WO2012175144A1

WO2012175144A1 - Blade server system and method for power management

Info

Publication number: WO2012175144A1
Application number: PCT/EP2011/060648
Authority: WO
Inventors: Julius LUHUR; Bernhard Schräder
Original assignee: Fujitsu Technology Solutions Intellectual Property GmbH
Current assignee: Fujitsu Technology Solutions Intellectual Property GmbH
Priority date: 2011-06-24
Filing date: 2011-06-24
Publication date: 2012-12-27
Anticipated expiration: 2013-12-24

Abstract

The invention relates to a blade server system (300), comprising an enclosure (310), at least one power supply unit (332, 334, 336), at least one management blade (350) having at least one controller (352), and at least one blade server (342, 344, 346, 348) having at least one processor (362, 364, 366, 368) and a remote management controller (372, 374, 376, 378). The controller (352) of the management blade (350) is configured to query the at least one power supply unit (332, 334, 336) to determine a total available power for the blade server system (300), to determine a power limit for the at least one blade server (342, 344, 346, 348). The remote management controller (372, 374, 376, 378) of the at least one blade server (342, 344, 346, 348) is configured to increase a counter associated with the blade server if an associated power monitoring means indicates that a current power consumption of the at least one blade server exceeds the stored power consumption limit. The controller (352) of the management blade (350) is further configured to determine a new power consumption limit for the at least one blade server (342, 344, 346, 348) based on the total available power and the value of the counter associated with the at least one blade server (342, 344, 346, 348) after a predetermined monitoring interval. The invention further relates to a method for power management and a data processing system (100).

Description

Blade server system and method for power management The present application relates to data processing systems in general. More specifically, the present invention relates to devices, systems and methods for power management within data processing systems such as blade server systems. Data processing systems are used for many applications and in many context. Depending on their use, the required processing power of data processing systems vary greatly. Furthermore, over time, more processing power may be required to satisfy the need of a specific application. For example, if a busi- ness grows, server computers used to store and process busi^¬ ness critical data may need to be upgraded in order to proc^¬ ess the increased amount of data. As a consequence, many data processing systems are designed in a modular fashion. One example of such a modular design of a data processing systems are so-called blade server systems. A blade server system typically comprises a number of so-called blade serv^¬ ers, each of which comprises at least one processor for proc^¬ essing data and associated working memory. Individual blade servers may be plugged into mounting positions typically ar^¬ ranged in a front portion of a so-called blade enclosure in order to increase the processing power of the overall blade server system. Furthermore, a blade server system typically comprises a num^¬ ber of so-called infrastructure modules, such as power supply units, network switches, storage modules and the like. The infrastructure modules supply common infrastructure services for all or at least some of the blade servers. Typically, at least one power supply unit and one network connection module are placed in mounting positions arranged in a back portion of a blade server enclosure.

Furthermore, known blade server systems typically also com^¬ prises a so-called blade management module for monitoring, configurating and administering the various plug-in modules of the blade server system. Finally, all of the modules of a blade server system are typically connected by a so-called midplane arranged in the blade enclosure.

In a conventional blade server system, the blade management module determines a current system configuration, typically during a start-up of the blade server system. For example, the blade management module may query the number and output power of all power supply units arranged in a blade server system. Furthermore, the blade management module may query the number and type of blade servers and other plug-in mod^¬ ules of the blade server system in order to determine their overall power consumption. The power consumption of individual modules may be provided by the modules themselves, may be determined by the blade management unit from a database stor^¬ ing module-specific power consumption values or may be speci^¬ fied manually by means of a user interface.

The blade management unit may then determine how many of the blade servers and other modules of the blade server system can be powered up using the available power supply units and activate them accordingly. If there is insufficient power provided by the power supply units in order to supply all blade servers, one or more server modules may not be acti^¬ vated . The described approach to power management ±n mouu± ar ua ua processing systems may result in an inefficient use of system resources. For example, one or more blade server modules may be deactivated because other blade server modules specify power requirements that exceed their actual power consump^¬ tion. Furthermore, modification of the allocation of a system power to individual components may result in the need for a manual configuration. A manual configuration of the power allocation within an electronic data processing system may be difficult and error prone and therefore may lead to further inefficiencies .

Therefore, there exists a need for improved apparatuses, sys^¬ tems and methods for power management in data processing sys- terns.

According to a first aspect of the invention, a blade server system is described. The blade server system comprises an en^¬ closure having a plurality of first mounting positions to ac- commodate a plurality of power supply units and a plurality of second mounting positions to accommodate a plurality of blade modules. The blade server system further comprises at least one power supply unit accommodated in one of the plu^¬ rality of first mounting positions, at least one management blade and at least one blade server accommodated in one of the plurality of second mounting positions. The management blade comprises at least one controller and the at least one blade server comprises at least one processor, a memory ac^¬ cessible to the at least one processor and a remote manage- ment controller. The controller of the management blade is configured to query the at least one power supply unit to de^¬ termine a total available power for the blade server system. The controller of the management blade is further configured to determine a power limit for the at leaau one jjr aue server and to provide the determined power limit to the remote man^¬ agement controller of the at least one blade server. The re^¬ mote management controller of the at least one blade server is configured to store the power limit provided by the con^¬ troller of the management blade and to increase a counter as^¬ sociated with the blade server if an associated power moni^¬ toring means indicate that a current power consumption of the blade server exceeds the stored power consumption limit. The controller of the management blade is further configured to determine a new power consumption limit for the at least one blade server based on the total available power and the value of the counter associated with the at least one blade server after a predetermined monitoring interval.

The described blade server system allows an adaptive and dy^¬ namic reallocation of total power of a blade enclosure to different blade servers. Therefore, the power allocation within the blade server system gracefully approaches a demand driven power requirements without manual interference. Fur^¬ thermore, the described system results in a very low overhead with respect to the management blade.

According to an embodiment, the at least one blade server comprises a power controller, wherein the power monitoring means are comprised in the power controller. Furthermore, the power controller may be adapted to provide a throttle signal to the processor. According to a second aspect of the present invention, a me^¬ thod for power management in a data processing system comprising a plurality of processing units and at least one power supply unit is provided. The method comprises the steps of determining a total available power pru v iueu jjy Liie au least one power supply unit, determining and storing an initial power limit for each one of the plurality of processing units, repeatedly comparing the current power consumption of each one of the plurality of processing units with the re^¬ spective stored power limit, increasing a counter associated with one of the plurality of processing units if the current power consumption of said one of the plurality of processing units exceeds the respective stored power limit, and deter- mining and storing a new power limit for at least one of the plurality of processing units after a predetermined period of time based on the total available power and the associated counter value of said one of the plurality of processing units .

In one embodiment, the method further comprises the steps of throttling one of the plurality of processing units if a cur^¬ rent power consumption of said one of the plurality of proc^¬ essing units exceeds the respective stored power limit.

In one embodiment of the described method, the step of deter^¬ mining and storing a new power limit comprises increasing the new power limit for said one of the plurality of processing units with respect to the previously stored power limit if the associated counter value exceeds a predetermined first threshold .

In another embodiment of the described method, in the step of determining and storing a new power limit, new power limits for at least a first and a second one of the plurality of processing units are determined and stored. The new power limit for the first one of the processing units is increased with respect to the associated previously stored power limit and the new power limit for the second one o r Liie pruce a a iny units is decreased with respect to the associated previously stored power limit, if the counter value associated with the first one of the processing units exceeds a predetermined first threshold and the counter value associated with the second one of the processing unit lies below a second thresh^¬ old. In such an embodiment, power previously allocated to a second processing unit can be transferred to a first process^¬ ing unit.

In another embodiment, the step of determining and storing a new power limit comprises a first and a second phase, wherein in the first phase, the amount of power available for distri^¬ bution is determined based on the total available power and, in a second phase, the determined power available is distrib^¬ uted between the plurality of processing units based on the counter values associated with the plurality of processing units . In one embodiment, in the second phase, the following two steps of are performed repeatedly while the available power is greater than zero. In a first step, the power available is reduced by an additional power budget and in a second step, a new power limit for one processing unit having an associated counter value exceeding a first threshold value is determined based on a sum of the current power limit of that processing unit and the additional power budget.

In a further embodiment, in the second phase, new power lim- its for the plurality of processing units are determined in a sequence determined by a priority level associated with the respective processing unit. According to different embodi^¬ ments, the priority level for each processing unit may be provided using a user interface of a management UII±L or may be determined in order of the associated counter values of the plurality of processing units by a management unit. In a further embodiment, in the first phase, the power avail^¬ able is determined based on the difference between the total available power and the sum of all power limits associated with the plurality of processing units. In a further embodi^¬ ment, a preset power consumption of other components of the data processing system is subtracted from the total available power .

In a further embodiment, in the first phase, for each proc^¬ essing unit having an associated counter value falling below a second threshold, a released power budget is added to the power available, and, in the second phase, a new power limit for said processing units having an associated counter value falling below a second threshold value is based on the cur^¬ rent power limit of the respective processing unit reduced by the released power budget.

According to a third aspect of the present invention, an electronic data processing system is described. The elec^¬ tronic data processing system comprises a plurality of proc- essing units, each processing unit comprising at least one processor. The system further comprises at least one power supply unit for providing at least one supply voltage, and a management unit, the management unit comprising at least one controller operatively coupled with the plurality of process- ing units and the at least one power supply unit. In the electronic data processing system, the management unit is adapted to determine a new power consumption limit for at least one of the plurality of processing units based on a to- ₀

tal available power provided by the at leaau one power a upp±y unit and a number of times a previously determined power con^¬ sumption limit was exceeded by said one of the plurality of processing units in a predetermined monitoring interval.

The described data processing system allows an adaptive, dy^¬ namic and incremental reallocation of power. Therefore, the power allocation within the data processing system gracefully approaches a demand driven power requirements without manual interference. Furthermore, the described system results in a very low overhead with respect to the management unit.

In one embodiment, the data processing system further comprises a plurality of power monitoring means. Each one of the power monitoring means is configured to compare a current power consumption of an associated processing unit of the plurality of processing units with a current power consumption limit provided by the management unit for the associated processing unit.

According to further embodiments, the processing units are configured to throttle their power consumption if the associ^¬ ated power monitoring means indicate that the current power consumption of the processing unit exceeds the current power consumption limit.

According to different embodiments, the processing units may be configured to throttle their power consumption by assert^¬ ing a throttling control signal for reducing the clock fre- quency of the at least one processor, by providing an ACPI control parameter and reducing the workload of a processor in response to the provided ACPI control parameter or by initi^¬ ating a shutdown operation of a processing unit. In a further embodiment, the management unit is adapted to allocate, in a first step, power to other infrastructure com^¬ ponents of the electronic data processing system. The power management unit is further adapted to allocate, in a second step, the remaining available power to the plurality of proc^¬ essing units based on the number of times the previously de^¬ termined power consumption limit was exceeded by a respective one of the plurality of processing units.

Further advantageous embodiments are described in the at^¬ tached claims and the description of detailed embodiments be^¬ low . The invention will be further described using a detailed de^¬ scription of presently preferred embodiments with respect to the attached drawings .

Figure 1 shows an electronic data processing system in accor- dance with an embodiment of the invention.

Figure 2 shows a flow chart of a method for power management in accordance with an embodiment of the invention. Figure 3 shows a schematic diagram of a blade server system in accordance with an embodiment of the invention.

Figure 4 shows a detailed flow chart of a method of power management for a blade server system.

Figure 5 shows the allocation of power within a blade server system before the maximum available power capacity of the blade server system is reached. Figure 6 shows the allocation of power within a blade server system once the maximum available power capacity is reached. Figure 7 shows a more detailed flow chart of a method for adaptive power budgeting.

Figure 8 shows flow chart of a method for power allocation in case the available power is unbounded.

Figure 9 shows a flow chart of a method for power allocation in case the available power is bounded.

Figure 1 shows an electronic data processing system 100 ac- cording to a first embodiment of the invention. The data processing system 100 comprises two power supply units 112 and 114. Furthermore, the data processing system 100 com^¬ prises three processing units 122, 124 and 126. The power supply units 112 and 114 and the data processing units 122, 124 and 126 are connected by means of a power distribution system 130 comprising a power management unit 132. Power monitors 142, 144 and 146, arranged between the power distri^¬ bution system 130 and the data processing units 122, 124 and 126, respectively, monitor the actual power consumption of the respective data processing unit 122, 124 or 126.

The components shown in Figure 1 may be arranged on a common printed circuit board (PCB) or may be arranged on different printed circuit boards of individual modules of the data processing system 100 such as a server computer. Some of the components shown in Figure 1 may be integrated with other components. For example, the power monitors 142, 144 and 146 may be build-in current monitors of a processor of the data processing units 122, 124 and 126 or, alt tiv ly, may jje integrated into the power distribution system 130. The compo^¬ nents of the data processing system 100 may be connected by one or more point-to-point connections, as indicated in Fig- ure 1, or may be connected by means of appropriated power and communication busses. The power management unit 132 may re^¬ trieve component-specific power values stored in the power supply units 112 and 144, as well as the data processing units 122, 124 and 126. Alternatively, the power management unit 132 may be connected to a non-volatile memory system for storage of default power values for some or all components of the data processing system 100. The operation of the data processing system 100 in general and the power management unit 132 in particular will be described next with reference to the flow chart presented in Figure 2.

Figure 2 shows a schematic flow chart of a method of power management as performed in the data processing system 100. In a first step 210, initial power limits PL are determined. For example, the power management unit 132 may query the system configuration in order to determine the amount and type of power supply units and data processing units present in the data processing system 100. The individual components 112, 114, 122, 124 and 126 may either explicitly provide informa- tion about the power provided or consumed or may return a type identifier used by the power management unit 132 to identify corresponding power values in a table with stored system configuration information. The initial power limits PL determined in step 210 may be determined based on the minimum or maximum power consumption of the data processing units

122, 124 or 126. Alternatively, the total power provided by the power supply units 112 and 114 may be evenly distributed between the data processing units 122, 124 and 126. The de- termined initial power limits PL may be suoreu e±Liier jjy uiie power management unit 132 or may be provided to the power monitors 142, 144 and 146 for monitoring. In a further step 220, the power monitors 142, 144 and 146 continuously monitor the power consumption P of the associ^¬ ated data processing units 122, 124 and 126, respectively. For example, the power monitors may compare an electric cur^¬ rent provided to each one of the data processing units 122, 124 and 126 with a threshold value based on the initial power limit PL.

If a current power consumption P exceeds the respective power limit PL, in a step 230, a respective counter associated with the respective data processing unit 122, 124 or 126 is in^¬ creased, before the method continues with step 240. Other^¬ wise, the method proceeds directly with method step 240.

In step 240, the power management unit 132 checks whether a current evaluation period has expired. In particular, it may compare a time t elapsed since the last determination of power limits PL with an monitoring interval T between subsequent re-evaluations of power limits PL. If the monitoring interval has not yet elapsed, the method continues, after an optional short delay of a few microsecond, with step 220. Otherwise, the method proceeds to step 250

In step 250 the total available power is determined. For ex^¬ ample, the power management unit 142 may query the total power available from the power supply units 112 and 114. In this way, the dynamic addition or removal, function or fail^¬ ure of individual power supply units may be detected even during operation of the data processing system 100. Alterna- tively, in a simple version of the data pruce a a iny a y a uem ι υ υ without a so-called hot-plug capability, the total available power determined during boot-up of the data processing system may be used as total available power.

In a subsequent step 260, new power limits PL are determined based on the current power limits PL and the determined available power. For example, if the power monitor 142 de^¬ tected that, in the system shown in Figure 1, the data proc- essing unit 122 has exceeded its current power limit PL twenty times within the monitoring interval while the power monitor 144 has only detected a single occurrence of the data processing unit 124 exceeding its previous power limit PL and the power monitor 146 has not detected any occurrence of ex- ceeding a preset power limit PL, the power management unit

132 may re-allocate part of the power previously allocated to the data processing unit 126 to the data processing unit 122. In consequence, in the given example, the new power limit PL for the data processing unit 122 may be increased by a preset amount, for example 20 watts, while the new power limit PL for the data processing unit 126 may be reduced by the equiv^¬ alent amount of a power. In the described example, the power limit PL of the data processing unit 124 remains unchanged in the next monitoring interval. After the new power limits PL have been determined in step 260, the method continuous in a step 220.

In various embodiments of the invention, different means may be applied in order to enforce the set power limits PL. For example, the individual data processing unit 122, 124 and 126 may react to the determined power limits PL by limiting their workload, reducing their operating frequency or the like. Alternatively, the power distribution system 130 may comprise means for actively limiting the provided power LO Liie pre ae u power limit. For example, the power distribution system 130 may limit the amount of current available to the individual data processing units 122, 124 and 126 or may deactivate any- one of the data processing units 122, 124 and 126 either temporarily or permanently in order to enforce the preset power limits. Alternatively, it may be acceptable for the data processing units 122, 124 and 126 to exceed their allocated power budget for a short time, e.g. until the end of the next evaluation period T.

Figure 3 shows a blade server system 300 according to a sec^¬ ond embodiment of the present invention. The blade server system 300 comprises a blade enclosure 310. A midplane 320 is arranged in the enclosure 310. he midplane 310 comprises a system management bus 322 and a power distribution network 324. The enclosure 310 further comprises a number of first mounting positions for the accommodation of power supply units and other infrastructure components. In the system shown in Figure 3, three power supply units 332, 334 and 336 are plugged into the enclosure 310. Furthermore, the enclo^¬ sure 310 comprises a number of second mounting positions ar^¬ ranged on the opposite side of the midplane 320 for the con^¬ nection of different blade units. In the system shown in Figure 3, four blade servers 342, 344, 346 and 348 are arranged in the blade server system 100. Furthermore, a management blade 350 is arranged in a further second mounting position and plugged into the midplane 320. In the described embodiment, each blade server 342, 344, 346 and 348 comprises a processor 362, 364, 366 and 368, respec^¬ tively. In addition, each blade server 342 to 348 comprises a remote management controller 372, 374, 376 and 378. Further- more, each blade server 342, 344, 346 ana J¾O com is s a power controller 382, 384, 386 and 388, respectively. The blade server 342, 344, 346 and 348 may comprise further com^¬ ponents, such as additional processors, memory modules, mem- ory devices such as hard disks, and the like which are not shown in Figure 3 for reasons of representational simplicity.

The management blade 350 comprises a microcontroller 352 and a non-volatile memory 354. In the non-volatile memory 354, a system management table comprising configuration data of all components of the blade server system 300 is permanently stored for use by the microcontroller 352. The microcontrol^¬ ler 352 can communicate with the remote management control^¬ lers 372, 374, 376 and 378 of the blade server 342, 344, 346 and 348 and may also query the capabilities of the power sup^¬ ply units 332, 334 and 336.

Next, the operation of the blade server system 300 is ex^¬ plained with respect to the flow chart of Figure 4. In Figure 4, the operations performed by the management blade 350 are shown on the left-hand side. Furthermore, the operations per^¬ formed by an individual remote management controllers, for example the remote management controllers 372 of the blade servers 342, are shown on the right-hand side.

In an initial phase 410, the enclosure 310 of the blade server system 300 is powered on in a first step 412. In a subsequent step 414, an initial chassis power budget calcula^¬ tion will be performed for the enclosure 310. For example, in step 414, the number, type and/or output power of the avail^¬ able power supply units 332, 334 and 336 may be determined. In addition, a power redundancy mode configuration may be checked by the management blade 350. For example, the blade server system 300 may be configured in a ao-ca±±eu z-t-i- configuration, wherein two power supply units 332 and 334 are used to provide the power required by the blade server system 300 while the power supply units 336 is kept as hot replace- ment unit in case either one of the power supply units 332 or 334 fails during operation. In this case, assuming that each one of the power supply units 332, 334 and 336 has a nominal output power capacity of 500 W, a total available power of 1000 W is determined in step 414.

Optionally, in step 414 the power consumed by unmanaged in^¬ frastructure units of the blade server system 300 is sub^¬ tracted from the total available power. For example, I/O com^¬ ponents such as network switches or storage area network (SAN) adapters plugged into the midplane 320, which have a more or less constant power consumption, are identified.

Based on the number and type of identified components, the management blade 350 may determine an amount of power con^¬ sumed by the unmanaged components of the blade server 300.

After the infrastructure units of the blade server system 300 including the power supply units 332 to 336 have been powered up, the management blade 350 will query some or all blade server 342, 344, 346 and 348 arranged in the second mounting positions. For each identified blade server, the respective remote management controllers 372 to 378 will report its ini^¬ tial power consumption value based on a factory-set value. The values provided in step 416 may correspond to a minimum power requirement of the blade server 342 required to boot up. For example, the reported initial power consumption value may correspond to the operation of a single processor of a multi processor blade server required to start up the blade server 342. In step 418, the management blade 350 compares the received initial power consumption values with the total available power budget. If the total available power budget is suffi- cient in order to activate the blade server 342, it will send a start command to the blade server 342. Otherwise, the man^¬ agement blade 350 does not issue a start command to the blade server 342. Furthermore, the management blade 350 may con^¬ tinue the process for the next blade server 344. Meanwhile, the blade server 342 may be held in a standby state.

In a step 420, after receiving a start command, the remote management controller of the blade server 342 initiates the power on sequence by activating a single processor of the blade server 342 in the given minimum configuration. Once the power on sequence is completed, in a step 422, the remote management controller 372 may calculate a real power budget value based on the actual configuration of the respective blade server 342. For example, the remote management control^¬ ler 372 may detect that a second processor is present in the respective blade server 342 and update its real power budget value correspondingly. The calculated real power budget value is provided back to the management blade 350 which recalcu^¬ lates its current power budget using the given real power budget value in a step 424.

The steps 416 to 424 are repeated for each blade server de^¬ tected in the blade server system 300. If the power budget does not suffice to power on all detected blade servers 342, 344, 346 and 348, some of the blade servers 342, 344, 346 and 348 may remain in a standby state and may only be activated once sufficient power becomes available for them. In a subsequent operational phase 430, the power a±±ocaueu LO individual blade servers 342, 344, 346 and 348 may be adapted dynamically . In a first step 432, the management blade 350 reads a con^¬ figuration setting determining whether adaptive power budgeting is enabled or not. If no adaptive power budgeting is en^¬ abled, the initially calculated power budget remains fixed in the operating phase 330. Otherwise, in a step 434, a maximum blade power limit PL for each blade server 342, 344, 346 and 348 is determined based on the initially computed power budget .

In a step 436, the value determined by the management blade 350 in step 434 for the blade server 342 is stored by the re^¬ mote management controller 372 or power controller 382. From then on, in a step 438, the power controller 382 monitors the power consumption of the blade server 342 independently from the management blade 350 in a continuous loop.

In a step 440 the current power consumption P of the blade server 342 is compared with the provided power limit PL. If the current power P exceeds the stored power limit PL, in a step 442, a control signal for throttling the processor 362 or one or more further processors of the blade server 342 is provided. For example, the so-called PROCHOT# pin of an Intel processor can be asserted which will force the respective processor to reduce its current power consumption with almost immediate effect. Alternatively, other means of reducing the power consumption of the blade server 342can be employed. For example, the processor 362 may be switched into a higher P- State using the ACPI-Interface . Furthermore, in step 444 a counter representing the number of occasions on which throt- tling had to be activated is increased. Oiiierwiae, Liie current power consumption P lies below the preset power limit PL, monitoring continues in step 438. Using a hardware implemented power controller 382 to imple^¬ ment the power monitoring and throttling operations has the advantage of fast reaction times. In particular, the power consumption of the blade server 342 may be controlled within a few microseconds, e.g. 50 ys . In this way, an overload situation can be avoided for the blade server system 300.

Meanwhile, the microcontroller 452 of the management blade 350 runs an adaptive power monitoring loop for the entire blade server system 300.

In a first step 450, the adaptive monitoring starts by ini^¬ tializing the monitoring loop. For example, the length of a monitoring interval T, e.g. 20ms, may be queried from the non-volatile memory 354.

Then, in a next step 452, the counter values provided for each blade server 342, 344, 346 and 348 are read. The counter values may be stored in a volatile memory of the remote man^¬ agement controllers 372 to 378, within the power controller 382 to 388 or may be stored in a memory or register of the management blade 350. Accordingly, the values may either be read locally or retrieved over the system management bus 322.

In a subsequent step 454, the read counter values are evalu- ated. For example, the management blade 350 may determine which one of the blade servers 342, 344, 346 and 348 exceeded their preset power limit PL most often. Based on the evaluation, in a subsequent auep ¾oo, Liie man^¬ agement blade 350 decides whether re-budgeting of the avail^¬ able power is necessary. For example, if a number of blade servers 342, 344, 346 and 348 exceeded their respective power limit PL within the given monitoring interval T by a certain threshold value, new power limits PL for some or all of the blade servers 342, 344, 346 and 348 may be determined in a step 434 and provided to the blade servers 342, 344, 346 and 348. If no re-budgeting of the available power is required, for example because none of the blade servers 342, 344, 346 and 348 exceeded its preset power limit PL, in a step 458, the management blade 350 waits for the next re-evaluation pe^¬ riod and then continues the method in step 450.

Figure 5 shows the development of the power distribution of a blade server system over time. In a first re-evaluation period 510, each one of eight blade servers BL#1 to BL#8 start with an initial, minimum power setting required to boot the respective blade server. Furthermore, a fixed amount of power is allocated to the remaining infrastructure units. In the example shown in Figure 5, there remains an additional, un^¬ used power headroom 520 between the maximum power available for the given enclosure and the combined infrastructure units ' power consumption and the initial power consumption of the blade servers BL# 1 to BL#8.

During operation of the blade server system, the power consumption of the individual blade servers is re-evaluated at regular intervals. In one re-evaluation period, the blade servers BL#2, BL#6 and BL#8 have exceeded their preset power limit. In response, the management blade has increased their power limit for the re-evaluation period 530 by allocating them more power from the available power headroom 520. In a further re-evaluation period 550, also the jjr aue se ve s

BL#1, BL#4 and BL#6 have their power allocation increased due to increased power requirements. Figure 6 shows the power distribution of the same blade serv^¬ er system after the total power consumption of the blade servers BL#1 to BL#8 and the general infrastructure units has reached the maximum power available for the enclosure of the blade server system.

In an evaluation period 610 the blade servers BL#1, Bl#3, BL#6 and BL#8 have a relatively large allocated power budget. In contrast, the blade servers BL#2 and BL#5 have a rela^¬ tively small power budget. The remaining blade servers BL#4, and BL#7 have an medium-sized power budget. The power budget 620, which exceeds the basic power consumption of the infra^¬ structure units may be redistributed in each subsequent evaluation period 630 and 650. For example, between the re- evaluation period 610 and the subsequent re-evaluation period 630, the power budget of the blade server BL#1 was reduced while the power budget of the blade server BL#2 was in^¬ creased. Similarly, between the re-evaluation periods 630 and 650, the power budget of the blade server BL#1 was further reduced, while the power budget of the blade server BL#3 was increased.

The subsequent flow charts of Figures 7 to 9 show the allo^¬ cation of power to the individual blade servers in more de^¬ tail .

In a first step 710, a current power budget of a blade server system is initially set to the power consumption of unmanaged infrastructure components including storage blades, connec- tion blades, fan modules and the like. Fui Liieimore, aome ini^¬ tial counter values used for the evaluation of the power budget are set to zero.

In a step 720 a loop starts by selecting a first detected blade server from all detected blade server in order to de^¬ termine the initial power budget for the first blade server.

In step 730, the current power budget is increased by the current power budget stored for the first blade server. For example, the value used in step 730 may be provided by a re^¬ mote management controller of the first blade server and may correspond to the initial power required for booting the first blade server or a previously set power limit PL. In a step 740, the method checks whether adaptive power budg^¬ eting is available for the current blade server. This may be determined by the presence of an appropriate power control^¬ ler, a configuration setting stored by a management blade, or a combination of both. If no adaptive power management is available, the initially provided power budget value is used for that particular blade server and the loop continues with the next detected blade server in step 720.

Otherwise, the method continuous for the first blade server in step 750. In step 750, an over count value indicating the number of times that the current power budget was exceeded by the first blade server in the last monitoring period is de^¬ termined. Furthermore, a counter is increased, which indi^¬ cates the number of blade servers included in the adaptive power budgeting.

In a further step 760, the management blade checks whether a active priority table is set for the blade server system. A present priority table may be used to manua±±y y ± ve a ii±y iier priority to blade servers running particularly important ap^¬ plications. If an active table is set in a further step 770, the table entry containing the over count values retrieved in step 750 is updated in the active priority table. Alterna^¬ tively, in a step 780, a corresponding table containing the number of over counts for the particular blade server as entry is created. This table is ordered by the number of over counts itself.

In a step 790, either the next available server blade is processed starting from step 720, or the loop ends if no fur^¬ ther blade servers are to be considered. Then, in a step 795, a decision is made whether the currently computed power budget exceeds the maximum power budget avail^¬ able for the enclosure of the blade server system or not. The maximum power budget available may either be the combined output of the active power supply units or correspond to a manually set power limit for the given blade enclosure. If the currently computed power budget does not exceed the total available maximum power, an unbounded reallocation of a power is performed as shown in Figure 5 and further detailed in the flow chart of Figure 8. Otherwise, if insufficient power is available to operate all blade servers according to their current needs, the bounded case as shown in Figure 6 and ex^¬ plained in more detail with respect to Figure 9 is performed.

In the unbounded case, in a first step 810, at first the available power is computed as the difference between the maximum available power of the enclosure and the current power budget. Then, in a loop, power is allocated to each blade server. For this purpose, a blade server having the highest priority according to the tables prepareu ±11 e±Liier step 770 or step 780 is selected in step 820.

In a step 830, an additional power budget for the presently selected blade server is computed as ten percent of its cur- rent power budget. In a step 840, the method checks whether the additional power budget lies below the available power computed in step 810. If this is the case, the power budget of the respective blade server is increased by the additional power budget. Furthermore, the current power budget of the entire enclosure is also increased by the additional power budget. The new power limit is then communicated to the re^¬ mote management controller of the respective blade server and the available power is reduced by the additional power budget. Otherwise, if the additional power budget exceeds the available power, in a step 860, the remaining available power is granted to the respective blade server in a similar way and the remaining variables are updated accordingly.

In a step 870, the loop continues by selecting the blade server having the next lower priority until all blade servers have been reallocated a new power budget or no further power is available for allocation. Then, the method ends the cur^¬ rent re-evaluation period in step 880. Figure 9 shows the corresponding flow chart for the case the current power budget has already reached the maximum power available for the enclosure of the blade server system.

In a step 910, the number of blade servers having an over count value of zero is determined. Furthermore, the available power for redistribution is set to an initial value of zero. _ _n

- ZD -

In a further step 920, the management blaue JOU c c a ±J_ uiie sum of all blade servers having an over count of zero is lar^¬ ger than zero. If this is not the case, i.e. if all blade servers have at least once exceeded their preset power limit, there is no power to be redistributed and the method ends im^¬ mediately in step 980.

However, if at least one blade server has not exceeded its previously set power limit, in a first phase 930, the amount of power available for redistribution is computed. In a step 932 a first blade server is selected in order to retrieve its associated over count value. In a step 934, it is checked whether the over count value of the respective blade server equals zero. If that is the case, in a step 936, ten percent of the previously allocated budget of that specific blade server is removed from the power budget of the selected server blade and is added to the budget of available power. If the counter of the associated blade server is larger than zero, the respective blade server is already using all of its associated power budget and the loop continues in step 938 with the selection of the next blade server.

In a second phase 940, the power budget available for reallo^¬ cation is distributed among the managed blade servers. For this purpose, in a step 942, a second loop starts in the or^¬ der of the priority of the respective blade servers stored in the previously determined priority tables.

In a step 944, the management blade 350 checks whether the current blade server has a over count value that exceeds zero. If not, the current blade server does not require any further power budget and the next blade server can be se^¬ lected in step 942. However, if the counter value of the cur- rent blade server exceeds zero, in a step J O, an auu±L±ona± power budget of ten percent of the current power budget of the respective blade server is computed.

In step 948, the management blade 350 checks whether the ad- ditional power budget exceeds the available power budget. If the additional power budget does not exceed the available power budget, in a step 950, a new power limit for the re^¬ spective blade server is computed based on the previously set power limit and the additional power budget. As described with respect to step 850, the remaining variables relating to the available power and the current power limit of the re^¬ spective blade server are updated. If the additional power budget exceeds the available power to be reallocated, the re^¬ maining power to be reallocated is allocated to the current blade server in a step 960. The allocation in step 960 corre^¬ sponds to the allocation described above with respect to step 860.

In a step 970, the reallocation of any remaining power is re- peated for the blade server having the next lower priority until all available power is distributed or no further blade servers need to be considered. Then, the method ends in a step 980. The method and systems described above with reference to Fig^¬ ures 1 to 9 describe various embodiments of particularly ad^¬ vantageous and currently preferred embodiments of the present invention. However, as it will be obvious to a person skilled in the art on data processing system, individual elements of different embodiments may be combined in order to carry out the invention. Furthermore, particular details given above W±LII re apec u LO the algorithms for determining and redistributing the available power may be changed in accordance with a policy setting provided by the manufacturer or operator of the blade server system. For example, rather than a proportional increase and decrease of power budgets, a fixed amount of additional power may be added or removed from any blade server.

In addition, instead of the described hardware throttling us^¬ ing a reduction in a operating frequency of a processor, other means of reducing the current power consumption of a component of a blade server system and may be employed. For example, a so-called soft throttling may be employed by up^¬ dating a P-state of a processer using the so-called advanced configuration and power interface (ACPI) of the blade servers in order to reduce their current power consumption. Furthermore, a fully software controlled throttling may be performed by means of an interface to an operating system running on an individual blade server.

Claims

1. A blade server system (300), comprising

an enclosure (310) having a plurality of first mounting positions to accommodate a plurality of power supply units (332, 334, 336) and a plurality of second mounting positions to accommodate a plurality of blade modules; at least one power supply unit (332, 334, 336) accommo^¬ dated in one of the plurality of first mounting posi^¬ tions ;

at least one management blade (350), the management blade (350) comprising at least one controller (352); at least one blade server (342, 344, 346, 348) accommo^¬ dated in one of the plurality of second mounting posi^¬ tions, the at least one blade server (342, 344, 346, 348) comprising at least one processor (362, 364, 366, 368), a memory accessible to the at least one processor (362, 364, 366, 368) and a remote management controller (372, 374, 376, 378);

wherein

the controller (352) of the management blade (350) is configured to query the at least one power supply unit (332, 334, 336) to determine a total available power for the blade server system (300), to determine a power limit for the at least one blade server (342, 344, 346, 348), and to provide the determined power limit to the remote management controller (372, 374, 376, 378) of the at least one blade server (342, 344, 346, 348);

the remote management controller (372, 374, 376, 378) of the at least one blade server (342, 344, 346, 348) is configured to store the power limit provided by the con^¬ troller (352) of the management blade (350) and to in^¬ crease a counter associated with the blade server if an associated power monitoring means inu±cauea LIICLL a cur^¬ rent power consumption of the at least one blade server exceeds the stored power consumption limit; and

the controller (352) of the management blade (350) is further configured to determine a new power consumption limit for the at least one blade server (342, 344, 346, 348) based on the total available power and the value of the counter associated with the at least one blade server (342, 344, 346, 348) after a predetermined monitoring interval .

The blade server system (300) of claim 1, wherein the at least one blade server (342, 344, 346, 348) com^¬ prises a power controller (382, 384, 386, 388), the associated power monitoring means is comprised in the power controller (382, 384, 386, 388) and the power controller (382, 384, 386, 388) is adapted to provide a throttling signal to the processor (362, 364, 366, 368).

A method for power management in a data processing sys^¬ tem (100) comprising a plurality of processing units (122, 124, 126) and at least one power supply unit (112, 114), the method comprising the steps:

determining a total available power provided by the at least one power supply unit (112, 114);

determining and storing an initial power limit for each one of the plurality of processing units (122, 124, 126) ;

repeatedly comparing the current power consumption of each one of the plurality processing units with the re^¬ spective stored power limits;

increasing a counter associated with one of the plural^¬ ity of processing units (122, 124, 126) if the current power consumption of said one of the p±ura±±uy o r proc^¬ essing units (122, 124, 126) exceeds the respective stored power limit; and

determining and storing a new power limit for at least one of the plurality of processing units (122, 124, 126) after a predetermined period of time based on the total available power and the associated counter value of said one of the plurality of processing units (122, 124, 126) .

The method of claim 3, further comprising

throttling one of the plurality of processing units (122, 124, 126) if a current power consumption of said one of the plurality of processing units (122, 124, 126) exceeds the respective stored power limit.

The method of claim 3 or 4, wherein, in the step of de^¬ termining and storing a new power limit, the new power limit for said one of the plurality of processing units is increased with respect to the previously stored power limit if the associated counter value exceeds the prede^¬ termined first threshold.

The method of claim 3 or 4, wherein, in the step of de^¬ termining and storing a new power limit, new power limits for at least a first and a second one of the plural^¬ ity of processing units (122, 124, 126) are determined and stored; and the new power limit for the first one of the processing units (122) is increased with respect to the associated previously stored power limit and the new power limit for the second one of the processing units (126) is decreased with respect to the associated previ^¬ ously stored power limit, if the counter value associ- ated with the first one of the proceaainy u ±m (-L^- .) exceeds the predetermined first threshold and the counter value associated with the second one of the processing units (126) lies below a second threshold.

The method of claim 3 or 4, wherein, in the step of de^¬ termining and storing a new power limit comprises a first and a second phase, wherein,

in the first phase (930), the amount of power available for distribution is determined based on the total avail^¬ able power; and,

in a second phase (940), the determined power available is distributed between a plurality of processing units (122, 124, 126) based on the counter values associated with the plurality of processing units (122, 124, 126) .

The method of claim 7, wherein, in the second phase (940), the following steps are performed repeatedly while the available power is greater than zero:

reducing the power available by an additional power budget; and

determine a new power limit for one processing unit (122, 124, 126) having an associated counter value exceeding a first threshold value based on the sum of the current power limit of that processing unit (122) and the additional power budget.

The method of claim 7 or 8, wherein, in the second phase (940), new power limits for the plurality of processing units (122, 124, 126) are determined in a sequence de^¬ termined by a priority level associated with the respec^¬ tive processing unit (122, 124, 126) .

The method of claim 9, wherein the priuiny ievei lux each processing unit is stored by a management unit (132) and may be adjusted by using a user interface of the management unit (132) .

The method of claim 9, wherein the priority level for each processing unit (122, 124, 126) is determined by a management unit (132) in order of the associated counter values of the plurality of processing units (122, 124, 126) .

The method of any one of claims 7 to 11, wherein in the first phase (930), the power available is determined based on the difference between the total available power and the sum of all power limits associated with the plurality of processing units (122, 124, 126) .

The method of any one of claims 7 to 12, wherein in the first phase (930), in the determination of the power available, a preset power consumption of other compo^¬ nents of the data processing system (100) is subtracted from the total available power.

The method of any one of claims 7 to 13, wherein, in the first phase (930), in the determination of the power available, for each processing unit (122, 124, 126) having an associated counter value below a second threshold value, a released power budget is added to the power available; and

a new power limit for said processing units (122,

124,126) having an associated counter value below a sec^¬ ond threshold value is based on the current power limit of the respective processing unit reuuceu jjy Liie re^¬ leased power budget.

An electronic data processing system (100), comprising: a plurality of processing units (122, 124, 126), each processing unit comprising at least one processor;

at least one power supply unit (112, 114) for providing at least one supply voltage; and

a management unit (132) comprising at least one control^¬ ler operatively coupled with the plurality of processing units (122, 124, 126) and the at least one power supply unit (112, 114);

wherein the management unit (132) is adapted to deter^¬ mine a new power consumption limit for at least one of the plurality of processing units (122, 124, 126) based on a total available power provided by the at least one power supply unit (112, 114) and the number of times a previously determined power consumption limit was ex^¬ ceeded by said one of the plurality of processing units (122, 124, 126) in a predetermined monitoring interval.

The electronic data processing system (100) of claim 15, further comprising:

a plurality of power monitoring means (142, 144, 146), each one of the power monitoring means configured to compare a current power consumption of an associated processing unit (122, 124, 126) of the plurality of processing units (122, 124, 126) with a current power consumption limit provided by the management unit (132) for the associated processing unit (122, 124, 126) .

The electronic data processing system (100) of claim 16, wherein the processing units (122, 124, 126) are config- ured to throttle their power consumption ±J_ Liie ciaauci- ated power monitoring means (142, 144, 146) indicate that the current power consumption of the processing unit (122, 124, 126) exceeds the current power consump- tion limit.

18. The electronic data processing system (100) of claim 17, wherein the processing units (122, 124, 126) are configured to throttle their power consumption by asserting a throttling control signal, and wherein the at least one processor of the respective processing units (122, 124, 126) is configured to reduce a clock frequency in re^¬ sponse to the assertion of the throttling control sig^¬ nal .

19. The electronic data processing system (100) of claim 17, wherein the processing units (122, 124, 126) are configured to throttle their power consumption by provision of an ACPI control parameter, and wherein the at least one processor of the respective processing units (122, 124,

126) is configured to reduce its workload in response to the provided ACPI controlled parameter.

20. The electronic data processing system (100) of claim 17, wherein the processing units (122, 124, 126) are configured to throttle their power consumption by initiation of a shutdown operation of the processing unit (122, 124, 126) .

21. The electronic data processing system (100) of any one of claims 15 to 20, wherein the management unit (132) is adapted to allocate, in a first step, power to other in^¬ frastructure components of the electronic data process- ing system and to allocate, in a seconu auep, a remain^¬ ing available power to the plurality of processing units (122, 124, 126) based on the number of times the previously determined power consumption limit was exceeded by a respective one of the plurality of processing units.