Tweet
Tweet

Slide 1

Slide 1 text

େن໛ػցֶश܇࿅ʹ ͓͚Δ*0ੑೳͷߴ଎Խ ۔୔࿨༸ ݐ෦मݟ ஜ೾େֶେֶӃγεςϜ৘ใ޻ֶݚڀՊ ஜ೾େֶܭࢉՊֶݚڀηϯλʔ ۔୔࿨༸ ݐ෦मݟେن໛ػցֶश܇࿅ʹ͓͚Δ*0ੑೳͷߴ଎Խ ৘ใॲཧֶ ձୈճ)1$ݚڀձใࠂ )1$ 7PM)1$ /P +VM !๺ւಓ๺ݟࢢຽձؗ

Slide 2

Slide 2 text

ൃද಺༰ ݚڀͷഎܠͱ໨త ਂ૚χϡʔϥϧωοτϫʔΫͷ֓ཁͱ܇࿅σʔλSFBEͷӨڹ SFBE*0Λߴ଎Խ͢ΔͨΊͷैདྷख๏ ఏҊख๏ ධՁ࣮ݧ ϕϯνϚʔΫධՁ ࣮ΞϓϦέʔγϣϯධՁ ؔ࿈ݚڀ ·ͱΊͱࠓޙͷ՝୊ 2

Slide 3

Slide 3 text

ݚڀͷഎܠʢ̍̎ʣ • ۙ೥ɼը૾෼ྨλεΫΛਂ૚χϡʔϥϧωοτϫʔΫʢҎ ԼDNNʣΛ༻͍ͯղ͘ख๏͕޿͘࠾༻͞Ε͍ͯΔ • ҰํͰɼը૾෼ྨͷϞσϧΛDNNͰ܇࿅͢Δࡍʹ͸ɼҰൠʹ਺ ສຕʢCifar100౳ʣʙ਺ඦສຕʢImageNetͳͲʣͷ܇࿅༻ͷը ૾Λread͢Δඞཁ͕͋Γɼread I/Oͷॴཁ࣌ؒ͸ແࢹͰ͖ͳ͍ ϘτϧωοΫͱͳΔ 3

Slide 4

Slide 4 text

ݚڀͷഎܠʢ̎̎ʣ $JGBS *NBHF/FUL ܇࿅ॲཧதͷ(16Χʔωϧ࣮ߦঢ়گ ܇࿅ॲཧதʢJUFSBUJPOʣͷ(16ར༻ঢ়گΛώʔτϚοϓͱͯ͠ՄࢹԽͨ͠ਤ ग़ల۔୔ɼݐ෦ɼਂ૚χϡʔϥϧωοτϫʔΫʹ͓͚Δ܇࿅ߴ଎ԽͷͨΊͷࣗಈ࠷దԽ ৘ใॲཧֶձୈճ)1$ݚڀձใࠂQਤ w ܇࿅σʔλͷSFBE͕཯଎͍ͯ͠Δέʔεͷ࣮ྫ 4

Slide 5

Slide 5 text

ݚڀͷ໨త w %//ʹ͓͚ΔSFBE*0ͷॴཁ࣌ؒΛ࡟ݮ͠ɼ܇࿅଎౓Λ վળ͢Δख๏ΛఏҊ͢Δ w ఏҊख๏ΛػցֶशϑϨʔϜϫʔΫͰ͋Δ$IBJOFSʹ࣮૷ ͠ɼ$IBJOFS͕ఏڙ͢ΔطଘػೳͱSFBE*0ͷੑೳൺֱΛ ߦ͏ 5

Slide 6

Slide 6 text

ैདྷख๏ʢ̍̎ʣ w ฒྻSFBE w ϛχόονόοϑΝϦϯά 6

Slide 7

Slide 7 text

ैདྷख๏ʢ̎̎ʣ w $IBJOFSʹ͓͍ͯ͸ฒྻSFBEͱϛχόονόοϑΝϦϯά ͷछྨΛ૊Έ߹Θͤͨػೳ͕ఏڙ͞Ε͍ͯΔ w ͔͠͠ɼ͜ͷΑ͏ͳ࢓૊ΈΛ༻͍ͯ΋ɼ࢖༻͢ΔετϨʔδͷ *0ੑೳʹΑͬͯ͸܇࿅σʔλͷSFBE͕(16ଆͷॲཧʹؒʹ߹ ΘͣɼશମΛ཯଎͢Δ 7

Slide 8

Slide 8 text

)1$Ϋϥελʹ͓͚Δ ετϨʔδ w Ұൠʹ)1$Ϋϥελͷܭࢉϊʔυʹ͸ϊʔυϩʔΧϧετϨʔδͱͯ͠/7.F44%౳ͷߴ଎ͳ ετϨʔδ͕౥ࡌ͞Ε͍ͯΔ͕ɼܭࢉδϣϒ͝ͱʹڞ༗ετϨʔδ͔ΒͷϑΝΠϧίϐʔ͕ඞཁ w ࢀߟͱͯ͠ɼ)1$Ϋϥελʮ$ZHOVTʯʹ͓͍ͯɼ*NBHF/FUͷը૾໿ສຕʢ߹ܭ(J#ʣ Λڞ༗ετϨʔδ͔ΒϊʔυϩʔΧϧετϨʔδʹσΟϨΫτϦ͝ͱίϐʔͨ͠ͱ͜Ζ ෼ʙ෼ఔ౓Λཁͨ͠ 8

Slide 9

Slide 9 text

ఏҊख๏ʢ̍̐ʣ ैདྷख๏ͱͷࠩ෼ w ڞ༗ετϨʔδ͔Β܇࿅σʔλΛฒྻʹϊʔυϩʔΧϧετ Ϩʔδ΁ϑΝΠϧίϐʔ͠ɼίϐʔ͞ΕͨϑΝΠϧ͔Βॱʹ SFBEͯ͠ϛχόονΛੜ੒͢Δ 9

Slide 10

Slide 10 text

ఏҊख๏ʢ̎̐ʣ w ఏҊख๏ʹΑΓɼϊʔυϩʔΧϧετϨʔδΛ࢖͏৔߹Ͱ΋܇ ࿅ॲཧ։࢝લͷϑΝΠϧίϐʔͷ׬ྃΛ଴ͨͣʹ܇࿅ॲཧ͕։ ࢝Ͱ͖Δ ैདྷख๏ ఏҊख๏ ϊʔυϩʔΧϧετϨʔδ΁ͷ ϑΝΠϧίϐʔ ϛχόονੜ੒ ܇࿅ॲཧ ܦա࣌ؒ 10

Slide 11

Slide 11 text

ఏҊख๏ʢ̏̐ʣ 11

Slide 12

Slide 12 text

ఏҊख๏ʢ̐̐ʣ JOEFYϦετͷྲྀΕ σʔλͷྲྀΕʢ࠶ܝʣ 12

Slide 13

Slide 13 text

ධՁ࣮ݧ؀ڥ w ຊݚڀͰͷධՁ͸ஜ೾େֶܭࢉՊֶηϯλʔʹͯӡ ༻͞Ε͍ͯΔ)1$Ϋϥελʮ$ZHOVTʯΛ࢖༻ͨ͠ $16 *OUFM9FPO(PME1SPDFTTPS $()[ Y .FNPSZ (J# (J#%%3&$$3%*..Y ϊʔυϩʔΧϧ ετϨʔδ *OUFM44%1$14FSJFT5# ڞ༗ετϨʔδ -VTUSF %%/&9"4DBMFS (16 /7*%*"5FTMB7(J#)#.1$*F Y 04 $FOU04-JOVYSFMFBTF $ZHOVT֎؍ 13

Slide 14

Slide 14 text

ධՁ࣮ݧ̍ʢ̍̎ʣ w ϝϞϦ্ʹϛχόονΩϡʔ͔ΒϛχόονΛ ݸ MPBE͢Δ·Ͱͷ࣌ؒΛܭଌ w ճܭଌͨ͠ฏۉ஋Λൺֱ͢Δ w ࢖༻͢Δσʔλ͸*NBHF/FUͷը૾໿ສຕΛ࢖༻͠ɼϛχ όοναΠζ͸ݻఆ ධՁύλʔϯ ࢖༻ετϨʔδ ख๏ ֤ॲཧͷ ੜ੒ϓϩηε਺ ϊʔυϩʔΧϧ ετϨʔδ 44% ैདྷख๏ ڞ༗ετϨʔδ -VTUSF ैདྷख๏ ఏҊख๏ -VTUSF44% ఏҊख๏ 14

Slide 15

Slide 15 text

ධՁ࣮ݧ̍ʢ̎̎ʣ w ఏҊख๏͸܇࿅σʔλϓϦϑΣονϓϩηε਺ʹൺྫͯ͠ੑೳ͕޲্͍ͯ͠Δ͕ɼϛχόον ੜ੒ϓϩηε਺ͱ͸΄΅૬͕ؔͳ͍ w ڞ༗ετϨʔδɼϊʔυϩʔΧϧετϨʔδڞʹϛχόονੜ੒ϓϩηε਺ʹൺྫͯ͠ੑೳ͕ ޲্͍ͯ͠Δɽ ఏҊख๏ ڞ༗ετϨʔδ ϊʔυϩʔΧϧετϨʔδ ఏҊख๏ ϓϦϑΣονϓϩηε਺ ϛχόονੜ੒ϓϩηε ఏҊख๏ʹ͓͍ͯϓϦϑΣονϓϩηε਺Λ มԽͤͨ͞ͱ͖ͷ݁Ռ શύλʔϯʹ͓͍ͯϛχόονੜ੒ϓϩηεΛ มԽͤͨ͞ͱ͖ͷ݁Ռ ʢఏҊख๏ͷϓϦϑΣονϓϩηε਺ʣ 'BTUFS 'BTUFS 15

Slide 16

Slide 16 text

ධՁ࣮ݧ̍ͷ ੑೳʹؔ͢Δߟ࡯ ࠷଎஋ JUFSBUJPOTFD όϯυ෯ <.#T> ߹ܭ࢖༻ϓϩηε਺ ϊʔυϩʔΧϧ ετϨʔδ ڞ༗ετϨʔδ ఏҊख๏ w ૝ఆ௨Γɼ࠷΋ετϨʔδσόΠεͱͯ͠ͷੑೳ͕ߴ͍44%͕࠷΋ߴ଎ w ఏҊख๏͸ڞ༗ετϨʔδΑΓ΋ߴ଎ͳ݁Ռ͕ಘΒΕɼϓϦϑΣονͷޮՌ͕֬ ೝͰ͖Δ w ڞ༗ετϨʔδͱఏҊख๏Λൺֱ͢ΔͱɼఏҊख๏ͷํ͕ΑΓগͳ͍ϓϩηε਺ Ͱڞ༗ετϨʔδΑΓ΋ߴ͍SFBE*0Λୡ੒͍ͯ͠Δ 16

Slide 17

Slide 17 text

ධՁ࣮ݧ̎ʢ̍̏ʣ w ࣮ΞϓϦέʔγϣϯͰͷධՁͱͯ͠$IBJOFSΛ༻͍ͯσʔλฒྻ܇࿅Λߦ͍ɼҎ ԼΛධՁ͢Δ w FQPDIͷσʔλฒྻ܇࿅ʹཁͨ͠߹ܭ࣌ؒ w ॲཧ಺༰ʹ͓͚Δϛχόονͷϩʔυ࣌ؒ w ධՁύλʔϯ͸ධՁ࣮ݧ̍ಉ༷ɼϊʔυ਺͸ ʢ113ݻఆʣ w Ϟσϧ͸3FT/FUͱ͍͏৞ΈࠐΈχϡʔϥϧωοτϫʔΫͷҰछΛ࢖༻ɼ࢖༻ ͢ΔσʔλɼϛχόοναΠζ͸ධՁ࣮ݧ̍ͱಉ͡ 17 w ˞༧ߘʹܝࡌͨ͠ʮFQPDIσʔλฒྻ܇࿅ʯͷධՁ݁Ռͱ͸ҎԼͷ఺͕ҟͳΓ·͢ w ॲཧ։࢝લʹESPQDBDIFTΛ༻͍ͯ1BHF$BDIFΛ࡟আ w ܇࿅͢ΔFQPDI਺Λ͔ΒʹมߋʢFQPDI໨Ҏ߱ͷ݁ՌʹมԽ͕ͳ͔ͬͨͨΊʣ w ίʔυͷ࠷దԽʢQSJOUσόοάʹΑΔΦʔόʔϔουΛۃྗ࡟আʣ

Slide 18

Slide 18 text

w σʔλฒྻ܇࿅͸.1*Ͱෳ਺ͷϓϩηεΛੜ੒͠ɼͦΕͧΕͷϓϩηε͕෼ׂ͞Εͨ ܇࿅σʔλΛSFBEͯ͠܇࿅͢Δ w ௨ৗͷ܇࿅ॲཧͱ͸ҟͳΓɼ#BDLXBSEͷ͋ͱʹ"MM3FEVDFΛ༻͍֤ͯϓϩηεؒͰ ಉظ্ͨ͠Ͱɼ֤ϓϩηεͷύϥϝʔλͷฏۉΛܭࢉͯ͠ڞ༗͢Δॲཧ͕ൃੜ͢Δ ධՁ࣮ݧ̎ʢ̎̏ʣ σʔλฒྻ܇࿅ʹ͓͚Δॲཧϑϩʔ 18

Slide 19

Slide 19 text

w ϊʔυ਺ͷ૿Ճͱͱ΋ʹͲͷख๏΋ੑೳ͕΄΅ઢܗʹεέʔϧ͍ͯ͠Δ w ఏҊख๏͸ͲͷέʔεͰ΋ڞ༗ετϨʔδΑΓ΋ߴ଎Ͱ͋Δͱ͍͏݁Ռ͕ಘ ΒΕͨ w ఏҊख๏ͷϊʔυϩʔΧϧετϨʔδʹର͢Δࠩ͸ ࠷େͰ໿ ඵࠩʢϊʔυʣɼ࠷খͰ໿ඵࠩʢϊʔυʣ ධՁ࣮ݧ̎ʢ̏̏ʣ ڞ༗ετϨʔδ ϊʔυϩʔΧϧετϨʔδ ఏҊख๏ ڞ༗ετϨʔδ ϊʔυϩʔΧϧετϨʔδ ఏҊख๏ 'BTUFS #FUUFS FQPDIͷσʔλฒྻ܇࿅ʹཁͨ͠ॲཧ࣌ؒ ϊʔυ਺ผͷ ̍ඵ͋ͨΓʹ܇࿅ॲཧͨ͠ը૾ͷຕ਺ 19

Slide 20

Slide 20 text

ධՁ࣮ݧ̎ʹ͓͚Δ ॲཧ࣌ؒͷ಺༁ʢ̍̑ʣ w ϊʔυʹ͓͚Δશ.1*ϓϩηεʢϓϩηεʣͷॲཧ࣌ؒΛੵΈ্͛ͨ஋ͷ ಺༁ΛݟΔͱɼFQPDI໨ͷϛχόονͷϩʔυ͸શମͷఔ౓Ͱ͋Γɼ FQPDI໨Ͱ͸΄ͱΜͲফࣦ͍ͯ͠Δ ఏҊख๏ʹ͓͚ΔϊʔυͰͷ શ.1*ϓϩηεͷશॲཧ࣌ؒ߹ܭ஋ ڞ༗ετϨʔδʹ͓͚ΔϊʔυͰͷ શ.1*ϓϩηεͷશॲཧ࣌ؒ߹ܭ஋ 'BTUFS 'BTUFS 20

Slide 21

Slide 21 text

ධՁ࣮ݧ̎ʹ͓͚Δ ॲཧ࣌ؒͷ಺༁ʢ̎̑ʣ w ఏҊख๏ͱϊʔυϩʔΧϧετϨʔδͰ͸FQPDI໨ʹͳΔͱ΄ͱΜͲॲཧ࣌ؒ ͕͔͔͓ͬͯΒͣɼϝϞϦ͔ΒͷΞΫηε࣌ؒͷΈ͕࿐ग़͍ͯ͠Δͱߟ͑ΒΕΔ w ڞ༗ετϨʔδͷΈ͕FQPDI໨ʹ͓͍ͯඵۙ͘Λཁ͍ͯ͠Δ͜ͱ͔Βɼε τϨʔδ͔ΒͷSFBE଴͕ͪ࣌ؒ͋Δఔ౓ଘࡏ͍ͯ͠Δ͜ͱ͕ߟ͑ΒΕΔ ϊʔυͰͷશ.1*ϓϩηεͷϛχόονͷϩʔυʹ͓͚Δ ॲཧ࣌ؒ߹ܭ஋ 'BTUFS ڞ༗ετϨʔδ ϊʔυϩʔΧϧετϨʔδ ఏҊख๏ ඵ ඵ 21

Slide 22

Slide 22 text

w FQPDIXϊʔυͰ܇࿅͍ͯ͠Δ࠷தʹɼ೚ҙʹબΜͩϊʔυʹ͓͚ ΔWNTUBUͰαϯϓϦϯάͨ͠XBʢ*0଴ͪ࣌ؒʣͷׂ߹ͷਪҠΛൺֱ w ϊʔυϩʔΧϧετϨʔδͱఏҊख๏͸FQPDI໨Ҏ߱͸΄ͱΜͲXBͷ஋ ͕ൃੜ͠ͳ͍͕ڞ༗ετϨʔδ͸FQPDI໨Ҏ߱΋XB͕Ұఆͷߴ͞ͰਪҠ FQPDI FQPDI FQPDI FQPDI FQPDI FQPDI ϊʔυϩʔΧϧετϨʔδʹ͓͚Δ XBͷਪҠ 22 ఏҊख๏ʹ͓͚Δ XBͷਪҠ ڞ༗ετϨʔδʹ͓͚Δ XBͷਪҠ ධՁ࣮ݧ̎ʹ͓͚Δ ॲཧ࣌ؒͷ಺༁ʢ̏̑ʣ

Slide 23

Slide 23 text

ධՁ࣮ݧ̎ʹ͓͚Δ ॲཧ࣌ؒͷ಺༁ʢ̐̑ʣ ϛχόονͷϩʔυͷ͕࣌ؒ JUFSBUJPOͷ܇࿅࣌ؒΑΓ΋௕͍৔߹ͷॲཧϑϩʔ ϛχόονͷϩʔυͷ͕࣌ؒ JUFSBUJPOͷ܇࿅࣌ؒΑΓ΋୹͍৔߹ͷॲཧϑϩʔ w ڞ༗ετϨʔδʹ͓͍ͯ͸܇࿅଎౓ΑΓ΋ϛχόονͷੜ੒࣌ؒͷํ͕௕͔͔͍ͬͯͨͨ͘Ίɼϛ χόονͷϩʔυ͕࣌ؒ͋Δఔ౓͔͔͍ͬͯͨͱߟ͑ΒΕΔʢࠨਤʣ w ҰํͰϊʔυϩʔΧϧετϨʔδɼఏҊख๏ʹ͓͍ͯ͸ৗʹϛχόον͕όοϑΝʹଘࡏ͍ͯ͠Δ ঢ়ଶʹͳ͍ͬͯͨͱߟ͑ΒΕΔʢӈਤʣ 23

Slide 24

Slide 24 text

ධՁ࣮ݧ̎ʹ͓͚Δ ॲཧ࣌ؒͷ಺༁ʢ̑̑ʣ ϊʔυͰͷશ.1*ϓϩηεͷ"MM3FEVDFͷ ॲཧ࣌ؒ߹ܭ஋ w "MM3FEVDFʹཁ͢Δॲཧ࣌ؒʹ͍ͭͯ͸ɼ྆FQPDIͱ΋ʹڞ༗ετϨʔδ͕࠷΋ॴཁ͕࣌ؒ௕͍ w ͜Ε͸ɼଞͷ̎ύλʔϯͱൺ΂ϛχόονϩʔυʹཁ͢Δ͕࣌ؒ௕͍ϓϩηε͕૿͑ͨ෼ɼ "MM3FEVDFͰશϓϩηεؒͰಉظΛͱΔͨΊʹ֤ϓϩηεͰಉظ଴͍ͪͯ͠Δ͕࣌ؒ௕ظԽͯ͠ ͍ΔͨΊͱߟ͑ΒΕΔ ϛχόονϩʔυʹΑ֤ͬͯϓϩηεͷ "MM3FEVDFͷಉظ଴ͪͷ͕࣌ؒ௕ظԽ͢Δྫ 'BTUFS ڞ༗ετϨʔδ ϊʔϩʔΧϧ ετϨʔδ ఏҊख๏ 24

Slide 25

Slide 25 text

ؔ࿈ݚڀ w ܇࿅σʔλઐ༻ͷ෼ࢄΩϟογϡαʔόΛઃܭ͠ɼ܇࿅ σʔλΛ͢΂ͯϝϞϦ্ʹΩϟογϡ͢Δ͜ͱͰSFBE࣌ ؒΛߴ଎Խ͢Δख๏ͷఏҊ<> w 5FOTPS'MPXʹ͓͍ͯ܇࿅σʔλͷฒྻSFBEͱϛχόον ͷόοϑΝϦϯάͷޮՌΛܭଌ͠ɼͦͷ༗ޮੑΛใࠂ<> [1] Zhu, Y., Chowdhury, F.,et al. “Entropy-Aware I/O Pipelining for Large-Scale Deep Learning on HPC Systems”, MASCOTS.2018.00023, pp.145-146 (2018) [2] X. Lu, H. Shi, M. H. Javed,et al. “Characterizing Deep Learning over Big Data (DLoBD) Stacks on RDMA-Capable Networks," 2017 IEEE 25th Annual Symposium on High- Performance Interconnects (HOTI), pp. 87-94 (2017) 25

Slide 26

Slide 26 text

·ͱΊ w ຊݚڀͰ͸ɼ௿଎ͳڞ༗ετϨʔδͱߴ଎ͳϊʔυϩʔΧϧετ ϨʔδΛ૊Έ߹Θͤͯɼػցֶशʹ͓͚ΔSFBE*0Λߴ଎Խ͢ Δख๏ΛఏҊͨ͠ w ϛχόονΛϩʔυ͢ΔϕϯνϚʔΫͰ͸ɼڞ༗ετϨʔ δͷΈΛ࢖༻͢Δ৔߹ͱൺ΂ͯ࠷େͰߴ଎ʹϛχόονΛ ϩʔυͰ͖ΔੑೳΛࣔͨ͠ w FQPDIͷσʔλฒྻ܇࿅ʹ͓͍ͯ͸ɼఏҊख๏͸ैདྷख๏ͱ มΘΒͳ͍εέʔϥϏϦςΟΛࣔ͠ɼશͯͷϊʔυ਺ʹ͓͍ͯڞ ༗ετϨʔδΑΓ΋܇࿅଎౓͕ߴ଎Ͱ͋Δͱ͍͏݁ՌΛࣔͨ͠ 26

Slide 27

Slide 27 text

ҎԼ༧උεϥΠυ 27

Slide 28

Slide 28 text

ϓϩηεGPSL࣌ͷ ઃఆʹ͍ͭͯ w $IBJOFSͷυΩϡϝϯτʹ*OpOJ#BOEͱ.VMUJQSPDFTT*UFSBUPS ͷ૊Έ߹ΘͤͰΫϥογϡ͢ΔՄೳੑ͕هࡌ͞Ε͍ͯΔ