A microprocessor is an electronic device capable of performing the processing and control functions for computing devices such as desktop computers, laptop computers, server computers, cell phones, laser printers, and so on. Typically, a microprocessor comprises a small plastic or ceramic package that contains and protects a small piece of semiconductor material that includes a complex integrated circuit. Leads connected to the integrated circuit are attached to pins that protrude from the package allowing the integrated circuit to be connected to other electronic devices and circuits. Microprocessors are usually plugged into or otherwise attached to a circuit board containing other electronic devices.
While a microprocessor integrated circuit typically includes only one computing unit, i.e., one processor, it is possible to include multiple processors in a microprocessor integrated circuit. The multiple processors, which are often referred to as “cores,” are included in the same piece of semiconductor material and connected to the microprocessor package pins. Having multiple cores increases the computing power of the microprocessor. For example, a microprocessor with four cores can provide almost the same amount of computing power as four single-core microprocessors. Harnessing the increased computing power that multiple-core microprocessors provide allows computing functions that previously required multiple computing devices to be performed with fewer computing devices.
For example, a server implemented across 32 traditional computing devices, i.e., a 32-way server, may be implemented by eight microprocessors, each having four cores. Taking the concept one step further, if each individual core is eight times more powerful than one of the 32 computing devices, a 32-way server may be implemented by one microprocessor with four cores. Reducing the number of microprocessors reduces the cost of the server, the amount of energy required to power the server, and the amount of maintenance the server requires.
The advantages of using multiple-core microprocessors is driving a trend toward “server consolidation.” Server consolidation is the process of taking multiple servers, possibly each providing a different service, and providing all of the services on one physical device, e.g., a four-core processor. While reducing costs, energy, and maintenance, consolidating servers has the effect of putting all of one's eggs into one basket. This puts a greater burden of reliability on the one physical device. If a server is implemented on many separate computing devices and a computing device fails, usually there are other computing devices that are able to take over for the failed computing device. The process of having one computing device take over for a failing computing device is referred to as “failover.” Techniques have been developed for traditional server configurations to perform failover in a controlled and orderly fashion to ensure that no data is lost and no ongoing processes are interrupted during the transition from the failing computing device to the replacement computing device.
In order to create multiple-core microprocessor servers that are as robust and reliable as single-core microprocessor servers, similar techniques are required.
This summary is provided to introduce a selection of concepts in a simplified form that are further described below in the Detailed Description. This summary is not intended to identify key features of the claimed subject matter, nor is it intended to be used as an aid in determining the scope of the claimed subject matter.
Transparently replacing an interrupt controlled processor with a replacement processor is disclosed. Rather than directing interrupts directly to processors, interrupts are directed to an unchangeable identifier mapped to a processor's identifier. Before the interrupts are directed to the replacement processor, the replacement processor's identifier is mapped to the unchangeable identifier. The interrupts are directed to the unchangeable identifier. The mapping of an unchangeable identifier to processor's identifiers rather than directly to processors allows processors to be transparently replaced.
A processor is replaced with a replacement processor by temporarily restricting the interrupts that are directed to the processor to be replaced; activating the replacement processor; mapping the replacement processor's identifier to the unchangeable identifier; isolating the processor to be replaced; and using the mapping of the unchangeable identifier to the replacement processor's identifier to direct subsequent interrupts to the replacement processor. Preferably, an intermediary, such as an I/O APIC, stores the unchangeable identifier.
The mapping of the unchangeable identifier to the replacement processor's identifier may be used for logical mode delivery, physical mode delivery, or interrupt mapping.
The foregoing aspects and many of the attendant advantages of this invention will become more readily appreciated as the same become better understood by reference to the following detailed description, when taken in conjunction with the accompanying drawings, wherein:
Functionally, a server is an entity on a network that provides a service. For example, a web page server provides a service that returns web pages in response to web page requests. Other exemplary servers are an email server that returns email messages for particular users, a video server that returns video clips from a video archive, etc. Physically, a server is a stand-alone, self-contained computing device that is often connected to other servers through networking or a similar protocol. Traditionally, there is a one-to-one mapping of a functional server to a physical server. For example, an email server is implemented on one traditional physical server. If a traditional physical email server fails, the email service can be restored by replacing the failed physical email server with another physical server.
An exemplary physical server contains a microprocessor, a memory controller, and memory blocks controlled by the memory controller. The memory controller and the memory blocks controlled by the memory controller are often referred to as a unit, i.e., a memory unit. Physical servers may also contain additional microprocessors, memory controllers, memory blocks, and other electronic devices such as interrupt processors. Hence, physical servers containing only a microprocessor and/or memory unit should be construed as exemplary and not limiting. As with many types of computing devices, the operation of a physical server is controlled by a software program called an operating system. A physical server executes the instructions contained in a copy of the operating system, i.e., an instance of the operating system.
Multiple-core microprocessors make it possible to implement more than one functional server on a physical server by partitioning the resources available on the physical server into individually manageable “partitions” comprising “partition units.” A partition unit comprises an electrically isolatable microprocessor, a memory unit, and/or perhaps other electronic devices, e.g., an interrupt processor. A partition comprises one or more partition units. Hence, a partition is an electrically isolatable set of partition units and electronic devices within a physical server that can run an independent instance of an operating system, i.e., a local operating system, to implement a functional server. Hereinafter, except where noted, the term “server” refers to a physical server.
Preferably, partitioning is dynamic. That is, partitioning is performed on active computing devices, i.e., computing devices that are energized and performing useful functions. Also preferably, partitioning is transparent. That is, partition units are assigned to, or removed from, partitions with little or no impact on the services the server provides. To support dynamic, transparent partitioning, partition units are managed as whole units and not subdivided. For example, a partition unit is moved into a partition as a unit. Therefore, when a partition unit is replaced, all of the devices in the partition unit are replaced. A server that is capable of being partitioned is a partitionable server. A server system, i.e., system, comprising partitionable servers is a partitionable system. A partitionable system provides flexibility in the number and configuration of partition units and electronic devices assigned to a partition. Partitionable systems support “server consolidation.”
Server consolidation is the process of taking multiple traditional servers, possibly each providing a different service, and providing all of the services on one partitionable server. While reducing cost, energy, and maintenance, consolidating service puts a greater burden of reliability on the partitionable server. Whereas a traditional server implemented on many separate computing devices usually has spare computing devices that are able to take over for failing computing devices, a partitionable server needs to look elsewhere for “backup” computing power. The process of having one computing device take over for a failing computing device is referred to as “failover.” Techniques have been developed for traditional server configurations to perform failover in a controlled and orderly fashion to ensure that no data is lost and no ongoing processes are interrupted during the transition from the failing computing device to the replacement computing device. In traditional server configurations the failing “computing device” was itself a server. Since servers connect to each other through a network and are not tightly tied together, work needed to be broken into small pieces and shared across the servers, i.e., packetized. This made it easy to replace a failing server since the failing server's work packets could be re-routed. With server consolidation the overhead of the packetizing of the work is gone, but so is the ease of completely removing a server. In order to implement servers on partitionable servers that are as robust and reliable as traditional servers, similar techniques are required.
It is impractical to make partitionable servers more reliable by notifying each of the high-level software applications when a failover is required. To enable high-level software applications to respond to such a notification would require that the computer code for each application be modified to adapt to the failover. Even notifying applications would probably not be enough to provide failover without a mechanism to replace a portion of a running server, which is not usually required in traditional server configurations. Instead, it is more practical and advantageous to involve only the lowest level software in the failover and allow the upper level software, e.g., applications, to behave as though no hardware change has happened.
An implementation of an orderly, low-level, partitionable server failover involves a global management entity and one or more local operating systems. Examples of a global management entity are a service processor (SP) and a baseboard management controller (BMC). An SP is a specialized microprocessor or microcontroller that manages electronic devices attached to a circuit board or motherboard, such as memory controllers and microprocessors. A BMC is also a specialized microcontroller embedded on a motherboard. In addition to managing electronic devices, a BMC monitors the input from sensors built into a computing system to report on and/or respond to parameters such as temperature, cooling fan speeds, power mode, operating system status, etc. Other electronic devices may fulfill the role of a global management entity. Hence, the use of an SP or BMC as a global management entity should be construed as exemplary and not limiting.
A local operating system is an instance of an operating system that runs on one partition. Partition units, which contain logical devices that represent one or more physical devices, are assigned to a specific partition to ensure that the logical devices cannot be shared with logical devices in other partitions, ensuring that a failure will be isolated to a single partition. Such a partition unit may indicate which physical addresses are serviced by a given memory controller and, thereby, map the physical memory addresses to the memory controller and to the physical partition unit containing the memory controller. More than one partition unit may be required to boot and operate a partition. Unused or failing partition units may be electrically isolated. Electrically isolating partition units is similar to removing a server from a group of traditional servers with the advantage that partition units may be dynamically reassigned to different partitions. Managing, e.g., adding or replacing, the partition units in a partitionable server allows a failover to be performed in a controlled and orderly fashion to ensure that the partitionable server is as robust and reliable as traditional servers.
An exemplary computing device 100 for implementing a partitionable server capable of supporting partitions and partition unit addition and/or replacement is illustrated in block diagram form in
In
The replacement of partition units may be understood by comparing the block diagram shown in
The partition 200a illustrated in
Partition 200a and partition 200b are in effect the same partition in that they have the same partition identifier (ID), the difference being that partition 200a is made up of a different set of partition units than is partition 200b. Prior to the transfer, the partition IDs of processors A, B and C were the ID of partition 200a/200b. The partition ID of processor D was different or not set, i.e., zeroed depending on the prior status of the partition unit including processor D. Regardless of the partition ID of processor D, as explained more fully below, after the transfer the partition ID of processor D becomes the ID of partition 200a/200b.
Replacing a partition unit involves identifying the hardware devices that require replacement and the replacement hardware devices. It is common for a processor, such as processor A 202, to have an Advanced Programmable Interrupt Controller ID (APIC ID) identifying the processor. Similarly, within a partition's local operating system, a memory unit's physical address uniquely identifies the memory unit. Within a partition's local operating system, such as partition 200a's local operating system, a processor's APIC ID is uniquely identifies the processor. A computing device, such as computing device 100, shown in
During a partition unit replacement, such as the partition unit replacement shown in
Partition unit IDs are a combination of the partition ID and a hardware device identifier such as an APIC ID for a processor or a physical address for a memory unit. For example, to create a unique global identifier for processor C 210, processor C 210's APIC ID is combined with partition 200a's partition ID. Similarly, to create a unique global identifier for memory 212, memory 212's physical address is combined with partition 200a's partition ID.
When a partition unit is replaced, each of the hardware devices in the partition unit is replaced. For example, as shown in
While a single processor and a single memory block, such as processor A 202 and memory 204, may comprise a partition unit, a partition unit may have other forms. A detailed view of an exemplary partition unit 400 having a different form is illustrated in
A logical device in a typical partition unit may be capable of notifying the local operating system of the device's status. Alternatively, or in addition, the local operating system controlling the partition unit may use predictive analysis to assess the status of the logical device and determine if the logical device might be failing and thus, may be a candidate for replacement. While a person, such as a system administrator, might check device status as a part of regular maintenance, it is preferable to have the hardware itself notify the local operating system of an impending failure. In some situations, it may be desirable to upgrade a processor from one model to another model or to add processors and/or memory to a system. While a system administrator may perform such functions, it is preferable to automate such replacements and additions by using explicitly programmed instructions or by periodically timed instructions that make use of partitions, partition units, and the ability of hardware to report status.
Processes, such as the processes in a local operating system, that are running on a processor that is to be replaced must be quiesced, i.e., put into an inactive state, because if the processor is in use, the processor's state is constantly changing. If the processor's state is changing, the processor cannot be safely and reliably replaced because the processor's state cannot be safely and reliably transferred. Therefore, a pause operation is executed by the processor that is to be replaced, e.g., a failing processor, to prevent the processor's state from changing. Those skilled in the art and others will appreciate that the process of pausing a processor to prevent the processor's state from changing is referred to as “quiescing” the processor. A system, such as the computing device 100 shown in
A partition unit, such as the partition unit 450 illustrated in
When an operating system starts, each logical processor is assigned a unique, initial APIC ID. The initial APIC ID is composed of the physical processor's ID and the logical processor's ID within the physical processor. An operating system may use initial APIC IDs to direct interrupts to particular processors. A device, e.g., a disc drive, may transmit an interrupt signal directly to the processor using a message signaled interrupt (MSI). A device interrupt may instead be routed into an intermediary software entity, i.e., an intermediary. The device generates a signal that is transmitted to the intermediary and the intermediary forwards the signal to the processor. An exemplary traditional intermediary is an Input/Output Advanced Programmable Interrupt Controller (I/O APIC). The block diagram in
The replacement of a partition unit, such as the partition unit 450 illustrated in
There are ways to allow the physical APIC ID of a processor to be unique yet still transfer the state of the processor, e.g., a failing processor, to another processor, e.g., a replacement processor. To other entities, the two processors will appear to be identical. Thus, for example, an interrupt directed to the failing processor will instead be directed to the replacement processor.
As described above, on many processors, the APIC ID, i.e., the physical and/or logical APIC ID, may be hardwired inside of the processor preventing the identity of the processor from being transferred. Rather than relying directly on the APIC ID to identify a processor, a processor's APIC ID, i.e., physical and/or logical APIC ID, may be hidden inside of a service, making it possible to present the other parts of the system with a service for doing operations that involve the APIC ID without direct reference to the APIC ID. In discussing such a service, it is helpful to divide a system, e.g., computing device 100 into two portions—a service processor (SP) portion and a non-SP portion. The SP portion is the combination of the SP 102, the SP firmware 104, and the routing table 106. The non-SP portion comprises the remaining items in the computing device 100. To transparently replace a processor, the non-SP portion is isolated from the identity of the processor allowing the processor identifier to be remapped. Remapping processor identifiers is accomplished by a set of instructions that may be stored in the SP portion, e.g., the SP firmware 104, or stored in other memory and pointed to by the SP firmware 104. The set of instructions for remapping processor identifiers is referred to hereafter as the “processor remapping service”.
Rather than relying directly on a processor's APIC ID to identify the processor, the processor's APIC ID is hidden inside of the processor remapping service, making it possible to present the non-SP portion with a service for doing operations that involve APIC IDs without direct reference to the APIC IDs. The processor remapping service assigns and accepts “unchangeable” APIC IDs, i.e., APIC IDs that do not change and are used by the non-SP portion. The processor remapping service converts the unchangeable APIC ID to the appropriate changeable APIC ID. Hence, during a processor replacement, whether the changeable APIC ID is physical or logical, the changeable APIC ID can be changed to refer to a replacement processor; thus, making the processor replacement transparent.
An exemplary processor remapping service provides three processes for transparently replacing processors: logical mode delivery, physical mode delivery, and interrupt remapping. By using one of the three processes during a processor replacement, the details of which processor is being used are abstracted out and hidden from the non-SP portion.
In the logical mode delivery process, the system, e.g., computing device 100, is configured such that the non-SP portion uses only logical APIC IDs and are not permitted to access or use physical APIC IDs. In the logical mode delivery process, the processor remapping service uses the logical APIC ID as a remapping register. A logical APIC ID of a processor can be programmed, i.e., changed, usually by low level software. The logical APIC ID is changeable and can be easily transferred to another processor without relying on I/O APICs and MSIs. By inserting the failing processor's logical APIC ID into the replacement processor, the failing processor's logical APIC ID is mapped to the replacement processor.
If logical mode delivery is not available in a computing device and/or system, physical mode delivery may be used to transparently replace processors. Physical mode delivery involves an intermediary. An exemplary intermediary is a redirection table in an I/O APIC. (See
In the physical mode delivery process an I/O APIC provides an interface through which an SP interacts with the I/O APIC. The physical mode delivery process takes advantage of the already existing I/O APIC interface to program a new physical ID, i.e., the physical ID of the replacement processor, into the I/O APIC's redirection table.
In the interrupt remapping process, the processor remapping service redirects interrupts from a processor that is replaced, e.g., processor A 202, to a replacement processor, e.g., processor D 214. Traditionally, devices that generate interrupts, such as a disc drive controller in the mass storage circuitry 126 of computing device 100, store the APIC ID of a processor that is intended to receive the interrupts, i.e., the destination processor. If the disc drive controller is “hardwired,” i.e., connected physically by a wire, circuit board trace, etc., to an I/O APIC, the disc drive controller sets the voltage on the wire to a level that triggers the programmed instructions in the I/O APIC's redirection table. Usually there are other devices that are connected into other entries in the I/O APIC's redirection table. Hence, there needs to be an interface at the connections of the I/O APIC to manage the arriving interrupts.
In the logical mode delivery, physical mode delivery, and interrupt remapping processes, the local operating system is quiesced and the state is transferred. It is particularly important the external devices are not aware of the physical APIC ID. Using logical delivery is the easiest way to accomplish that. If a device is “hardwired,” i.e., connected physically by a wire, circuit board trace, etc., to an I/O APIC, the device sets the voltage on the wire to a level that triggers the programming in the I/O APIC. It is likely that there are other devices that hook into other entries in the redirection table in the I/O APIC. The I/O APIC is, in effect, a shared state repository. Hence, there needs to be some interface at the connections of the I/O APIC to manage the shared state repository, i.e., the I/O APIC.
An exemplary process for replacing a processor, e.g., a failing processor, is illustrated in
Continuing in
While the preferred embodiment of the invention has been illustrated and described, it will be appreciated that various changes can be made therein without departing from the spirit and scope of the invention.
This application claims the benefit of U.S. Provisional Patent Applications “Transparent Replacement of a System CPU,” No. 60/866,821, filed Nov. 21, 2006; “Driver Model for Replacing Core System Hardware,” No. 60/866,817, filed Nov. 21, 2006; and “Replacing System Hardware,” No. 60/866,815, filed Nov. 21, 2006; and U.S. Nonprovisional Patent Applications “Replacing System Hardware,” No. ______ , filed concurrently herewith (Attorney Docket No. MSFT-1-28517); “Driver Model for Replacing Core System Hardware,” No. ______ , filed concurrently herewith (Attorney Docket No. MSFT-1-28519); and “Correlating Hardware Devices Between Local Operating System and Global Management Entity,” No. ______ , filed concurrently herewith (Attorney Docket No. MSFT-1-28307), the subject matter of which is also incorporated herein by reference.
Number | Date | Country | |
---|---|---|---|
60866821 | Nov 2006 | US | |
60866817 | Nov 2006 | US | |
60866815 | Nov 2006 | US |