
yuj1osm
1
160
Go で言うところのアレは TypeScript で言うとコレ / Kyoto.なんか #7
susisu
7
2k
Oracle Base Database Service:サービス概要のご紹介
oracle4engineer
PRO
2
20k
Preferred Networks (PFN) とLLM Post-Training チームの紹介 / 第4回 関東Kaggler会 スポンサーセッション
pfn
PRO
1
270
Goでマークダウンの独自記法を実装する
lag129
0
230
広島銀行におけるAWS活用の取り組みについて
masakimori
0
160
RAID6 を楔形文字で組んで現代人を怖がらせましょう(実装編)
mimifuwa
1
320
Figma + Storybook + PlaywrightのMCPを使ったフロントエンド開発
yug1224
10
3.3k
Featured
See All Featured
GraphQLとの向き合い方2022年版
quramy
49
14k
How STYLIGHT went responsive
nonsquared
100
5.8k
Let's Do A Bunch of Simple Stuff to Make Websites Faster
chriscoyier
507
140k
How to Ace a Technical Interview
jacobian
279
23k
VelocityConf: Rendering Performance Case Studies
addyosmani
332
24k
No one is an island. Learnings from fostering a developers community.
thoeni
21
3.4k
Code Reviewing Like a Champion
maltzj
525
40k
I Don’t Have Time: Getting Over the Fear to Launch Your Podcast
jcasabona
33
2.4k
Navigating Team Friction
lara
189
15k
10 Git Anti Patterns You Should be Aware of
lemiorhan
PRO
656
61k
Performance Is Good for Brains [We Love Speed 2024]
tammyeverts
10
1k
Understanding Cognitive Biases in Performance Measurement
bluesmoon
29
1.9k
Transcript
ΫϥυσʔληϯλʔωοτϫʔΫͷ l͍·zͱ l͜Ε͔Βz "*.-)1$ωοτϫʔΫՊձ .BTBZVLJ,PCBZBTIJ
ωοτϫʔΫΛਐԽͤ͞ΔΩʔϑΝΫλʔԿ͔ $PNQVUJOH*OOPWBUJPO 4UPSBHF*OOPWBUJPO /FUXPSLJOH*OOPWBUJPO • ࢄػցֶश • "*$IJQͷߴԽ ίϯϐϡʔςΟϯάͱετϨʔδͷٕज़ֵ৽͕ωοτϫʔΫʹਐԽѹΛ͔͚Δ •
ࢄετϨʔδ • /7.F 4$. 1.&. • +PC$PNQMFUJPO5JNFݮ • Ϟσϧ։ൃͰͷϨΠςϯγݮ
͜Ε·ͰͷίϯϐϡʔςΟϯάͱσʔληϯλʔωοτϫʔΫ ϞϊϦε͔ΒϚΠΫϩαʔϏε ϕΞϝλϧ͔Β7.ɺͦͯ͠ίϯςφ ίϯϐϡʔςΟϯάͷਐԽγεςϜϦιʔεͷࢄஔΛՄೳʹͨ͠ɻ ωοτϫʔΫͷੑೳ͕γεςϜશମͷύϑΥʔϚϯεʹେ͖͘Өڹ͢Δɻ
͜Ε·ͰͷίϯϐϡʔςΟϯάͱσʔληϯλʔωοτϫʔΫ $MPVE/BUJWF%$/FUXPSL 8FC4DBMF యܕతͳ l'BCSJDz l$MPTzͱݺΕΔϨΨγʔτϙϩδ ίϞσΟςΟ&UIFSOFUσόΠεʹΑΔεέʔϧΞτ *1ϕʔεͷύεࢄʹΑΔ#JTFDUJPO #BOEXJEUIར༻ ଟ༷ͳϑϩʔͷࠞࡏ
$PNQVUJOH 4UPSBHF FUD $16r$FOUSJD"SDIJUFDUVSF
͍·զʑΛऔΓר͘ঢ়گͱ՝ େنݴޠϞσϧ ੜ"* ࣗવݴޠॲཧ 5SBOTGPSNFSϞσϧ ࢄػցֶश ͖ଇʹΑΔೖϞσϧͷେنԽ ˠ ઐ༻ΞΫηϥϨʔλ
(16 ͱΫϥελؒ௨৴ͷ֬อ ޫి༥߹ ΦʔϧϑΥτχΫε Ԇͱফඅిྗͷݮ ߴେ༰ྔ௨৴ ˠ ಋମ։ൃ ઐ֎ͷͨΊຊࢿྉͷείʔϓ֎
$16 ͜Ε·ͰͷίϯϐϡʔςΟϯάͱσʔληϯλʔ جຊతʹ͕ͯ͢ $16Ͱ࣮ߦ͞Ε͍ͯͨ )ZQFSWJTPST *01SPDFTT /FUXPSL4UBDL ςφϯτίʔυ ΞϓϦέʔγϣϯ /*$
"QQ LT *0 )7 /FUXPSL େྔͷσʔλసૹΛ͏ϫʔΫϩʔυͳͲͰɺ $16͕ϘτϧωοΫʹͳΔ ςφϯτॲཧʹׯব $16ϦιʔεՄೳͳݶΓ"QQ ςφϯτ Ͱར༻͍ͨ͠ 1$*F
%16*16 ͜Ε͔ΒͷίϯϐϡʔςΟϯάͱσʔληϯλʔ %JTUSJCVUFE%JTBHHSFHBUFE$PNQVUJOH $16 /*$ "QQ LT *0 )7 /FUXPSL
1$*F $16 /*$ "QQ LT *0 )7 /FUXPSL 1$*F (1(16 "*.- 1$*F 1$ *F
͜Ε͔ΒͷίϯϐϡʔςΟϯάͱσʔληϯλʔ %JTUSJCVUFE%JTBHHSFHBUFE$PNQVUJOH $16 /*$ "QQ LT *0 )7 /FUXPSL 1$*F
(1(16 "*.- 1$*F %16*16 $16 /*$ "QQ LT *0 )7 /FUXPSL 1$*F 4UPSBHF /7.F 1$*F 8FC4DBMF'BCSJD 'SPOUFOE/FUXPSL 4DIFEVMFE'BCSJD "*.- #BDLFOE/FUXPSL $16 /*$ "QQ LT *0 )7 /FUXPSL 1$*F $16 /*$ "QQ LT *0 )7 /FUXPSL 1$*F 1$*F 1$*F
͜Ε͔ΒͷίϯϐϡʔςΟϯάͱσʔληϯλʔ 3%."#BTFE*OUFSDPOFDU $16 /*$ "QQ LT *0 )7 /FUXPSL 1$*F
(1(16 "*.- 1$*F %16*16 $16 /*$ "QQ LT *0 )7 /FUXPSL 1$*F 4UPSBHF /7.F 1$*F 4DIFEVMFE'BCSJD "*.- #BDLFOE/FUXPSL $16 /*$ "QQ LT *0 )7 /FUXPSL 1$*F $16 /*$ "QQ LT *0 )7 /FUXPSL 1$*F 1$*F 1$*F 8FC4DBMF'BCSJD 'SPOUFOE/FUXPSL
͜Ε͔ΒͷίϯϐϡʔςΟϯάͱσʔληϯλʔ $PNQVUJOH 4UPSBHF 3%."/FUXPSL
͜Ε͔ΒͷίϯϐϡʔςΟϯάͱσʔληϯλʔ )PU$IJQT5FTMBجௐߨԋ l#FZPOE$PNQVUF&OBCMJOH"*5ISPVHI4ZTUFN*OUFHSBUJPOzΑΓ
͜Ε͔ΒͷίϯϐϡʔςΟϯάͱσʔληϯλʔ l'VUVSFEBUBDFOUFSTPGBMMLJOETXJMMCFCVJMUMJLFIJHIQFSGPSNBODFDPNQVUFST z /WJEJB$&0 +FOTFO)VBOH
͜Ε͔ΒͷίϯϐϡʔςΟϯάͱσʔληϯλʔ
֤ϫʔΫϩʔυ͕ωοτϫʔΫʹٻΊΔ͜ͱ )1$ ԆͷӨڹΛඇৗʹड͚͍͢ΞϓϦέʔγϣϯͰδολճආ͍ͨ͠ ԆΛ࣮ݱ͢ΔͨΊʹɺେྔͷύέοτΛΩϡʔʹೖΕΔ͜ͱආ͚Δ %$2$/ʹ͓͚ΔύέοτϚʔΩϯάʹΑΔΞάϨογϒίϯτϩʔϧඞਢ "*.- )1$ͱରۃʹ͋Δػցֶशͷ(16ΞϓϦέʔγϣϯ ϨΠςϯγ͕·͍͕͠ɺͦΕΑΓ༻Մೳͳ࠷େεϧʔϓοτ͕ཁٻ͞ΕΔ %$2$/ͷύέοτϚʔΩϯάϥϯμϜͰঃʑʹ૿Ճͤ͞Δ
8FC4DBMF/FUXPSLͱ "*.-/FUXPSLͷҧ͍ 8FC4DBMF/FUXPSL ଟ༷ͰෆۉҰͳଟͷϑϩʔ͕ࠞࡏ͠ɺݸผͷϑϩʔͷτϥϑΟοΫখ͍͞ ΄ͱΜͲͷϑϩϯτΤϯυ௨৴5$1ʹґଘͰ͖ΔͨΊύέοτϩεʹඇৗʹ༰ "*.-/FUXPSL ଟͷδϣϒ͕ಉظ͠ͳ͕Βू߹ԋࢉΛߦ͍ɺߴଳҬ෯ͷ௨৴͕ඞཁʹͳΔ ͯ͢ͷδϣϒ͕࠷͍δϣϒʹΑͬͯ͞ΕΔ 3%."ΧʔωϧόΠύεͷͨΊ5$1ʹґଘͰ͖ͣɺύέοτϩεʹහײ "*.-Ͱඇৗʹ৴པੑͷߴ͍ύέοτసૹΛߦ͏ωοτϫʔΫΛ֬อ͢Δ͕ඞཁ͋Δ
8FC4DBMF/FUXPSLͷ՝ $PNQVUF 4UPSBHF "*.- /POTDIFEVMFE'BCSJD "*.-ϫʔΫϩʔυʹ࠷దԽ͞Ε͍ͯͳ͍ ϑϩʔΛ۠ผ͠ͳ͍ͨΊ3%."ΛࢧԉͰ͖ͳ͍ ड৴Ωϡʔͷঢ়ଶΛݩʹͨ͠ϑϩʔࢄ͕ߦΘΕͳ͍ͨΊಛఆͷΩϡʔ͕ڝ߹͢Δ ϚϧνςφϯτΛఏڙ͢ΔΫϥυඇৗʹl-PTTZzͳωοτϫʔΫ
8FC4DBMF/FUXPSLͷ՝ 3%."ͱύέοτଛࣦͷ՝ ௨ৗͷ*1ωοτϫʔΫͰ͋ΕύέοτυϩοϓΛ5$1ͷΈͳͲͰճ෮ͤ͞Δ 3%."ͷΩϡʔͰͦΕෆՄೳͰ͋ΓɺύέοτଛࣦʹΑΔੑೳԼ͕ஶ͍͠ • 3%."ͷ࠶ૹϋʔυΣΞ࣮ • (P#BDL/͕ύϑΥʔϚϯεͷԼΛট͘ • 3%."ϩεϨεωοτϫʔΫΛલఏͱͯ͠ߟҊ͞Εͨख๏
#VJMEJOHB'VUVSF1SPPG$MPVE*OGSBTUSVDUVSF"6OJGJFE"SDIJUFDUVSFGPS/FUXPSL 4FDVSJUZ BOE4UPSBHF4FSWJDFT
"*.-ϫʔΫϩʔυʹඞཁͳωοτϫʔΫ *OUFSDPOOFDUJPO/FUXPSL #BDLFOE/FUXPSL 'VMM#JTFDUJPO #BOEXJEUI -PTT-FTT'MPX$POUSPM %ZOBNJD-PBE#BMBODJOH ಠཱͨ͠Ϋϥελؒ௨৴༻ͷཪ໘ωοτϫʔΫ (16ؒͷಉظ "MMUP"MMͳͲͷू߹ԋࢉ
"*.- 8FC
ݱࡏ࠾༻͍ͯ͠Δ (16*OUFSDPOOFDU/FUXPSL5PQPMPHZ 3BJM0QUJNJ[FE5PQPMPHZ ಉ͡)$"/*$ಉ͡-FBGTXJUDIʹଓ ௨ৗ࠷ύεͰͷ௨৴͕Մೳɺঢ়گʹԠͯ͡ΞϓϦέʔγϣϯ͕࠷దͳ/*$Λબ Leaf Rail 1 Leaf Rail
2 Leaf Rail 3 Leaf Rail 4 Leaf Rail 5 Leaf Rail 6 Leaf Rail 7 Leaf Rail 8 GPU Server GPU Server GPU Server GPU Server
ϥοΫσβΠϯͷมߋ 5P3 5PQPG3BDL ͔Β &P3 &OEPG3PX ిݯͷ੍ͰεΠονωοτϫʔΫϥοΫʹू͢Δඞཁ͕͋Δ 5P3 TXJUDIͷΘΓʹύονύωϧΛஔ͢Δઃܭ
αʔόϥοΫ(16ͱ$PPMJOH&RVJQNFOUͰ༗ GPU Server Patch Panel GPU Server GPU Server Patch Panel GPU Server GPU Server Patch Panel GPU Server GPU Server Patch Panel GPU Server NW Switch Patch Panel NW Switch NW Switch Patch Panel NW Switch
*OUFSDPOOFDUJPO/FUXPSL )1$501 ʹ͓͚Δ*OUFSDPOOFDU'BNJMZ4IBSF *OGJOJCBOE &UIFSOFU IUUQTEPDTOWJEJBDPNEHYEHYIVTFSHVJEFJOUSPEVDUJPOUPEHYIIUNM
&UIFSOFUWT*OGJOJCBOE IUUQTQDOBOPHPSHTUBUJDQVCMJTIFENFFUJOHT/"/0(@$BSEPOB@5PXBSET@)ZQFSTDBMF@)JHI@WQEG
"*.-ʹ͓͚Δ*OUFSDPOOFDUͷબج४ʢҰྫʣ *OGJOJCBOE ͱʹ͔͘ԆͰͷੑೳΛٻΊΔ શʹดͨ͡ΫϥελڥͰͳ͍ 'BUUSFFͰͳ͍τϙϩδΛ࠾༻͢Δ %SBHPOGMZͳͲ 3P$&W طଘͷࢿ࢈Λ׆༻͠ɺઐ༻ϋʔυΣΞͷೋॏࢿΛճආ͍ͨ͠ ΫϥυڥͰͷϚϧνςφϯγʔ͕ཁ݅ʹ͋Δ ωοτϫʔΫͦͷͷΛεέʔϧΞτ͍ͤͨ͞
*OGJOJCBOE 3P$& ٕज़ղઆ *OGJOJCBOE3P$&ษڧձͰղઆࡁΈ աڈࢿྉΛࢀর
ϩεϨεΠʔαωοτͰඞཁͳνϡʔχϯά %$2$/ &$/ͷύϥϝʔλௐ ϓϥΠϕʔτΫϥυڥͰϢʔβ ςφϯτ ͷϫʔΫϩʔυʹ߹Θͤͯௐ͢Δ ߴͳࣝͱܦݧ͕ඞཁ ,NJO ,NBY 1NBY
,NJO 1NBY )1$8PSLMPBET (168PSLMPBET (FOFSBM8PSLMPBET ,NBY
'SPOUFOEͱ #BDLFOEͷωοτϫʔΫڞ௨Ͱྑ͍ͷ͔ʁ 3%."ωοτϫʔΫͱඇ3%."ωοτϫʔΫ໌֬ʹ͢Δ͖ ߏஙͱཧίετ্͕Δ͕τϨʔυΦϑ͕େ͖͍ ڞ௨ͷωοτϫʔΫͰेͳΩϡʔΛ֬อͰ͖ͳ͍߹͕͋Δ ඞཁͳͷ'Ϛγϯͱͦͷઐ༻ίʔεͰ͋ͬͯɺҰൠंͱߴಓ࿏Ͱͳ͍ "*.- 8FC IUUQTXXXVTFOJYPSHDPOGFSFODFOTEJQSFTFOUBUJPOCBJ
࣮ࡍͷ3P$&WϩεϨενϡʔχϯά ඇެ։
&UIFSOFU4XJUDIJOH"4*$ͷಈ • 5CQT 'MPXMFU %-#ػೳαϙʔτ ͷDIJQ͕ଓʑొ • (16(*OUFSDPOOFDUͷੈք ඇެ։
((CQTͷ՝ ඇެ։
ϨΠςϯγͱεέʔϧΞττϨʔυΦϑ ωοτϫʔΫͷαΠζͱԆͷτϨʔυΦϑ͕ଘࡏ͢Δ $MPT5PQPMPHZ *OEJSFDU5PQPMPHZ ͷආ͚ΒΕͳ͍՝ Ѽઌϊʔυ·Ͱͷϗοϓ͕૿Ճ͢ΕϨΠςϯγ૿͑Δ ͕ɺ΄ͱΜͲͷϢʔεέʔεͰʹͳΒͳ͍ ඇৗʹେنͳ)1$(16ڥͰɺ͜ΕΒʹηϯγςΟϒͳέʔεʹͲ͏ରԠ͢Δ͔
$MPTͰͳ͍τϙϩδͷݕ౼ https://pc.nanog.org/static/published/meetings/NANOG74/1763/20181003_Martin_Routing_In_Dense_v1.pdf
%SBHPOGMZ 5PQPMPHZ IUUQTEBUBUSBDLFSJFUGPSHNFFUJOHNBUFSJBMTTMJEFTSUHXHSPVUJOHJOESBHPOGMZUPQPMPHJFT
%SBHPOGMZ 5PQPMPHZ IUUQTJFFFYQMPSFJFFFPSHEPDVNFOU MMFBGSPVUFST QIPTUTQFSMFBGSPVUFST T ITQJOFSPVUFST LSPVUFSSBEJY /HSPVQ OVNCFSPGIPTUTJOUIFHSPVQ
QMTI QIL /HSPVQ QM L L
%SBHPOGMZ 5PQPMPHZݕূڥ ඇެ։
Կ͕Γ͍ͯͳ͍ͷ͔ʁ ͍·զʑʹඞཁͳςΫϊϩδʔ "EBQUJWF3PVUJOH &$.1ͷภΓΛແ͘͢%-#ͷػೳ͚ͩͰෆे .JOJNBMͱ /PONJOJNBMΛಉ࣌ʹར༻͢Δࢄํࣜ 6("-ͷΑ͏ͳΞϓϩʔνΛ*13PVUJOHͰཱࣗࢄతʹߦ͍͍ͨ IUUQTJFFFYQMPSFJFFFPSHEPDVNFOU
"*.-ωοτϫʔΫઃܭɾӡ༻ͷ͠͞ • ిݯɾۭௐઃඋͷݶք ਫྫྷઃඋͷసʹ͍Δ • ωοτϫʔΫΛৗʹ࠷େߏͰઃܭ͢Δඞཁ͕͋Δ • ܭࢉίετͱΠϯϑϥߏஙͷܦࡁ߹ཧੑ • ߴੑೳڥ͚ͷߴͳϑϩʔ੍ޚΛ͢Δඞཁ͕͋Δ
• ϢʔβͷϫʔΫϩʔυ͝ͱͷಛੑΛཧղ͠ͳ͚ΕͳΒͳ͍ • ϋʔυΣΞϦιʔεͷநԽͱϚϧνςφϯγʔͷ͠͞ • ωοτϫʔΫͱίϯϐϡʔςΟϯάͷ໌֬ͳઢҾͰ͖ͳ͍
͜͜Ͱ͍ٞͨ͜͠ͱ • ʹΈͳ͞ΜͲ͏ͯ͠·͔͢ʁ