Direct liquid cooling by bull Laurent Cargemel Server Design & Development June 27th, 2012
Green datacenter characteristics PUE small as possible 1 Energy reuse : warmer liquid may open new usage 65 C outlet temperatures open «central heating» market Reduce liquid cooling footprint within the datacenter Mobull, new racks Today Bull focuses on smaller PUE by using liquid cooling 2 Current DLC offer 2nd generation improvements Research activities
Water cooling & PUE Water cooled blade don t need icecold water Cooling with hot water improves significantly the PUE Cool doors AirCooled 10(20) kw/rack Room 20 C A/C water PUE 1.7 712 C Direct Liquid Cooling 40 kw/rack 60 kw/rack Room 20 C 27 C Room up to 27 C Water 712 C 1419 C Water 3040 C PUE 1.51.6 1.41.5 PUE 1.11.2 A/C 3
Bullx direct liquid cooling technology Cooperative Power Chassis (CPC) Chassis Partitioning of water cooling and AC voltage for safety reasons 54V DC to the chassis 18 bisockets blades Switch IB & Ethernet embedded 7U 100% water cooled fan less Noiseless Hydraulic Chassis (HYC) Provide water cooling for the 5 chassis Redundant 2N at rack level (Cheaper than at blade level ) 100% liquid cooled 42U Bull rack (19 ) HYC 2 Bisockets nodes per blade 4
Bullx dlc Main features Bullx dlc technology allows to implement more powerful blades Double EP node blades, Single EP node blade plus accelerators 2kW blade with 1 EP node plus custom accelerators It supports largest MIC adaptors Maintainability remains as it was in an air flow design Maximum rack water inlet temperature: Direct access to CPU sockets Direct access to Dimms Direct access to disks Type of blade 2 Bisockets nodes blade 1 EP node with GPU CPU power Node power EP 130W 820W 35 C EP 115W 790W 40 C MIC 300W 1400W 40 C > 2000W 20 C 1 EP node with accelerators 5 Water T (customer loop)
Bullx dlc Very accurate Blade cooling control (1) Blade Water inlet temperature is regulated by setting appropriate inlet valve openness Water Redundancy is provided by putting 2 Hydraulic Controller per rack. One HYC is active at a time Regulation worst case is when one active HYC fails: Tcase CPU variation is much smaller than water inlet temperature variation : 23 C Figure shows real test results Influe nce d'une é lév ation de T e au sur la T de s CPU 10 Tcase CPU 8 T d'entrée d'eau dans les lames Degré ( C) 6 4 2 0 2 50 100 150 200 4 T e m p s (s ) 6 250 300
Bullx dlc Blade cooling control accuracy (2) Other source of variation: 34 Start at warm or cold water condition, blade power variations (+/ 1 C around reference temperature) Rac k c onsumption Water flow is regulated depending on the rack configuration 34 T inlet blade Temperatures ( C) Power (kw) W ater tem perature control Scrutation 10s Kp=20 Ki=0,5 Regul 30 C Resulting measured water flow variation : (1,4 à 1,7 l/min) Related Tcase CPU variation (1,4 C) 32 32 30 30 28 28 26 26 24 24 22 22 20 0 2000 4000 6000 8000 10000 Time (s) Blade water flow control (differential pressure) T c a s e C P U e n fo n c tio n d e s v a ria tio n m a x im u m d e d é b it d a n s le s la m e s A rm o i re 4 0 k W 4 5 P l a q u e s fro id e s W i n x e n T e a u c l ie n t 3 5 C From 5 to 1 chassis, and from 1 to 5 chassis 300 5 1,6 E c art Tc as e C P U 1,4 250 1,2 3 100 Number of chassis Differential pressure 2 1 Ecart / nominal ( C) 150 Number of chassis Différential pressure (kpa) 4 200 0,8 0,6 0,4 0,2 50 0 0 1 0 10 20 30 40 50 60 70 80 1,4 7 1,4 5 1,5 1,5 5 D é b it (l/m in ) Time (min) 1,6 1,6 5 1,7 20 12000
Bullx dlc technology 2nd generation (1) Today, a good solution is based on free cooling (warm water system) Tomorrow, a warmer outlet temperature may help for energy reuse Here is the split of the temperature difference between CPU case and water Losses by exchange type Margin 10% Series water path 13% Conduction 45% TIM 13% Exchange w. Fluid 19% 8 2 main candidates for an optimization are the heat spreader geometry and the heat exchange at cold plate level
Bullx dlc technology 2nd generation (2) Example of heat exchange optimization at coldplate level More turbulence involves the heat exchange between cold plate en fluid Several simulations done with different shapes of fins Intensity of turbulence (%) Cold plate bullx Exchange area Fluid path 9 Fluid speed spectrum (m/s)
Bullx dlc research activities (1) Experiment new fluid : Nanofluid 12 kw Blade immersion while saving easy access to the different removable parts 12 kw blade without liquid connectors Copper nanoparticles fluid has been tested : 1 C improvement might be expected Illustration 2: Nanofluides vus au microscope électronique: (a) éthylène glycol + cuivre à 0,5 %; (b) eau + alumine ; (c) eau + or à 2 nm; (d) eau + nanotubes Critère 2: Performances thermiques "Pouvoir caloprteur" à 45 C 14,0 12,0 ΔTio (eau)_45 C ΔTio (T10)_45 C ΔTio (T10/ZnO)_45 C ΔTio (Eau/CuO)_45 C ΔT (inout) [ C] 10,0 8,0 6,0 4,0 2,0 0,0 0,6 0,8 1,0 1,2 1,4 1,6 Débit [l/min] 10 1,8 2,0 2,2 2,4 2,6
Bullx dlc research activities (2) Water cooling with no liquid in the board Advantage : No hydraulic disconnection during maintenance operations (security & cost improvement) Technical stakes : Electronic components are far from the cooling liquid Thermal : The usage of heat pipes is mandatory to drive heat toward the thermal exchange surface of the blade Mechanical : Need perfect alignment between blade and cold plate in order to use very thin high performance TIM Mechanical Test vehicle has been done : first results are promising Poignée amovible Coque de la lame Coque du châssis Zone de contact CPU TIM CPU Plaque froide CPU CPU Effort Dissipateur à caloduc Refroidissement Zone d'appui caloduc Design principle that ensures a uniform compression of the TIM 11 Design of a 400mm high performance heat pipe
Make datacenter easier Current Mobull Racks layout uses warm and cold corridors scheme AC main connection + Fire allarm and protection + Access control Thermal insulation Hydraulic pipes Floor Shock absorber is needed to insure full container transportation: container floor 12