Transcript
2015 LENOVO INTERNAL. ALL RIGHTS RESERVED.
Torsten Bloth –
[email protected]
That‘s me
Torsten Bloth HPC Systems Architect leading architect for the LRZ SuperMUC Phase 2 supercomputer. started with IBM back in 2005 and was part of the IBM System x transition into Lenovo 2014. lives in beautiful Potsdam and it took him a while to travel to Lugano ;)
2
2015 LENOVO INTERNAL. ALL RIGHTS RESERVED.
Torsten Bloth –
[email protected]
There is No Agenda
Lenovo and HPC LRZ SuperMUC Phase 1 Review LRZ SuperMUC Phase 2 – Overview and current status – Technology used – Advantages of Water Cooling Technology
4
2015 LENOVO INTERNAL. ALL RIGHTS RESERVED.
Torsten Bloth –
[email protected]
Lenovo
Who we are
2015 LENOVO INTERNAL. ALL RIGHTS RESERVED.
Torsten Bloth –
[email protected]
Who is Lenovo?
A $39 billion, Fortune 500 technology company – Publicly listed/traded on the Hong Kong Stock Exchange – 60,000+ employees serving clients in 160+ countries A global company – Two headquarters in Raleigh, N.C and Beijing, China – Major research centers in the U.S, Japan, China – Manufacturing in U.S., China, India, Brazil, Mexico Invests in innovation – Ranked as one of Top 25 most innovative companies – #1 in worldwide PC market share – #2 in worldwide PC & tablet market share
6
2015 LENOVO INTERNAL. ALL RIGHTS RESERVED.
Torsten Bloth –
[email protected]
Lenovo Enterprise System Portfolio High-end rack systems 4/8 socket enterprise-class x86 performance, resiliency, security
Dense systems Optimize space-constrained data centers with extreme performance and energy efficiency
Converged/Blade systems Integration across IBM assets in systems and SW for maximum client optimization and value
1P & 2P Rack & Tower systems Broad rack and tower portfolio to meet a wide range of client needs from infrastructure to technical computing
System x Cloud 7
2015 LENOVO INTERNAL. ALL RIGHTS RESERVED.
SOLUTIONS Analytics Torsten Bloth –
[email protected]
Technical Computing
HPC Storage Portfolio Home Grown + Legacy Lenovo + IBM OEM Offerings
Direct Attach GSS 21/22
Network Attach GSS 24/26 8
2015 LENOVO INTERNAL. ALL RIGHTS RESERVED.
Torsten Bloth –
[email protected]
IBM Storwize V3700/V7000 Unified
HPC Center Opening
a permanent HPC benchmarking and R&D center in Stuttgart, Germany Technology partners will gain access to a state-of-the-art benchmarking facility The center will provide HPC support to Lenovo partners worldwide Industry partners – Intel, IBM, Mellanox, NVIDIA Client partners - Leibniz-Rechenzentrum (LRZ), Science & Technology Facilities Council: Hartree Centre, Barcelona Supercomputing Centre (BSC), Cineca, Rechenzentrum Garching (RZG), Forschungszentrum Jülich (FZJ), Distributed Research utilizing Advanced Computing (DiRAC) ISV partners – ScaleMP, Allinea, PathScale
9
2015 LENOVO INTERNAL. ALL RIGHTS RESERVED.
Torsten Bloth –
[email protected]
SuperMUC Phase 1
Review
2015 LENOVO INTERNAL. ALL RIGHTS RESERVED.
Torsten Bloth –
[email protected]
LRZ SuperMUC – Introduction
2 HPC AC presentations from Klaus Gottschalk – 2011: SuperMUC HPC System – 2013: LRZ SuperMUC – One Year of Operation
11
2015 LENOVO INTERNAL. ALL RIGHTS RESERVED.
Torsten Bloth –
[email protected]
SuperMUC Phase 1 Review
12
2015 LENOVO INTERNAL. ALL RIGHTS RESERVED.
Torsten Bloth –
[email protected]
SuperMUC Phase 1 Review
3 PetaFLOP/s Peak Performance 9288 water cooled nodes with 2 Intel SB-EP 207 nodes with 4 Intel WSM-EX CPUs 324 TB Memory Mellanox InfiniBand FDR10 Interconnect Large File Space for multiple purpose 10 PiB + 200GiB/s GPFS @ DDN SFA12k 2 PiB NAS Storage with 10GiByte/s Innovative Technology for Energy Effective Computing Warm Water Cooling Energy Aware Scheduling with xCAT and LL Huge Monitoring instances PUE ~1.15 through all season free-cooling capability No GPGPUs or other Accelerator Technology 13
2015 LENOVO INTERNAL. ALL RIGHTS RESERVED.
Torsten Bloth –
[email protected]
How do 10k IB cables look like?
14
2015 LENOVO INTERNAL. ALL RIGHTS RESERVED.
Torsten Bloth –
[email protected]
9288 IBM System x iDataPlex dx360 M4 – 43997256 Components – 18576 Intel SandyBridge-EP – 148608 CPU Cores –8.08m² CMOS
– 74304 4GB DIMMs 11868 Optical Cables
192640m
23736 QSFP Connectors 15
2015 LENOVO INTERNAL. ALL RIGHTS RESERVED.
5040 3TB disks, RAID6 Tubes and Hoses – 690m Stainless Steel – 4039m EPDM – 34153m Copper – 7.9m³ Water Mass: 194100 kg Hardware and 5year Maintenance, Energy, and Support: € 83000000 Torsten Bloth –
[email protected]
SuperMUC Phase 2
Solution Overview
2015 LENOVO INTERNAL. ALL RIGHTS RESERVED.
Torsten Bloth –
[email protected]
Heck! Moores Law?
Adding ~3PFLOP/s, Direct Water Cooled Island Design almost equal to SuperMUC Phase I 6 Compute + 1 I/O islands
$WORK parallel file system on GSS Technology • Adding 6PB net capacity • 100GB/s
3096 compute nodes nx360 M5 WCT InfiniBand FDR14 IO island with GPFS backend and management servers
18
2015 LENOVO INTERNAL. ALL RIGHTS RESERVED.
$HOME File System • Adding 2PB
Torsten Bloth –
[email protected]
From Chip to System System 6 Domains 3096 Server Domain 8 Racks 516 Server 508 TF/s 33.024TB RAM
Chassis 12 Server ~12 TF/s 768GB RAM
Server 2 Chips 986 GF/s 64GB RAM
Chip ~493 GF/s
19
2015 LENOVO INTERNAL. ALL RIGHTS RESERVED.
Torsten Bloth –
[email protected]
Compute Server
3096 Compute Nodes Lenovo NeXtScale nx360M5 WCT 2 x Intel E5-2697 v3 2.6GHz 14c 64GB Memory Mellanox Connect-IB Single Port Diskless Direct Water Cooled
20
2015 LENOVO INTERNAL. ALL RIGHTS RESERVED.
Torsten Bloth –
[email protected]
InfiniBand Concept #1
.......
#42
3
4:1
1:1 516 (compute nodes)
21
516
2015 LENOVO INTERNAL. ALL RIGHTS RESERVED.
516
516
Torsten Bloth –
[email protected]
516
516
IO Components
InfiniBand Concept – Phase 1 + Phase 2
Fabric #1
36p FDR10 #1
Fabric #2
36p FDR10 #18
36p FDR10 #126
18p leaf #7
Thin compute Edge Switch #2 18p leaf #8
18p leaf #36
18p leaf #1
18p leaf #7
Thin compute Edge Switch #19 18p leaf #8
18p leaf #36
18p leaf #1
18p leaf #7
I/O #1 Edge Switch #20 18p leaf #8
1 n01
18p leaf #18
18p leaf #1
18p leaf #7
Thin compute Edge Switch #1 18p leaf #8
18p leaf #36
18p leaf #1
18p leaf #7
Thin compute Edge Switch #6 18p leaf #8
18p leaf #36
18p leaf #1
n516
SuperMUC Phase I 20 islands (1+18+1)
n01
n516
2015 LENOVO INTERNAL. ALL RIGHTS RESERVED.
Torsten Bloth –
[email protected]
18p leaf #7
I/O #2 Edge Switch #7 18p leaf #8
1
GPFS available everyhwhere via Multi-Homed GPFS
22
36p FDR #42
3
1
18p leaf #1
36p FDR #6
36p FDR #1
SuperMUC Phase II 6+1 islands
18p leaf #21
SuperMUC Phase 2
The Technology
2015 LENOVO INTERNAL. ALL RIGHTS RESERVED.
Torsten Bloth –
[email protected]
Technology Overview - NeXtScale Chassis
Standard Rack
Compute
Storage
Acceleration
Water Cooled Node
24
2015 LENOVO INTERNAL. ALL RIGHTS RESERVED.
Torsten Bloth –
[email protected]
The Chassis IBM NeXtScale n1200 WCT Enclosure
System infrastructure Simple architecture
Water Cool Chassis 6U Chassis, 6 bays Each bay houses a full wide, 2node tray (12 nodes per 6U chassis) Up to 6x 900W or 1300W power supplies N+N or N+1 configurations No fans except PSUs Fan and Power Controller Drip sensor, error LEDs, and web link for detecting water leaks No built in networking
Front View shown with 12 compute nodes installed (6 trays)
3x power supplies
Rear fillers/EMC Fan and Power shields Controller
Rear View 25
2015 LENOVO INTERNAL. ALL RIGHTS RESERVED.
Torsten Bloth –
[email protected]
3x power supplies
Rear fillers/EMC shields
The Node
System infrastructure Simple architecture
Water Cool Compute Node 2 compute nodes per full wide 1U tray Water circulated through cooling tubes for component level cooling Dual socket Intel E5-2600 v3 processors (up to 165W) 16x DIMM slots (DDR4, 2133MHz) InfiniBand support: FDR: ConnectX-3 (ML2) FDR: Connect-IB (PCIe) QDR (PCIe) Onboard GbE NICs
nx360 M5 WCT Compute Tray (2 nodes)
Dual-port ML2 (IB/Ethernet)
CPU with liquid 16x DIMM slots cooled heatsink x16 ML2 slot
2015 LENOVO INTERNAL. ALL RIGHTS RESERVED.
1 GbE ports
Power, LEDs
PCI slot for Infiniband
Cooling tubes Water Inlet
PCI slot for Connect IB x16 ML2 slot PCI slot for Connect IB
26
Labeling tag
Torsten Bloth –
[email protected]
Water Outlet
How it works
27
2015 LENOVO INTERNAL. ALL RIGHTS RESERVED.
Torsten Bloth –
[email protected]
SuperMUC Phase 2
Why Water Cool Technology
2015 LENOVO INTERNAL. ALL RIGHTS RESERVED.
Torsten Bloth –
[email protected]
Temperature Impact on Processor Performance E5-2697 v3 145W Junction Temp vs Performance on NeXtScale WCT
• 1,015
18 °C Inlet
45 °C Inlet
995
n101
Linpack Score (Gflops)
n102
•
n104 975
n105
Example: Linpack scores across a range of temperatures: • 12 sample processors running on NeXtScale System WCT Linpack scores remain mostly flat for junction temperatures in the range that water cooling operates.
n106
Air Cooled
n107
•
The Linpack scores drop significantly when junction temperature is in range that air cooling operates.
•
Conclusion: Water Cooling enables the highest performance possible for each processor SKU at any water inlet temperature under 45°C
n108
955
n109 n110 n111 935 n112 Tj,max
915 20
30
40
50
60
70
80
90
100
Junction Temperature, Tj (C)
Vinod Kamath 29
2015 LENOVO INTERNAL. ALL RIGHTS RESERVED.
Torsten Bloth –
[email protected]
Chip Junction vs Water Inlet Temperature – nx360 M5 WCT
• Linpack scores remain high and relatively stable for chip junction temperatures <60°C Achieved using <45°C inlet water • Performance drops off significantly at higher junction temps Typical for Air Cool n101
n102 Linpack
Linpack
T_water_inlet
Tj
Score
Tj
Score
(°C)
(°C)
(Gflops)
(°C)
(Gflops)
8
28
934.5
27
948.4
18
35
935.2
33
948.3
24
41
934.7
42
948.1
35
54
931.9
52
946.3
45
60
926.7
60
944.7
55
68
921.3
73
938.5
55 1
75
918.9
79
936.1
Note: Typical flow is 0.5 lpm, 551 is 0.25 lpm/node 30
2015 LENOVO INTERNAL. ALL RIGHTS RESERVED.
Torsten Bloth –
[email protected]
Vinod Kamath
High Power Processor Support in WCT E5-2698A v3 Junction Temp vs Performance on NeXtScale WCT 1,150
18 °C Inlet
NeXtScale WCT can cool the Xeon E52698A processor due to greater thermal capabilities
35 °C Inlet
1,100
Air Cooled n125
Linpack Score (Gflops)
1,050
n126
n127 1,000
• 165W, 2.8GHz, 16 cores • Highest Linpack performance per node of any Xeon processor at 1.083 Tflops • Highest Gflops/Watt of any Xeon processor at 2.45 GFlops/Watt • Cannot be cooled by air • Processor throttles at 65°C junction temperature (below air cool range)
n128
n129 950
n130
• NeXtScale WCT cools 165W processors with inlet temperatures up to 35°C • No chillers required
Tj,max 900
2x E5-2698a v3
2x E5-2697 v3
2x E5-2690 v3
HPL (GF)
1083
907
813
Power (W)
441
402
385
Perf/Watt (MF/W)
2457
2256
2111
850 20
30
40
50
60
70
80
90
100
Junction Temperature, Tj (C) 31
2015 LENOVO INTERNAL. ALL RIGHTS RESERVED.
Torsten Bloth –
[email protected]
Vinod Kamath
Lower Power Consumption with WCT E5-2695 v2 CPU 115W
air water = 45°C
Socket Power Is relatively flat for junction temperatures in the range that water cooling operates (18-45°C) Increases significantly when junction temperature is in range that air cooling operates.
water = 18°C
Result: Save 5% power when cooling with water up to 45°C
Power saving is about 5% per node Vinod Kamath 32
2015 LENOVO INTERNAL. ALL RIGHTS RESERVED.
Torsten Bloth –
[email protected]
Hot Water Cooling How Direct Water Cooling Makes Your Data Center More Efficient and Lowers Costs Chillers not required for most geographies – Due to inlet water temperature of 18°C to 45°C – Reduce CAPEX for new data centers
40% energy savings in datacenter due to no fans or chillers. Compute node power consumption reduced ~ 10% due to lower component temperatures (~5%) and no fans (~5%) Power Usage Effectiveness PTotal / PIT: PUE ~ 1.1 possible with NeXtScale WCT – 1.1 PUE achieved at LRZ installation – 1.5 PUE is typical of a very efficient air cooled datacenter.
85-90% Heat recovery is enabled by the compute node design – Heat energy absorbed may be reutilized for heating buildings in the winter – Energy Reuse Effectiveness (PTotal – PReuse) / PIT: ERE ~ 0.3 Vinod Kamath 33
2015 LENOVO INTERNAL. ALL RIGHTS RESERVED.
Torsten Bloth –
[email protected]
34
2015 LENOVO INTERNAL. ALL RIGHTS RESERVED.
Torsten Bloth –
[email protected]
35
2015 LENOVO INTERNAL. ALL RIGHTS RESERVED.
Torsten Bloth –
[email protected]
36
2015 LENOVO INTERNAL. ALL RIGHTS RESERVED.
Torsten Bloth –
[email protected]