1), (i) · 2018. 9. 25. · ? 9 • v $ tsqxoktwu •:*v • 1993 3& ; 0 ad+ 3 • 1995 3&...
TRANSCRIPT
$�)%(�'&)�(1), (I)��#) �������!)"*
� � � ��
���������+�,������� �����
2018/9/25 $�)%(�'&)� (1), (I) 1
�� (���
1. �����
2. ����
3. ����� �
4. ������
5. �(!' ��%()����*
2018/9/25 �"�(#'�&$(� (1), (I) 2
@4W/-
•I�7<3W026TNR)L#]QZWJ��EW�9^5�O]1. ���.iwjxWhxmxdwnsxj *%2. ����%TMPIW�K�3. I�7<3�%
•�=�9W5�V[\J�@)W�CW02^"$VAZ]MT^/-TO]
•<31�_t`_wh?�@4zbkeuxD)• http://www.compsci-alliance.jp• �H_t`_whYW,D^yhttp://www.compsci-alliance.jp/�)�G/
• ;�^(QPX��>^+8• -U'�&BJ�FfwpgarSW+:• !���hmdwW_bawl+8
2018/9/25 hmdwovctqwc (1), (I) 3
���(.40/�%(���#
•/1-625,436,798• �������7����8
•/1-625,436,7�8• �������� ���7����8
•���(��!��� (�� �+��%"*$
•���)!��!���(����%$• ��!���&+'!�(��� %$
2018/9/25 /1-625,436, (1), (I) 4
?�9�• ��v� $�tSQX OKTWu
• :*v• 1993�3& � ;���0��AD+��3 (
• 1995�3& � ;�����C0��253<7)3��#��>4��
• 1998�3& � ;�����C0��253<7)3��#��>4��
• 1998�4& '��3����A����3 ?�
• 2002�4&�2007�3& '��3��^qimrc_Z[qa�A ?�
• 2007�4&�2008�11& 6,��<73�25bqcr 25�
• 2008�12&w2013�11& 6,��`ael����253%"
• 2013�12&w2015�11& '��� ���1bqcr .�%"
• 2015�12&w/� ���1bqcr %"
• 2016�2&�/� t�!u��C��8253D+8���#
• EGPU^qimreYq]FEah^q�Jgdfpr\FQPR�BN-@�s
• =KISE�$�FN]]VOGWGWMHUL
2018/9/25 ah^qjo]nkq] (1), (I) 5
:3�-i��<;,)j
1. 9 25�(��)o LKSgR2. 10 2�
l � ��&F�"%0i��j
3. 10 9�oRWPg�'=�l fOKg�#BURVYfOb]�5
4. 10 16�l A�4YfOb\gO�$F�+li>�^_cBdhYJgfhcgOj (��)
5. 10 23�l A�4YfOb\gO�$F�+2iM`TQaXfTN�j
6. 10 30�l 5 -ZNVd.F� �
2018/9/25 RWPgYfOb\gO (1), (I)
7. 11 6�l HD�$F� �
8. 11 20�l 5 -5 .F� �ilj
9. 11 27�l 5 k5 .F� �imj
10. 12 4�l pq�7$ilj
l PgURV9@(6
11. 12 11�l pq�7$imj
12. 12 18�l pq�7$injB?�!;�
13. 1 8�l RB-HC8EB*/1��
e[hVCIGPgURV9@i2�o
2019�2 4�i j24� ��
6
,���
1. �'D0<SR-0T+9a^d[H=C�l %��
l +;K+:MF,�81NR
l >LCI 0T+:�*JG7e+;G7MF�:>f
2. XcZY[-0T+:l �.eg�ji�e��ffJ4 ��H2�3e��f
l @QUE��>RPITB:SK4�'a^d[�H�
• gIO4hIO4gEhI��4I/�86R5
• gEhI��T(A?��J4�!4%��D�&T��>R5
• �)I'�8#"4$�"H[`V=CM=7
2018/9/25 Y\Xc]bW`_cW (1), (I) 7
2016���1. 2016(�28)� S1S2
• ������/ 37�• )&,#�� �/ 22�
• ��/1���/13���/4���/3����� +"!#�� ��/3�-3�������.2. 2016(�28)� A1A2
• ������/ 9�• )&,#�� �/ 4�
• �/4����� +"!#�� ��/2�-2�������.3. 2017(�29)� S1S2
• ������/ 60�• )&,#�� �/ 40�
• ��: 3���/18���/13���/3����� +"!#�� ��/5�-3���������.4. 2017(�29)� A1A2
• ������/ 21�• )&,#�� �/ 11�
• �/6���: 1���: 3����� +"!#�� ��/1�-1�������.5. 2018(�30)� S1S2
• ������/ 37�• )&,#�� �/ 15�
• ��: 1���/5���: 8���: 0����� +"!#�� ��/5�-2�������.
2018/9/25 !$ +%*�('+� (1), (I) 8
6*R!�
•r24mYFDSFortran24nNtvs^%?L�+•VN_OC;���R5:�1�M/0
• wusxR��afkc;�T;ej`ihR��C]BKL?P?N�F?<
• P\UD/0-^"[FI?
• wusxR��afkcC]BKL?P?�;ej`ihR��C]BKL?P?�S;��Q;9�&Q�A\Z@8�FXGRM;7�PD.?LDJE?<
• �^,�m3��#�Q�Fpol'�n
• HR) q����$S(50l �;VW�C=�><
2018/9/25 bdakej`igk` (1), (I) 9
��/�4• ���#2���>L?/=MCM;LDHM?OOakforest-PACS=MCM;LDHM?<=@GP5��'0(
• 10�9�O�P/���.�687LA�-C=KMB5�,��(3��,( �)� '+$*&!
• ���.GPU9I=? (Reedbush-H)1�!0( �"%��N
2018/9/25 =C;LEJ:IFL: (1), (I) 10
�������
•����$'* �������������*)+!������
l ��&�!�PDF#��(http://www.cspp.cc.u-tokyo.ac.jp/hanawa/class/
2018/9/25 �"�*$)�&%*� (1), (I) 11
�/�• Mlmkrh+gF:Y`�,Se��`��b[N
• ��% K'#�&K9$�� 0
• "�(�KHJGI;=<F:@=?<B?@AA<KHJGI;=?F:ECD;@=?<B?@AACK*��s><=A�>�=D�K=CB7
• O �`)�P• lmkr`24�[WL��hSfdWU24VZQbWL
•lmkra�_�RgS•lmkra]i^�.c[K^X6U3-[Tg`S•���1K��`58\�!��K^]
2018/9/25 lmkrnqjporj (1), (I) 12
�)�|#-�}• Uorlzsykvtzk�5~��'`JLIa�-~V
• $! �" .S
• �%�SIMFKCAB9>9;=9:@<>?=9>S(���<:;=�=�;<�S �DE?87<::6
• W��a&�X
• G02_/3• G02SHPQRQNOC:02anzsxsykvuZ��•��ixmwpubS�_gYfe][3�•�4,a��h�^jq{•��b�+T�d^����1*h�c��\a�5�
2018/9/25 orlzsykvtzk (1), (I) 13
�&�} )�~• \��uzmxv{m�2�o{uyuzmxwc�gITORGJdITORBDD]
• !� �� +
• ���"�ZFKCH687A3;8:7=9;<=:ZFKCH68:A3@>?6;8:7=9;<=<Z% �� 978<�<�9<�
• ^��e#�_• D./ZESUWUMR@7./c-0• D./ZESUWUMR@7./e,�eo{uyuzmxw`���*}qk{z|r��~
•�1(e��j�bls|• LQRNSXV3JD )�*4DYPXQR$5[ptn{ci )�*[•��f'[hb��uzmxv{mj�g��ae�2�
2018/9/25 ptn{uzmxv{m (1), (I) 14
�Q7• ���4 �D h b]��.R������ h�
• ^E*> <�32 D� OS�B=!A <�C��32 ��D"® S�cE�@ / ��32��D"® S�';9� J0�32 ��("® S� �H:A¬5LU�1 S�
• �©�I�GT)85±ljkjijnimjd° � �± so° £«�4±lpl`�vxtw±rpqhnhmmrhjloqrhp° #�±mgrrj� ;mgqjj�¯Koefd
• �;7�C-�• u|}~}z{XZ�W[• 4 �¨�§�¥��4+����%�[6• ;\P��$������9 �a�W?�FT��,W?�uuy��«������V�4 YM�¨�§�¥�� «
• �$��N²�N�&_�����4 YM�"������
2018/9/25 �¡�ª¢©�¦¤ª� (1), (I) 15
�(�k\aZichYgdiY�6lM�%�#
• �2-M�9$.�VW`jBE3�• �(�J;<QS9�!V%;E��chYgdiYM$.�JBH�,G@RUH;T
• �=T��0+!>:TAJ> �
• ��M$.chYgeM�%LF;H1. ������&]i^jMFX10\jajZibfj^[\_eIDMOO�%CT
2. ')�MPCXg\^kMPI>�%I?TPMlI�%CT3. ���M��*M\jajZibfj^I�%CT
• �/MPCV%;H9kMPIINK;l4"chYgeI$.CTk�L4"chYgeM85�M17l
2018/9/25 \aZichYgdiY (1), (I) 16
�������� �������
2018/9/25 ��������� (1), (I) 17
�ª�ª�©�£ª���• �#LU.a��©�£ª����pkf��©�£ª����p• 2M��T��p
• C��6j§�¦�@O(U��|[O=�u�• RiH��kPC�®jc�k®� f�¡¢¥��|[O=
• m��A5����_3>n=>mb�_3PD�n«/�¬=>
mb�_3PD��WN�����A5��W�Z���}t^B��-V����K�n«$*28%117181R?E<K�¬�Z�!]���¦g�[O=
• N¯9N�h�°���¦g�[O=�o{~k�e6j(Us12.5���¤@O�`r���
• �ª�ª�©�£ª�"�,S ($*26%37311Gw��z)
• I dFQ� 3. u�,S�50TFLOPS��D\H6j(U�8y��ª�ª�©�£ª�ª�"��dFv��sku�!]Q��&X�'xYJyu��y�l
• C�k����y�~��ª�ª�©�£ª��� [O=• :���)��I�©�s+8y�Reedbush�ª�ª�©�£ª���� kOakforest-
PACS�ª�ª�©�£ª���� �k� [O=
2018/9/25 ���©�¨�¤�©� (1), (I) 18
�4kGPU 0;�`r~w�q!!
S^PmB�. (1)��B:D<BSn^nPm_gnUC��("�r��BCPUK�%:H7>=/�&K��
• MPP (Massively Parallel Processor)• S^Pm� �r��,PRIMEHPC-FXRinT0Cray XCRinT@?• � B\nZ-[VYlnN(MmUP[NY): ��,TOFU, Cray Aries (�)�8I@2)
• NhSU• ��BQn]K[VYlnN=#�9<1RSWd9;FB• \nZ-[VYlnNrIntel OmniPath, InfiniBand@?PfXLWL'
• EthernetC1EG�JI@2��CoS^Pm>9<Cp$�A+2r• aNYj
• NEC SXRinT• efiRSWd�*BPSY3�4:5H
• SMP (Symmetric Multi Processor)• ��efibRmrHP (�SGI) UVRinT0256CPU!�E=q7I���46=4@2
2018/9/25 S^Pm`kOhcmO (1), (I) 19
���"�� (2)
• ���� #�����
• GPU (NVIDIA Tesla)• PEZY-SC2• Intel Xeon Phi (Knights
Corner)• NEC SX-Aurora
TSUBASA• �
• PCI Express�#�=> ��"���#�
2018/9/25 ���"�!���"� (1), (I) 20
��ITC�GPU�����"�Reedbush-L���
HWLWGUMRWI4�*?��• TFLOPSYKSVNTJOH%
Tera Floating Point Operations Per SecondZ• \�#6\�7���Y�����Z,\FLOPS&
• KYCTZ8\,[[[Y Z%MYQBZ8\,[[[,[[[Y��Z%GYDBZ8\,[[[,[[[,[[[Y��Z%TYKSZ8\,[[[,[[[,[[[,[[[Y��Z
• 2+=%��#6���7�������7�,)? /5&
• PFLOPSY:IVNTJOHZ• \�#60.\�Y.*Z�7�������7�,)?&
• '�GUMRWI(Y2012�9��"�%11.2PFLOPSZ
2018/9/25 HLGUOTFSPUF (1), (I)
l PC7���8_l 4.2GH`Y\�#642��7ETJE���Z503%;0\ETJE)1>\�7�������,[email protected]&
l Intel Core i7 (Skylake)48%4GA%\ETJE416�7���� �,4-?74%4.2 GHz * 16��������/Hz * 4GA = 268.8 GFLOPS
l Cray-\8160MFLOPS& \^][��7HLGU<>%PC7�,1680�$!X
21
Sg[gRf\cgUD,���• $-\gP�(hTheoretical Peak Performancei
• ZgYMNL�(7GE=9<?�(3
• jPeWPC�)A9H ���# �7G'<?FLOPS�K�6;B8�53
• ��(hEffective Performancei• �G7D_fV`gPT]XMOLK�)<@�)�1K+!3
• >D_fV`gP^eQdbC�JI@5H ���#"'K'3
• ��D�K�C'<?FLOPS�D;B3
• /&���%�D�*_fV`gPA4HLINPACK2���0� (CG
�iD�*_fV`gPA4HHPCG8�:�JI@5H3
• ���#��D_fV`gPF��h�.i
2018/9/25 S[Rf^eQdafQ (1), (I) 22
0=3=/<49=1'����
• LINPACK6<27=-&'500�)&':<,<.
• #������'� (���)
• ��4�&10�• ����%"�($��
Linpack��
1�500�'�!
500�
1�
�/<49=1
��� (TaihuLight)
��2 (Tianhe-2)�/<49=1
SequoiaTitan
http://www.top500.org/ *+
03/<5;.:8<. (1), (I) 232018/9/25
LR:4��• �Intel�4���?RFOQLR:(��*."�����4��3�,8�!�#
$���CDI4 �5"&6-TWP 0U�328%
• )9'7�+/"
$J;<NINBD@4�5"&6-TWP 0U�328%
• ��3681"�V�0TS�128#• ��35"EGRF�
2018/9/25 AH>OIN=MKO= (1), (I) 24
7J?J4IADJ8*EI1I3(1)
• TOP500Khttp://www.top500.org/L• LINPACK*�#-����.��%'�*500�+)*EI1I3
•�0J2F96����M;=5J�>927@G�* Jack Dongarra ��$��
•� 6�@</:* ��ISC!11�@�* ��SC)��K�*""500,�(�&L
2018/9/25 7?4IBH3ECI3 (1), (I) 25
-E5E+D6?E/!@D(D*(2)
• Green500Fhttp://www.green500.org/G
• Top500 ���%�,-2<!������#���!@D(D*
• Linpack���!��& ����/��=FLOPS/W
• HPCG (http://www.hpcg-benchmark.org/)
•����(CG�) "$8D0:E)
•����!����!Linpack"#=>A').-�93B41) ���!'7A ��
2018/9/25 -5+D7C*@;D* (1), (I) 26
+<1<);27<,"8;';((3)
• Graph500=http://graph500.org/>• ��(83��&���TEPS (Traversed Edges Per Second)�8;';(
• ���� (BFS)!����2017/11�$SSSP (Single Source Shortest Paths)#��!
• Green Graph500 ��"#�%=�� ?>• IO500 (http://www.io500.org/)
• +.9<*"��=��"����>• 6,-<,�� (IOPS)•0;/ (GB/sec)
• 2017/11�$
2018/9/25 +1);4:(85;( (1), (I) 27
28
http://www.top500.org/
Site Computer/Year Vendor CoresRmax
(TFLOPS)Rpeak
(TFLOPS)Power(kW)
1 Summit, 2018, USADOE/SC/Oak Ridge National Laboratory
IBM Power System AC922, IBM POWER9 22C 3.07GHz, NVIDIA Volta GV100, Dual-rail Mellanox EDR Infiniband
2,282,544 122,300(= 122.3 PF)
187,659 8,806
2 Sunway TaihuLight, 2016, ChinaNational Supercomputing Center in Wuxi
Sunway MPP, Sunway SW26010 260C 1.45GHz, Sunway 10,649,600 93,015 125,436 15,371
3 Sieera, 2018, USADOE/NNSA/LLNL
IBM Power System S922LC, IBM POWER9 22C 3.1GHz, NVIDIA Volta GV100, Dual-rail Mellanox EDR Infiniband
1,572,480 71,610 119,194
4Tianhe-2A, 2018, ChinaNational Super Computer Center in Guangzhou
TH-IVB-FEP Cluster, Intel Xeon E5-2692v2 12C 2.2GHz, TH Express-2, Matrix-2000 4,981,760 61,445 100,679 18,482
5ABCI (AI Bridging Cloud Infrastructure), 2018, JapanNational Institute of Advanced Industrial Science and Technology (AIST)
PRIMERGY CX2550 M4, Xeon Gold 6148 20C 2.4GHz, NVIDIA Tesla V100 SXM2, Infiniband EDR
391,680 19,880 32,577 1,649
6Piz Daint, 2017, SwitzerlandSwiss National Supercomputing Centre (CSCS)
Cray XC50, Xeon E5-2690v3 12C 2.6GHz, Aries interconnect , NVIDIA Tesla P100 361,760 19,590 25,326 2,272
7 Titan, 2012, USADOE/SC/Oak Ridge National Laboratory
Cray XK7, Opteron 6274 16C 2.200GHz, Cray Gemini interconnect, NVIDIA K20x 560,640 17,590 27,113 8,209
8 Sequoia, 2011, USADOE/NNSA/LLNL
BlueGene/Q, Power BQC 16C 1.60 GHz, Custom 1,572,864 17,173 20,133 7,890
9 Trinity, 2017, USADOE/NNSA/LANL/SNL
Cray XC40, Intel Xeon Phi 7250 68C 1.4GHz, Aries interconnect
979,968 14,137 43,903 3,844
10 Cori, 2016, JapanDOE/SC/LBNL/NERSC
Cray XC40, Intel Xeon Phi 7250 68C 1.4GHz, Aries interconnect 622,336 14,016 27,881 3,939
12Oakforest-PACS, 2016, JapanJoint Center for Advanced High Performance Computing
PRIMERGY CX1640 M1, Intel Xeon Phi 7250 68C 1.4GHz, Intel Omni-Path 556,104 13,556 24,913 2,719
51th TOP500 List (June, 2018) Rmax: Performance of Linpack (TFLOPS)Rpeak: Peak Performance (TFLOPS), Power: kW
2018/9/25 ��������� (1), (I)
29
http://www.hpcg-benchmark.org/
HPCG Ranking (June, 2018)Computer Cores HPL Rmax
(Pflop/s)TOP500
RankHPCG
(Pflop/s) Peak
1 Summit 2,392,000 122.300 1 2.926 1.5%2 Sierra 835,584 71.610 3 1.796 1.5%3 K computer 705,024 10.510 16 0.603 5.3%4 Trinity 979,072 14.137 9 0.546 1.8%
5 Piz Daint 361,760 19.590 6 0.486 1.9%
6 Sunway TaihuLight 10,649,600 93.015 2 0.481 0.4%
7 Oakforest-PACS 557,056 13.555 12 0.385 1.5%8 Cori 632,400 13.832 10 0.355 1.3%9 Tera-1000-2 522,240 11.965 14 0.334 1.4%10 Sequoia 1,572,864 17.173 8 0.330 1.6%
2018/9/25 ��������� (1), (I)
30
Green 500 Ranking (June, 2018)TOP 500
Rank System Cores HPL Rmax(Pflop/s)
Power(MW) GFLOPS/W
1 359 Shoubu system B, Japan 794,400 858. 47 18.404
2 419 Suiren2, Japan 762,624 798. 47 16.835
3 385 Sakura, Japan 794,400 825. 50 16.657
4 227 DGX SaturnV Volta, USA 22,440 1,070. 97 15.113
5 1 Summit, USA 2,282,544 122,300. 8,806 13.889
6 19 TSUBAME3.0, Japan 135,828 8,125. 792 13.704
7 287 AIST AI Cloud, Japan 23,400 961. 76 12.681
8 5 ABCI, Japan 391,680 19,880. 1,649 12.054
9 255 MareNostrum P9 CTE, Spain 19,440 1,018. 86 11.865
10 171 RAIDEN GPU, Japan 35,360 1,213. 107 11.363
13 411 Reedbush-L, U.Tokyo, Japan 16,640 806. 79 10.167
19 414 Reedbush-H, U.Tokyo, Japan 17,760 802. 94 8.575
http://www.top500.org/
2018/9/25 ��������� (1), (I)
2018/9/25 ��������� (1), (I) 31
IO 500 Ranking (Nov., 2017)
Site Computer File system Client nodes IO500 Score BW(GiB/s)
MDkIOP/s)
1 JCAHPC, Japan Oakforest-PACS DDN IME 204 101.48 471.25 21.85
2 KAUST, Saudi Shaheen2 Cray DataWarp 300 70.90 151.53 33.17
3 KAUST, Saudi Shaheen2 Lustre 1000 41.00 54.17 31.03
4 JSC, Germany JURON BeeGFS 8 35.77 14.24 89.83
5 DKRZ, Germany Mistral Lustre 100 32.15 22.77 45.39
6 IBM, USA Sonasad Spectrum Scale 10 21.63 4.57 102.38
7 Fraunhofer, Germany Seislab BeeGFS 24 18.75 5.13 68.58
8 PNNL, USA EMSL Cascade Lustre 126 11.17 4.88 25.57
9 SNL, USA Serrano Spectrum Scale 16 4.25 0.65 27.98
http://www.io500.org/
Top5009;,;- =2018�6�!�>*+�+��+6/;
l 16�"$R-CCS: �.;48<210.5 PFLOPS
l 19����: TSUBAME3.08.125 PFLOPS
l 25,26��'�5.73 PFLOPS x2
l 32� ���?ITO4.54 PFLOPS
l 50� JAXA: SORA-MA3.15 PFLOPS
l 54��(�: Camphor23.05 PFLOPS
l 56����: FX1002.91 PFLOPS
l 61� )�$: JFRS-12.78 PFLOPS
l 77��&�$: FX1002.38 PFLOPS
l 83�%���: ATERUI II2.08 PFLOPS…
• �����+6/;l 180� �$: Sekirei
1.178 PFLOPSl 411�� �#1;2<:
Reedbush-L0.805 PFLOPS
l 414�� �#1;2<: Reedbush-H
0.802 PFLOPSl 436� �$: Sekirei-ACC
0.864 PFLOPS
2018/9/25 03.;5:-97;- (1), (I) 32
Sunway Taihulight�����• ��-��(Wuxi) � �#�,�,�. (NRCPC)• $.� 125.4 PF, Linpack 93.0 PF, 40960!.�• Sunway� SW26010
• 260���'�.��%+���(1+64��)*4�)��, 1.45GHz
• $.��/3.06TF• '(*",��/136.5 GB/s
• �,�� ��/InfiniBand FDR
2018/9/25 �#�,%+�)&,� (1), (I) 33
��/Top500, HPCWire Japan, PCwatch
Piz Daint @ CSCS• ��� ETH Zurich��������� �• ��� 33.8 PF, Linpack 19.5 PF (2017 upgrade)• 5,320 (P100 + Xeon Haswell) + 1,431 Xeon Broadwell
• Cray XC50 + XC40
2018/9/25 �������� (1), (I) 34
���https://www.cscs.ch/publications/news/piz-daint-one-of-the-most-powerful-supercomputers-in-the-world/
Cori @ NERSC• NERSC: ���!-�1�(DoE)/1.0�#1�.1��� �(LBNL)�1��• National Energy Research Scientific Computing Center
• 9,688 Intel Xeon Phi (KNL), %1��� 30 PF + 2,388 Intel Xeon (Haswell)• Cray XC40 �� )
• Gerty Cori: �����*,�����"1'-����
2018/9/25 �$�0&/�+(0� (1), (I) 35
�#1*.2&4K-Computer5@�� R-CCS• �� ������%1&24��,2' !/1(5
• 8678�9�����
• ;<=:SPARC64 VIIIfx(CPU�� 128GFLOPS)
• 2011�11�TOP500�LINPACK�
• ���:11.280 PLOPS• ���:10.510 PFLOPS ��:93.13
$)#1+0"/-1" (1), (I)
��:������� 4http://www.aics.riken.jp/index.html 5
2018/9/25 36
����-!-�,"'-�2���$'*-�3
•���������$'*-�3• NEC SX-ACE• 5,120�-�• 0�-����4��• ���20.1PFLOPS
• %&) ,��1.3PB/sec
�!�,#+�($,� (1), (I)
��2 ������.http://www.jamstec.go.jp/es/jp/system/index.html/
2018/9/25 37
������ TSUBAME3.0• HPE ICE-XA
• CPU: Intel Xeon E5-2680v4 2.4 GHz (14 cores) x 2(Hyperthreading enabled)
• GPU: NVIDIA Tesla P100 x 4
• Memory: 256GB• 540�
� ������� (1), (I)
��http://www.t3.gsic.titech.ac.jp/sites/default/files/guidance.pdf
2018/9/25 38
��.@/B$-3)@2�$��+-08E6*&'=CGD
39
FY11 12 13 14 15 16 17 18 19 20 21 22 23 24 25
Yayoi: Hitachi SR16000/M1IBM Power-7
54.9 TFLOPS, 11.2 TB
Reedbush, HPEBroadwell + Pascal
1.93 PFLOPS
T2K Tokyo140TF, 31.3TB
Oakforest-PACSFujitsu, Intel KNL25PFLOPS, 919.3TB
BDEC System50+ PFLOPS (?)
Oakleaf-FX: Fujitsu PRIMEHPC FX10, SPARC64 IXfx1.13 PFLOPS, 150 TB
Oakbridge-FX136.2 TFLOPS, 18.4 TB
Reedbush-L HPE
1.43 PFLOPS
Oakbridge-IIIntel/AMD/P9/ARM CPU only
5-10 PFLOPS
Big Data & Extreme Computing
������-B3B)@4:B/
92B)%����-B3B)@4:B/C!�� "�F���A��D
1B/� A+7:>B+;@��-B3B)@4:B/
���,;5���������#��-B3B)@4:B/
2018/9/25 -3)@6?(<7@( (1), (I)
2 (or 4))+-4!��• Oakleaf-FX (��� PRIMEHPC FX10)
• 1.135 PF, �(:05<,��, 2012�4�� 2018�3�• Oakbridge-FX (��� PRIMEHPC FX10)
• 136.2 TF, "�# ��>168�#?, 2014�4�� 2018�3�
• Reedbush (HPE, Intel BDW + NVIDIA P100 (Pascal))• ��ITC�&GPU)+-4, DDN IME (Burst Buffer)• .<,��;)358<)6:�+</<(:05<,>2016�7�� 2020�6�?• Reedbush-U: CPU only, 420 nodes, 508 TF (2016�7�)• Reedbush-H: 120 nodes, 2 GPUs/node: 1.42 PF (2017�3�)
• "�#*61 ����� ���%��+</<(:05<,>2017�10�@2020�6�?RB• Reedbush-L: 64 nodes, 4 GPUs/node: 1.43 PF (2017�10�)
• Oakforest-PACS (OFP) (���$Intel Xeon Phi (KNL))• JCAHPC (���CCS=��ITC)
402018/9/25 +/(:29'73:' (1), (I)
Reedbush����Reedbush-U2016�7�1� ���2016�9�1� ������
Reedbush-H2017�3�1� ���2017�4�3� ������
Reedbush-L2017�10�2� ���2017�11�1� ������
41
Top500: RB-L 291�@Nov. 2017RB-H 203�@Jun. 2017RB-U 361�@Nov. 2016
Green500: RB-L 11�@Nov. 2017RB-H 11�@Jun. 2017
2018/9/25 ����������� (1), (I)
2018/9/25 ��������� (1), (I) 42
外部接続ルータ 1Gigabit/10Gigabit Ethernet Network
InterConnect( 4x EDR InfiniBand) InterConnect( 4x EDR InfiniBand)
ログインノード群SGI Rackable C1110-GP2
6nodes
NFS Filesystem16TB
Lustre FilesystemDDN SFA14KE x3set
5.04PB
高速キャッシュDDN IME14K x6set
209TB
NAS Storage24TB
E5-2680v4 2.4GHz 14core,256GiB Mem
管理サーバ群SGI Rackable C1110-GP2
9nodes
GbE SW
x6
x6 x2x2(for PBS)
Reedbush-Hx240 (FDRx2/node)
Reebush-Ux420
x36(IME:6x6) x24(OSS(VM):x 12 x2)
x4(MDS:x 2)
x12
高速キャッシュDDN IME240 x8set
153.6 TB
管理用補助サーバ
SGI RackableC1110-GP2 x2
x16(IME:8x2) x2x12
x8 x2x10(Ctrl:8,MDS:2)
Reedbush-Lx128( EDR x2/node) x4
x6 x9x4
x9
x64
x120
x420
x9
Management port
管理コンソールMac Pro
電力管理サーバ
電力計器
Reedbush-USGI Rackable C2112-4GP3420 nodes, 508.03TFLOPS・CPU : E5-2695v4 2.1GHz 18core
Reedbush-HSGI Rackable C1102-GP8120 nodes, 240GPUs, 1.418PFLOPS・CPU : E5-2695v4 2.1GHz 18core ・GPU : NVIDIA Tesla P100 SXM2 x2/node
Reedbush-LSGI Rackable C1102-GP864 nodes, 256GPUs, 1.434PFLOPS・CPU : E5-2695v4 2.1GHz 18core ・GPU : NVIDIA Tesla P100 SXM2 x4/node
E5-2680v4 2.4GHz 14core,128GiB Mem
ライフ/管理ネットワーク 1Gigabit/10Gigabit Ethernet Network
Reedbush������
43
Reedbush-U Reedbush-H Reedbush-L
CPU/node Intel Xeon E5-2695v4 (Broadwell-EP, 2.1GHz, 18core) x 2sockets (1.210 TF), 256 GiB (153.6GB/sec)
GPU - NVIDIA Tesla P100 (Pascal, 5.3TF, 720GB/sec, 16GiB)
Infiniband EDR FDR2ch EDR2ch
�!�� 420 120 64
GPU� - 240 (=1202) 256 (=644)
�!���(TFLOPS) 509 1,417
(145 + 1,272)1,433
(76.8 + 1,358)
���� ��(TB/sec) 64.5 191.2
(18.4+172.8)194.2
(9.83+184.3)
���� 2016.07 2017.03 2017.10
2018/9/25 �� ����� � (1), (I)
Oakforest-PACS (OFP)• 2016�12�1�&�6�• 8,208 Intel Xeon/Phi (KNL)9Q[H�025PFLOPS
• ��4:�+
• TOP 500 9�\��2�]^HPCG 6�\��2�]\2017�11�]
•��(�HPC #�3(JCAHPC: Joint Center for Advanced High Performance Computing)• )"��2*%�$'MYN[• ������ #MYN[
• �����GUYPLC������ #MYN[�B9�!7C�/:��@A>?32<EL[P[JYQV[NKLOTF3.;^��(C�1 8�02* #F�+Z5�<E=DC,-
• http://jcahpc.jp
442018/9/25 LPJYRXIWSYI (1), (I)
Oakforest-PACS ��
��.�(5�=B<J08L:C7H=AKB@:=15�3��26� 5������!>J?L,M�N4& 5��-�*'$�#%>J?L,PO��/�)"��5� �9FJB=
��+�www.jiji.com
452018/9/25 =B<JDI;GEJ; (1), (I) 45
Oakforest-PACSの特徴 (1/2)• ��(8&
• 1(8& 68��<3TFLOPS�8,208(8&= 25 PFLOPS
• /039MCDRAM9��<16GB:;DDR49��<96GB::
• (8&���
• +4)�#� 17)7&����Fat-Tree'$%68�
• � ����-3�8 17����<�!1,��
• Intel Omni-Path Architecture
462018/9/25 "*�7-5�2.7� (1), (I) 46
Oakforest-PACS �� 47%,!9/8 519 (1), (I)2018/9/25
�-:����� 25 PFLOPS*:)� 8,208��*:)
Product ��� PRIMERGY CX600 M1 (2U) + CX1640 M1 x 8node
/8&'" Intel® Xeon Phi™ 7250;��!:): Knights Landing<68 !��1.4 GHz
236 �+9) 16 GB, MCDRAM,�� 490 GB/sec�+9) 96 GB, DDR4-2400, -:� 115.2
GB/sec�����
Product Intel® Omni-Path Architecture69�� 100 Gbps(08$ .7+�&�#49+9)Fat-tree�
47
Oakforest-PACS '� H2 / 2I• <,-CI/O
• ��<,-C236?:Lustre 26PB
• <,-C.@52A236?HDDN IMEIJ1TB/sec+ #*���,�1PB• ����F;507G4��F����&(��
• ��!�• Green 500%(��6�• LinpackJ 2.72 MW
• 4,986 MFLOPS/WHOFPI• 830 MFLOPS/WH�I
��<,-C236?
<,-C.@52A236?
B5/�$)1209G8'"��
482018/9/25 3:1E=D0B>E0 (1), (I) 48
Oakforest-PACS ���1 �2��&��.� ")
Type Lustre File System��� 26.2 PBProduct DataDirect Networks SFA14KE�$0# 500 GB/sec
��&��.�*!�+� ")
Type Burst Buffer, Infinite Memory Engine (by DDN)
��� 940 TB (NVMe SSD,%-"����)Product DataDirect Networks IME14K�$0# 1,560 GB/sec
����� 4.2MW1�����2�,!� 102
492018/9/25 %�0'/�,(0� (1), (I) 49
Oakforest-PACS >RZTIJG• OS: Red Hat Enterprise Linux (fLHgWiU)-
CentOS 1C? McKernel (#�WiU-���!)• McKernel: ��AICS:*��>`ViNG 4OS
• Linux��-Linux=�@&)-biO\fLd_=�0D�+<7
•]QT�NgYaiS=A�'6ED��.• NgXHdjGCC, Intel Compiler, XcalableMP
• XcalableMP: ��AICS;���:��*��>�\fLd^gL"%
• CBFortran:$(6E9NiU=���F0D5;:-�!>,/�G\eMiPcgF �=*�8D5;2:3D.
• dH[dehG\eMiPcgj Ki\gRiQRZTIJG• ppOpen-HPC, OpenFOAM, ABINIT-MP, PHASE system, FrontFlow/blue,
LAPACK, ScaLAPACK, PETSc, METIS, SuperLU etc.
502018/9/25 QXNg\fLd^gL (1), (I) 50
FX10� ,����
Memory Memory Memory
�012�����
Core#1
Core#2
Core#3
Core#0
2018/9/25 �!�+"*�(#+� (1), (I) 51
-������
Core#13
Core#14
Core#15
Core#12…
L2 (16������12MB)
L1 L1 L1 L1 L1 L1 L1 L1: L1�,��&��'32KB
85GB/=(8Byte�1333MHz�8 channel)
DDR3 DIMMMemory
4GB �2� 4GB �2� 4GB �2� 4GB �2�
,����$%) .8GB�4/32GB
20GB/
TOFUNetwork
ICC
FX10����"$TOFU��#2018/9/25 ��� ����� � (1), (I)
52
�!� �!�
�!��!�
�!�
�!��!�
�!��!�
�!�
�!�
�!�
$TOFU��
6����5GB/�"���#
���!��
1TOFU���� �
�!�
?TOFU��
?TOFU��
?TOFU��
?TOFU��
?TOFU��
?TOFU��
?TOFU��
?TOFU��
?TOFU��
?TOFU��
?TOFU��
?TOFU��
?TOFU��
?TOFU��
?TOFU��
?TOFU��
?TOFU��
?TOFU��
FX10&���=?TOFU���& �>2018/9/25 03-;4:,95;, (1), (I) 53
?TOFU��
?TOFU��
?TOFU��
?TOFU��
?TOFU��
?TOFU��
?TOFU��
?TOFU��
?TOFU��
@��
l8<.�*�+#�X��Y��Z�% �!��&1TOFU#���&?TOFU'����!)�(�=@�2<90�>l ����� �"'
l X�'2<90l Y�'61/7l Z�'61/7(�'�2<90
%$�!�(�
Reedbush-U�#� �!���• �� �������� ������� ��
=> NUMA (Non-Uniform Memory Access)(FX10�����)
Intel Xeon E5-2695 v4 (Broadwell-
EP)
QPI76.8GB/s
76.8GB/s
IB EDRHCA
15.7 GB/s
DDR4��
128GB
76.8GB/s 76.8GB/s
Intel Xeon E5-2695 v4 (Broadwell-
EP)QPIDDR4DDR4DDR4
DDR4DDR4DDR4DDR4
�� 128GB
G3 x16
2018/9/25 ���"�!���"� (1), (I) 54
Broadwell-EP��
Memory Memory Memory
2018/9/25 ���' &�$!'� (1), (I) 55
)��������
76.8 GB/�=(8Byte 2400MHz 4 channel)
DDR4DIMM Memory
16GB 2� 16GB 2� 16GB 2� 16GB 2�
�������"#%�*16GB 8+128GB
Core#0
L1
L2 L3
Core#1
L1
L2 L3
Core#2
L1
L2 L3
Core#3
L1
L2 L3
Core#4
L1
L2 L3
Core#5
L1
L2 L3
Core#6
L1
L2 L3
Core#7
L1
L2 L3
Core#8
L1
L2 L3
Core#9
L1
L2 L3
Core#10
L1
L2 L3
Core#11
L1
L2 L3
Core#12
L1
L2 L3
Core#13
L1
L2 L3
Core#14
L1
L2 L3
Core#15
L1
L2 L3
Core#16
L1
L2 L3
Core#17
L1
L2 L3
QPI x2 PCIe �����L1�(�: 2KB, L2: 256KB, L3: 2.5MB(��) => L3 ����45MB
Reedbush-U(���• >D<0614BF<F:.�#Fat Tree�
• &(,�'��;G:.�/$+��'���$�� ��
• Mellanox InfiniBand EDR 4x CS7500: 648@G9• ��)36@G95087 (SB7800). (36+18)��*-!"+(%��• RB-H)+�1 �RB-L%)���'��
2018/9/25 5=3F?E2CAF2 (1), (I) 56
181 3619 5437
Downlink: 18
. . . . . . . . .
Uplink: 18
. . .
. . .
36@G9Leaf 5087
36�
36@G9Spine5087
18�
648@G9Director5087 1�(��
Reedbush-H��������
NVIDIA Pascal
NVIDIA Pascal
NVLinK20 GB/s
Intel Xeon E5-2695 v4 (Broadwell-
EP)
NVLinK20 GB/s
QPI76.8GB/s
76.8GB/s
IB FDRHCA
G3
x16 15.7 GB/s 15.7 GB/s
DDR4���128G
B
EDR switch
EDR
76.8GB/s 76.8GB/s
Intel Xeon E5-2695 v4 (Broadwell-
EP)QPIDDR4DDR4DDR4
DDR4DDR4DDR4DDR4
���128G
B
PCIe swG
3 x16
PCIe sw
G3
x16
G3
x16
G3 x16
G3 x16
IB FDRHCA
2018/9/25 ������� �� (1), (I) 57
Oakforest-PACS ��'�• Intel Xeon Phi (Knights Landing)
• 1�'�1����
• MCDRAM: �&���'����&��!"$16GB
+ DDR4!"$
2018/9/25 ���&�%�# &� (1), (I)
Knights Landing Overview
Chip: 36 Tiles interconnected by 2D Mesh Tile: 2 Cores + 2 VPU/core + 1 MB L2 Memory: MCDRAM: 16 GB on-package; High BW DDR4: 6 channels @ 2400 up to 384GB IO: 36 lanes PCIe Gen3. 4 lanes of DMI for chipset Node: 1-Socket only Fabric: Omni-Path on-package (not shown) Vector Peak Perf: 3+TF DP and 6+TF SP Flops Scalar Perf: ~3x over Knights Corner Streams Triad (GB/s): MCDRAM : 400+; DDR: 90+
TILE
4
2 VPU
Core
2 VPU
Core
1MB L2
CHA
Package
Source Intel: All products, computer systems, dates and figures specified are preliminary based on current expectations, and are subject to change without notice. KNL data are preliminary based on current expectations and are subject to change without notice. 1Binary Compatible with Intel Xeon processors using Haswell Instruction Set (except TSX). 2Bandwidth numbers are based on STREAM-like memory access pattern when MCDRAM used as flat memory. Results have been estimated based on internal Intel analysis and are provided for informational purposes only. Any difference in system hardware or software design or configuration may affect actual performance. Omni-path not shown
EDC EDC PCIe Gen 3
EDC EDC
Tile
DDR MC DDR MC
EDC EDC misc EDC EDC
36 Tiles connected by
2D Mesh Interconnect
MCDRAM MCDRAM MCDRAM MCDRAM
3
DDR4
CHANNELS
3
DDR4
CHANNELS
MCDRAM MCDRAM MCDRAM MCDRAM
DMI
2 x161 x4
X4 DMI
HotChips27
KNL�#����
Knights Landing: Next Intel® Xeon Phi™ Processor
First self-boot Intel® Xeon Phi™ processor that is binary compatible with main line IA. Boots standard OS.
Significant improvement in scalar and vector performance
Integration of Memory on package: innovative memory architecture for high bandwidth and high capacity
Integration of Fabric on package
Potential future options subject to change without notice. All timeframes, features, products and dates are preliminary forecasts and subject to change without further notification.
Three products
KNL Self-Boot KNL Self-Boot w/ Fabric KNL Card
(Baseline) (Fabric Integrated) (PCIe-Card)
Intel® Many-Core Processor targeted for HPC and Supercomputing
2 VPU 2 VPU
Core Core
1MB
L2
MCDRAM: 490GB/��� (��)DDR4: 115.2 GB/�=(8Byte�2400MHz�6 channel)
����� �!"$*16GB�6+96GB
58
Oakforest-PACS: Intel Omni-Path Architecture %)+<C:-4.1AE:E8Fat-tree�
768 port Director
Switch
12�(Source by Intel)
48 port Edge Switch
362 �
22
241 4825 7249
Uplink: 24
Downlink: 24
. . . . . . . . .
037' +!<C:-4.1AE:E8,��• 136@�����%(���� �,��• ��$��G2A=%"+��9F8�*�#&���!��
592018/9/25 3;0E>D/B?E/ (1), (I)
��9F8B5.%��
�����!N`PaOakforest-PACSMaVaJ`W[aPLMQZ?�,'b0./2�1�/��c
• VaOT^JaMb�-c• 100,000� d 1�8UaS(��)4��512UaSA;�&�• 3�A;
• I^aXJaM• 400,000� (�� 480,000�) d 1� 8UaSb��c4��2048UaSA;
• ��@45RaH`6;+ • 9*BUaS�3360�324�-?5RaH`68�7DFE
• VaOT^JaM@2UaS"�• ��UaSA;@4RaH`�(��81.0• ��UaSG)7E<4)7:@4�(��82.0>=E• ��$?\aK@Reedbush<?"�RaH`#&C�%
2018/9/25 MVJ`X_I]Y`I (1), (I) 60
�����!K]M^ReedbushJ^S^G]TX^MIJNW;�,'_2018�4�1�`
• S^LQ[G^J_�-`• 150,000� a RB-U: 4R^P_��`/��16R^P=7
RB-H: 1R^P_��`/�(�� 2.5xRB-L: 1R^P_��`/�(�� 4.0x
• F[^UG^J• 300,000�a 1� 4R^P_��`/��128R^P=7/
RB-Ha 1R^P_��`/O^E]��<U;2.5�RB-La 1R^P_��`/O^E]��<U;4.0�
• RB-U;>�� 360,000� a 1� 4R^P_��`/��128R^P=7• RB-H;>�� 216,000� a 1� 1R^P_��`• RB-L;>�� 360,000� a 1� 1R^P_��`
• ��</0O^E]17+ • 5*>R^P�.360�.24�-;0O^E]14�3@CB• ��R^P=7</O^E]�(��41.0 (H<2.5, L< 4.0)• ��R^PD)3B8/)36</�(��42�:9B• ��$;Y^H<Oakforest-PACS8;"�O^E]#&?�%• R^P��?2A
2018/9/25 JSG]U\FZV]F (1), (I) 61
JPY (=Watt)/GFLOPS RateSmaller is better (efficient)
62
System JPY/GFLOPSOakleaf/Oakbridge-FX (Fujitsu)(Fujitsu PRIMEHPC FX10) 125
Reedbush-U (SGI)(Intel BDW) 62.0
Reedbush-H (SGI)(Intel BDW+NVIDIA P100) 17.1
Oakforest-PACS (Fujitsu)(Intel Xeon Phi/Knights Landing) 16.5
2018/9/25 ��������� (1), (I)
kpg|q{fwt|fW�7
v}iVTQR�
• �8U:3d^`0!AV5c_Oa
• :3VFFagkod^`�ISHa
• �M!AD�3S^`�IW:3GSHa
• kpg|Wjknu*�d-9LD*�V�cOPq{fwt|f
• �zryW���d�H�NDeyhxluW��
?.�VTQR]�
• <@Uxm}kUWSD#�V�QRZLE• RB-UX,!)S90%>IW .+~,�/VXZ[($�
• 100�Wsj|W .�+d10%�JbY10��LPKTV
2018/9/25 kpg|q{fwt|f (1), (I)
• OFPW��• � �"C1.1�D
5��6CS72.2�D2'�T��
• B%���A44�~1;D&=\�D3.2 MWI_E
63
�RYGVTZG;@^• "���?RYGVU\���&T ]D)_�?���D�89)T / _ >5C2;*
• ���-:@��*• �'@):0C.<,.@)����?��\EXIWKU]: 01 (43/$,• EXIWKU�)��>�:0=+%�?��• #�?6A?F[PSMO?��
• #��7�/B�&• N[L !�&
JQHZRYGVTZG (1), (I) 64
T
T / _
2018/9/25
MPI27• Message Passing Interface
• ULJ\G%��6W@QWX �
• TCP/IP6KBLN%�435��_./2"$7��• � �!�+�
• RZJH�6)��FHMT]Massively Parallel Processing (MPP)FHMT^5�*;
• �����RZJH81 #5�*:<0);
• ��+��• API]Application Programming Interface^6���
• HB\WPXM?(��+')• %���=V\E+"$-;,259;>YDXIT6�&�+�
2018/9/25 HOC[RZAWS[A (1), (I) 65
Oakforest-PACS �����
2018/9/25 ��� ����� (1), (I) 66
��� ���������
2018/9/25 ����������� (1), (I) 67