One Dell Way R Round Rockk, Texas 78 8682 www.dell.com DEL LL ENTER RPRISE WHITEPAP W PER THERM MAL DESIG D GN OF O TH HE DEL LL™ POWE P ERED DGE™ M-S SERIE ES K.C. Coxe C Enterp prise Therm mal Enginee ering THIS WHITE PAPER ISS FOR INFOR RMATIONAL PURPOSES P ONLY, AND MAY M CONTAIN N TYPOGRAPH HICAL ERRORS AND A TECHNIC CAL INACCUR RACIES. THE CONTENT ISS PROVIDED AS IS, WITH HOUT EXPRESSS OR IMPLIED WARRANTIES W S OF ANY KIND D. AMD and d Opteron arre trademarkks of Advanced Micro Deevices, Inc. Intel and Xeeon are regisstered trademarks of Intel Co orporation. Other O tradem marks and trad de names maay be used in n this documeent to refer to either e the enttities claimingg the marks and a names or o their produ ucts. Dell disclaims proprietary interest in n the marks and names of others. For more information, contact Dell.. Informatio on in this doccument is sub bject to changge without no otice. 2|Page 1 INT TRODU UCTION N With energy costs rising, energy efficiency e in IT equipmentt has quickly become a major m focus of o Dell customers. One of th he key avenu ues of addresssing energy efficiency att the platform m level lies in the developm ment of a high hly efficient th hermal solution. System cooling c can co onsume a siggnificant portiion of overall system power, but it doesn’t need to. The launcch of the De ell™ PowerEd dge™ M1000ee Modular Enclosure and d PowerEdge M600 and M605 blades maarks an evolu utionary leap in thermal deesign efficien ncy by Dell. The T thermal architecture a w within the system stretches from the layyout of each h blade to Chassis Managgement Conttroller (CMC)) that he fan modules that provvide the airflo ow. This pap per will walkk the user thrrough manages cooling to th some of the t features of o the thermal design and d discuss the underlying principals p and d design goalss that are intend ded to result in one of the most thermaally efficient servers s on thee market. A reader seeking a mo ore detailed description d off the featuress of the PoweerEdge M100 00e and assocciated blades and peripheralss should read the Dell M-Series Architecture Whitep paper availablle at dell.com m. 2 DE ESIGN GOALS G The Dell M1000e M is designed to reliiably operatee at ambient conditions c up p to 35°C (95°°F) and at altitudes of up to 10,000 feet (3,048m). In addition, the system supports s red dundant cooling. The theermal performance requirem ments to meeet such condittions, however, are not tyypical of mosst deploymen nts. A more typ pical data cen nter could be supplying 15-17°C air from f the floor tiles, a much m more benign b operatingg condition. Even a dataa center opeerating at “higher” tempeeratures arou und 25°C offfers a significantt opportunityy for reducingg power-to-co ool vs. a 35°C environmentt. The challeenge that Dell faced in thee developing the t M1000e was w to develo op a solution that is both: • Reliable operation at worstt-case, redund dancy lost conditions • Optimized O for typical data center c conditions w detail how w Dell accom mplished thiss in hardware design as well as in syystem The sections below will thermal management m through the Integrated I Deell Remote Acccess Controller (iDRAC) an nd CMC. 3 CO OOLING G SOLU UTION OVERV VIEW The sections below are intended to provide a general understanding of the major elements of o the thermal solution in the e Dell PowerEEdge M1000ee Modular Encclosure. Thesse sections arre not all-inclusive, but will ho opefully provvide the readeer an appreciaation of the engineering e th hat has gone into the design. 3|Page Figure 1:: Front View of PowerEdg ge M1000e wiith 1 Half-Heigh 16 ht Blade Encllosures Fig gure 2: Rear View V of PoweerEdge M10000e 3.1 What W prov vides th he coolin ng The PoweerEdge M1000 0e is cooled by b 9 custom fan f modules. The fan module was dessigned to meeet the specific co ooling require ements of the M1000e, in ncluding its airflow operatting point and d power efficciency requiremeents. Each fan f is hot sw wappable and d can be indiividually speeed controlled d to help opttimize system co ooling during normal run time or if redu undancy is losst. F Figure 3: Dell M1000e Fan n Module Cooling iss managed byy a combinatiion of blade-level (iDRAC)) and enclosu ure-level (CMC C) controllerss (See Figure 4). The thermal control algo orithm runnin ng on each blades’ iDRAC is i an evolutio onary step forrward from the BMC that run ns on the Delll PowerEdge 9G servers. It has the cap pability to anaalyze the hard dware configurattion on each h blade, the thermal cond dition createed by softwarre load on th he blade, and the ambient temperature t ommunicate the t blade’s sp pecific cooling requiremen nts to of the blade and then co the enclossure. 4|Page The enclo osure (CMC),, in turn, is designed to o interpret each e of the 16 blades’ inputs, as well as temperature sensors in the M1000 0e infrastructure (such as in the IO Modules and thee control pan nel on the front of the M100 00e), and set each fan to the t lowest sp peed possiblee – helping to o minimize airflow and poweer consumptio on – to meett the cooling requirementss of the systeem. The CMC C controls thee fans in zones, with each zo one mapped to t the coolin ng of a group of blades baased on locattion in the ch hassis. This desiggn allows the CMC to increease only thee fans needed d to cool hottter blades orr blades with more power-inttensive hardw ware configurrations, whilee leaving the rest r of the fan ns at a lower power and airflow saving levvel. Fiigure 4: Fan Control C Block k Diagram 3.2 Aiir Manag gement The M1000e makes use u of parallel air paths to cool the blades, IO Modules, M and d power sup pplies. h of the major subsystem ms receives ambient a air to t help avoid pre-heating the Because of this, each modules in the back of the system. Providing ambient air to the IO Modules and d Power Sup pplies, despite th heir locationss in the rear of the system, was a maajor focus of the early sysstem architectural engineering, with the e goal of allowing the modules m to consume lesss airflow while w still meeeting ure requiremeents. component temperatu The serveer modules arre cooled with h traditional front-to-backk cooling. As shown in Figgure 5, the fro ont of the system m is dominatted by inlet area for the in ndividual blad des (green higghlighted areea). The air passes p through the server mo odules, througgh vent holess in the midp plane, and is then t drawn in nto the fans which w t air from the chassis. There are plenums p both h upstream of o the midplane, between the exhaust the 5|Page midplane and the blades, and dow wnstream of the midplan ne, between the t midplanee and the fan ns, to distributee the cooling potential p from m the three columns of fan ns across the server modu ules. Serrver Module e Inlet (M1 1000e frontt view) Server Mo odule Cooling Air Proffile (M M1000e side e view) Figure 5: 5 Dell Blade Server Modu ule Cooling Path P The Poweer Supplies, lo ocated in thee rear of the system, use basic front-tto-back coolin ng, but draw their inlet air frrom a duct located beneatth the server modules, as seen s in the green highlight in the left side of Figure 6. This insuress that the power supplies receive amb bient temperature air, wh hich minimizees the v of airr needed to co ool the power supply. required volume Power Sup pply Inlet ((M1000e fro ont view) Power Sup pply Cooling g Air Profile e (M10 000e side view) v Figuree 6: M1000e Power P Supply y Cooling Patth The IO Modules use a bypass duct to draw amb bient air from m the front of o the system to the IO Module inlet, as seen s in Figure e 7. This duct is located above the seerver modulees (again high hlighted in grreen). This cool air is then drawn d down through the IO Modules in a top-to-b bottom flow path and intto the b the midplane m and d fans. From this plenum,, the air is exxhausted from m the system. The plenum between CMC and Integrated Ke eyboard/Videeo/Mouse (iKVM) are also cooled by airr in this flow path. p 6|Page IO Module Inlet (M M1000e fron nt view) IOM Lo ocations and d Airflow Direction (M1000e back b view) IO Coo oling Air Pro ofile (M100 00e side vie ew) Figu ure 7: M1000ee IO Module Cooling C Path h 4 DE ESIGN CONSID C DERATIIONS AFFECT A TING PO OWERTO O-COOL L 4.1 Im mpedanc ce, Airflo ow, and Fan Effficiency y Developin ng a fan to cool the M1 1000e and provide p scalability for thee next severral generatio ons of computing architecturre provided a unique challenge. The faan needed to o meet an aggressive operrating point, yett fall within a much tighter power budget than off-the-shelf fans could provide. o p Thiss was accomplisshed by identtifying a fan efficiency e targget that requ uired custom technology to be develop ped to fit the neeeds of the M1 1000e. This section s will prrovide a brieff overview of what fan effiiciency is, and d why it is criticaal to an efficie ent thermal design. d There is somewhat s of a misconcepttion about ho ow fans perfo orm and are specified. s A fan f purchased at a local electtronics store for a desktop p PC that is ad dvertised as an a “80 CFM” fan f will not move m 80 CFM of air when insttalled in a PC.. This is called a “free air” measurement of the fan performancee, and is how much 7|Page air the fan n moves whe en sitting on a table with nothing n occluding the inlett or outlet (su uch as the intternal components of a computer). This method of describing a faan is so comm monplace thaat fan vendorrs will even list this value on n their datash heets, even though t it is only o a partiall specification n of a fan’s actual a airflow ou utput. The half of o the fan equ uation that isn’t often disccussed is impeedance. Impedance is thee opposition to t the flow of air, a is measurred in pressu ure, and in US U customaryy units is rep presented byy inches of water, w (Pascals in n SI derived units). u The more m impedance that a faan is loaded with, w the lesss airflow it acctually moves. A fan in free air has zero impedance. In an electrronics system m, however, the t fan runs at an operatingg point of ‘airfflow at a giveen impedancee.’ The chartt below is typ pical of a fan that might be b found in a desktop com mputer or tow wer server. In this chart, thee free-air ratin ng is 116 CFM M, but the acttual operatingg point is closser to 80 CFM M at 0.19 inch hes of water (in.w.g.). 0.6 Fan Curve Impedance (in.w.g.) 0.5 0.4 Operatting Point 0.3 Freee Air 0.2 System Imp pedance Curve 0.1 0 0 20 40 60 80 100 120 140 Airflow (CFFM) Figu ure 8: Examplle Fan Perform mance Curvee Why is th his importantt? A thermaal design enggineer knowss how much airflow is reequired to co ool an electroniccs enclosure and can measure the im mpedance of the enclosurre to determ mine the operrating point required of the system s fan. The operatin ng point of th he system fan is a significcant driving factor f for how much m power is needed to cool the systtem. The op perating pointt of a fan actu ually describees the amount of o energy (wh hich can be represented r as power) th hat needs to be put into the air to fo orce it through a system. Wh hen compared d to the electtrical power required r to ru un the fan, th he efficiency of o the fan can bee determined d. 8|Page η= Worrk out Operrating Flow wrate × Operrating Presssure = Worrkin Innput Powerr Eq. 1 Fan efficieency is depen ndent on how w the fan is built b and dessigned, from the shape off the blades to t the quality off the motor, and a can vary significantly,, even from one o fan to an nother in the same form faactor. What doees all this meaan with respeect to cooling power? There are two im mportant poin nts: 1. A higher efficiency fan wiill consume proportionally p y less powerr than a low efficiency fa an. A higher efficien ncy fan design n, leading to a lower pow wer fan, results in more power availab ble for co omputing. Alternatively, A o poweer consumptio on, which caan be it can result in lower overall trranslated into o cost savingss in the deployment. Thee higher efficciency may co ome at the co ost of m more expensivve componen nts to build the t fan, so itt is critical th hat a thermaal design enggineer seelect a fan with a material cost approp priate to the amount of en nergy that may be saved in i the deeployment. 2. Iff a system ca an be designeed with a low wer work outp put requirem ment, the fan will consume less po ower. We have discussed how the operating point is determined byy the airflow w and im mpedance of the system. While the volume of airflo ow required to t cool an eleectronics enclosure iss somewhat flexible, it often o needs to be maintained at a minimum m levvel to contro ol the teemperature rise r through the t system (m more on that later). Assuming a minim mum fixed airflow, im mpedance is the t other variiable that can n affect the po ower-to-cool. These principals guide ed not only development d of the M100 00e fan, but the t close relaationship witth the impedancce drove a lo ow-impedance design as well. The effficiency of the M1000e fan f is significcantly higher thaan typical 92m mm or 80mm m form factor fans used in many m PCs currrently on thee market. Thiis was achieved through opttimization off the aerodyynamics, mottor, and mottor controller design witthin a custom fo orm factor. This fan iss coupled with a blade deesign with up p to half the impedance of o competitivve products on o the market an nd a chassis midplane thaat underwentt several desiign iterationss to help optimize airflow. The blade encclosure and layouts were studied exteensively priorr to building the first funcctional system ms to understan nd what facto ors contributted to the impedance of the t module, and targetingg certain areaas for improvem ment. These layout studies resulted in i componen nt placements made for airflow, a as well w as subtle chaanges in the industrial dessign of the blaade, all to help drive lowerr impedance and a a lower power p thermal design. d 4.2 Aiirflow an nd Temp perature e Rise There aree numerous benefits b a cusstomer can reealize by loweering fan speeds and increeasing the exxhaust air tempeerature from a server. To illustrate this effect, conssider the simplification off a Dell PowerEdge 1950 servver into basic thermal build ding blocks: an a airflow ressistance, a heeat source, a power p distrib bution system, and a fan (Figu ures 9 and 10 0). 9|Page Figure 9:: PowerEdge 1950 Figurre 10: Block-D Diagram Simp plification off a Server's Airflow A System m me the serverr is under a constant c load d that requirees a 300W input. With th he fan In this exaample, assum running at a 9,000 RPM M, there is an 8°C rise thro ough the systtem. By deccreasing the fan f speed to 3000 RPM, the exit air tem mperature increases to 22°C above thee inlet tempeerature, but the t system power p decreasess 55W! 10 | P a g e Figure 11: 1 Comparison of Fan Speed, Power Draw, D and Exh haust Tempeerature at Hig gh (top) and Low L (botto om) Fan Speed ds The reaso on for the draamatic power decrease beetween the high h and low fan speeds in Figure 11 iss that fan power and RPM do o not have a linear l relation nship. In factt, power chan nges as a cubiic function off a fan speed chaange, so small decreases in i fan speed can have sign nificant impacts on the po ower consum mption of a fan (EEq 2). Poweriniitial ⎛⎜ RPM M initial = ⎜ RPM Powerfinal M final fi ⎝ ⎞ ⎟ ⎟ ⎠ 3 (Eq 2) As an example, and ussing (Eq 2), a 10% reductio on in fan RPM M can result in n a 27% reduction in fan power p consumpttion! Because of this relattionship, there was a consscious effort made in the M1000e design to minimize fan speed and increase exhaust air temperature, t , with the intended beneefit of significcantly reducing system powe er and airflow consumption. All of the t componeents downstrream in the warm w airflow haave been desiigned to meet the high tem mperature requirements of o this design condition. o the high teemperature rise r that strettches beyond d the enclosu ure-level efficciency An expectted benefit of improvem ments in the M1000e is th hat the capaccity of the air conditionerrs in a data center c can acctually increase with w a higherr temperaturee rise as the CRAC (Computer Room Air A Conditioneer) operates closer c to its opttimum return n temperaturre. Care musst be taken when w selectin ng air conditioning to pro operly match it to a datace enter deployyment, but the t higher reeturn tempeerature to th he CRAC (exxhaust temperature from the system) can have benefitss beyond pow wer-to-cool att the IT enclosure level. 11 | P a g e 4.3 Co ompone ent Selec ction With a ch hassis designe ed for therm mal efficiency,, there are sttill several op pportunities for a custom mer to lower pow wer and airflo ow consumption by selectiing easier-to--cool compon nents and blad de configurattions. 0, for examplle, is available with processors rangingg from 40W (Intel® Xeon® 5148LV) to 120W 1 The M600 (Intel Xeo on X5355) the ermal design power levelss. In addition to the pow wer savings asssociated witth the processorr power, the lower preheaat placed on downstream m componentss by the 5148 8LV as oppossed to the X5355 5 provides an n opportunityy for airflow reductions ranging r from 5% to 40% at a a 25°C am mbient temperature, depending on other components c in the blade. Memory configuration c ns can also afffect cooling power and airflow a requirrements. Nottably, the M6 600 is more efficient coolingg memory wh hen 4 higher capacity DIM MMS are useed in place off 8 lower cap pacity DIMMs to o achieve a given memo ory capacity.. Airflow reeductions of up to 25% can be achiieved, depending on memorry capacity reequired and other o compo onents in the blade. Thiss is a result of o the ed in every other o slot in the t memory bank, b and thee resulting reeduction in DIMMDIMMs beeing populate to-DIMM heating. 5 CA ASE STU UDY: POWERE EDGE M600 M V M60 VS. 05 VS. PO OWERED DGE 19 950-III What doees the custom mer get for a well-designeed blade servver with cooliing efficiencyy designed-in from the produ uct’s inceptio on? Considerr an M600 deeployment an nd a deploym ment of 16 PowerEdge 19 950 III servers with similar hardware, particularly: • (2 2) Intel Xeon 5130 5 processsors • (4 4) 2GB Fully Buffered B DIMM Ms c an M605 M deploym ment of 16 blades with And also consider • (2 2) AMD Opterron™ 2218 HE H processorss • (4 4) 2GB DDR2 2Rx4 DIMMs 12 | P a g e Table 1: Compariso on of Deploym ment of 16 M600 M Blades vs. v M605 Blad des vs. 16 PowerrEdge 1950-IIII Servers Num mber of Servers Racck-Space to Deploy D Serrver Density Tottal AC Power Required1 Fan Power1, 2 Airfflow Consum med by Deployyment3 Ave erage Airflow w per Server3 Po owerEdge 19 950 III 16 6 16 6U 1.0 0 Servers/U 47 785W 53 30W 49 95 CFM 31 1 CFM PowerEd dge M600 16 10U 1.6 Serveers/U 3835W 80W 370 CFM 23 CFM Pow werEdge M60 05 16 10U 1.6 Servers/U S 3660 0W 65W W 340 CFM 21 CFM Po ower represen nted in AC Watts W “at the wall” after power p supply y efficiency iss considered and d assumes a SPECjbb S work kload. 2 Fa an power is in ncluded in to otal AC power, but is brok ken out here for f compariso on of cooling efficienccy. 3 Ai irflow consum mption is bassed on a 25°C C ambient tem mperature to the server an nd will vary as temperature varies. v 1 From Tab ble1, a deployyment of 16 M600 bladees deploymen nt consumes 19% less po ower and 29% % less airflow th han the PowerEdge 1950 IIII deployment. Of the 845 5W in power savings, 450W W is a result of o the efficiencyy in the coolin ng solution. The PoweerEdge M605 deployment shows additional power and a airflow savings, madee possible by more efficient cooling c in the e parallel layo out of the CP PU and memo ory in the airrflow, as welll as power saavings associated d with DDR2 as compared to FBD memory. 6 CO ONCLUS SIONS Designingg the cooling solution forr IT equipment can be ass simple as selecting s a faan from a catalog, installing it in a computer, and turrning it on. A basic appro oach like that, however, will w not resullt in a solution that t is optim mized for po ower, airflow w, or coolingg. The coolling solution deployed in n the PowerEdgge M1000e Modular M Enclosure and itts associated d blades werre developed with the go oal of maximizin ng energy effficiency and minimizing airflow requirements wh hile still supp porting worstt-case operatingg conditions. Tools such as high efficciency fans, low-impedancce system deesign, advancces in fan control, and systems management m a low airflo ow/high temperature rise are designed d to help min nimize the poweer consumptio on of the M1 1000e therm mal solution and a allow mo ore of the customer’s avaailable power to be dedicated d to computin ng, as opposed to cooling. 13 | P a g e