Optimized function execution for a multiprocessor computer system

[0002] This application is based on U.S. Provisional Application Ser. No. 60/057,223, filed Aug. 29, 1997.

[0003] This invention relates generally to computer operating systems. More particularly, this invention relates to the scheduling by an operating system of functions for execution on specified processors or specified nodes of a multiprocessor computer system.

[0004] Multiprocessor computers by definition contain multiple processors (also referred to as CPUs) that can execute multiple parts of a computer program or multiple distinct programs simultaneously, in a manner known as parallel computing. In general, multiprocessor computers execute multithreaded-programs or single-threaded programs faster than conventional single processor computers, such as personal computers (PCs), that must execute programs sequentially. The actual performance advantage is a function of a number of factors, including the degree to which parts of a multithreaded-program or multiple distinct programs can be executed in parallel and the architecture of the particular multiprocessor computer at hand.

[0005] Multiprocessor computers may be classified by how they share information among the processors. Shared-memory multiprocessor computers offer a common physical memory address space that all processors can access. Multiple processes or multiple threads within the same process can communicate through shared variables in memory that allow them to read or write to the same memory location in the computer. Message passing multiprocessor computers, in contrast, have a separate memory space for each processor, requiring processes in such a system to communicate through explicit messages to each other.

[0006] Shared-memory multiprocessor computers may further be classified by how the memory is physically organized. In distributed shared-memory computers, the memory is divided into modules physically placed near each processor. Although all of the memory modules are globally accessible, a processor can access memory placed nearby faster than memory placed remotely. Because the memory access time differs based on memory location, distributed shared memory systems are often called non-uniform memory access (NUMA) machines. By contrast, in centralized shared-memory computers, the memory is physically in one location. Centralized shared-memory computers are called uniform memory access (UMA) machines because the memory is equidistant in time from each of the processors. Both forms of memory organization typically use high-speed cache memory in conjunction with main memory to reduce execution time.

[0007] Multiprocessor computers with distributed shared memory are often organized into multiple nodes with one or more processors per node. The nodes interface with each other through a memory-interconnect network by using a protocol, such as the protocol described in the Scalable Coherent Interface (SCI)(IEEE 1596). UMA machines, in contrast, typically use a bus for interconnecting all of the processors.

[0008] Further information on multiprocessor computer systems in general and NUMA machines in particular can be found in a number of works including Computer Architecture: A Quantitative Approach(2^ndEd. 1996), by D. Patterson and J. Hennessy, which is hereby incorporated by reference.

[0009] A single operating system typically controls the operation of a multinode multiprocessor computer with distributed shared memory. Examples of suitable operating systems include UNIX-based operating systems such as DYNIX/ptx, BSD, SVR4, UnixWare, PC UNIX, etc. For general information on such operating systems, see Bach, M. J., The Design of the UNIX Operating System,Prentice-Hall, 1986; Vahalia, U., Unix Internals: The New Frontier,Prentice-Hall, 1996; McKusick, M., et al., The Design and Implementation of the4.4 BSD Operating System,Addison-Wesley, 1996, which are all hereby incorporated by reference.

[0010] One purpose of such an operating system is to schedule the execution of functions that perform periodic tasks. These periodic tasks include retransmitting dropped network packets, scheduling and managing memory, monitoring devices to avoid losing interrupts, polling devices that do not support interrupts, and remote but immediate execution of a function on a specified CPU or node. This scheduling is often done through a timeout mechanism such as the UNIX timeout system call that defers a specified function's execution for a specified time or a remote execution mechanism that specifies on which CPU or node a function will be immediately executed.

[0011] Prior operating systems for UMA multiprocessor computers have used a number of timeout mechanisms that defer a function's execution a specified number of system clock ticks. A relevant mechanism to this invention includes a global priority queue data structure in the centralized memory, such as shown in FIG. 1. This data structure groups timeout handlers (timeout cells (tocells) A0, D1, F0, etc., that represent specified functions that are scheduled for execution) into “buckets.” All tocells in a given “future events” bucket are connected in a circular doubly-linked list and are scheduled to execute, or “fire,” on the same system clock tick, modulo the number of buckets. The array of buckets is indexed by the lower order bits of a system variable tonow that is incremented on each clock tick. All the tocells in the bucket after the one indexed by tonow are scheduled to fire on the next clock tick, modulo the number of buckets. On that next clock tick, these tocells are added to the end of the group of tocells in the “current events” bucket, where they are processed by any processor (CPU) that is currently available. The tocell is identified by letter (which represents a pointer to the memory containing the cell), and the associated number indicates the number of clock ticks still remaining after the tocell is placed in the current events bucket. If the number is zero, then processing the tocell will cause it to fire. If the number is nonzero, such as two or three, the CPU processing the tocell will place it in the proper future events bucket two or three ahead of the one pointed to by the index, with the number reset to zero. Thus in FIG. 1 when tocell C2 is placed on the current events list, it will not fire but will be moved two future events buckets ahead, to the bucket that contains D1 and E0, and its count set to zero.

[0012] Access to the global priority queue is controlled by a global “gate” (also known as a “lock”) a system variable that is set by a CPU to indicate that the queue is currently being accessed. This lock ensures that only one CPU at a time can access the future events buckets through the timeout system call to add or remove tocells, thereby preventing data corruption that would otherwise occur if multiple CPUs attempted to schedule functions simultaneously.

[0013] Although this timeout mechanism has served adequately for UMA machines, it has a number of drawbacks. The use of a global lock results in poor system behavior during overload, i.e., when multiple CPUs are simultaneously attempting to schedule functions. The global lock also severely limits the scalability of the timeout mechanism. That is, the performance of the timeout mechanism degrades as the number of CPUs in the computer increases. And in certain circumstances, a sequence of interrupts can cause multiple instances of a scheduled specified function for a tocell to be executing concurrently. This can be avoided only by using a much more complicated locking design for the lock.

[0014] An objective of the invention, therefore, is to provide a method and means for scheduling functions for execution on a multiprocessor computer without the delay inherent in prior approaches. Another objective of the invention is to provide such a method and means that permits functions to be scheduled for execution on a specified CPU (which could be the scheduling CPU or another CPU) or on a specified node of a multiprocessor computer. Yet another objective of the invention is to provide such a method and means that balances CPU loads by directing the execution of functions from overloaded CPUs to available CPUs of a multiprocessor computer.

[0015] In accordance with the invention, a shared memory multiprocessor computer includes, for each of the processors, an associated data structure stored in shared memory for scheduling the execution of specified functions by the processor. A scheduling function responsive to a request by a processor adds a specified function to a processor's data structure. The use of separate data structures for each processor avoids the delay inherent in the prior approach of using a single, global data structure for scheduling all functions for execution.

[0016] In one embodiment of the invention, the scheduling function is responsive to a request by a processor to add the specified function to the data structure of a processor other than the requesting processor. In another embodiment of the invention, the multiprocessor computer includes a multiple of nodes each of which has multiple processors. In this embodiment the scheduling function is responsive to a request by a processor to add the specified function to the data structure of a processor on a specified node.

[0017] The processor's data structure may take many forms. In a preferred embodiment of the invention, the data structure has first and second data structures. The first data structure includes a field containing a list of second data structures identifying specified functions to be immediately executed by the processor. The second data structure exists for each scheduled specified and includes a first field identifying the first data structure for association and a second field identifying a specified function for execution.

[0018] In methods associated with the invention, specified functions can be moved from one CPU to another if a CPU becomes unavailable or overloaded.

[0019] The foregoing and other objectives, features, and advantages of the invention will become more apparent from the following detailed description of a preferred embodiment which proceeds with reference to the accompanying drawings.

[0020] FIG. 1 is a block diagram of a data structure used in a prior art timeout mechanism.

[0021] FIG. 2 is a block diagram of a multinode multiprocessor computer system with a system interconnect coupling the nodes together.

[0022] FIG. 3 is a block diagram of one of the nodes of the computer system of FIG. 2.

[0023] FIG. 4 is a block diagram of how an operating system for the system of FIG. 2 operates with multiple processes and input and output devices.

[0024] FIG. 5 is a block diagram of an overall data structure used in a timeout mechanism in accordance with the invention for scheduling functions on a multiprocessor computer.

[0025] FIG. 6 is a more detailed diagram of the data structure of FIG. 5, showing the fields of the constituent data structures and a tocell in a future events bucket.

[0026] FIG. 7 is the same diagram of the data structure as FIG. 6, with the tocell now in the current events bucket.

[0027] FIG. 8 is a block diagram of an overall data structure used in a remote execution mechanism in accordance with the invention.

[0028] FIG. 9 is a block diagram of a process for measuring interrupt loads and for throttling interrupts

[0029] FIG. 10 is chart of the step response of a filtered interrupt load.

[0030] FIG. 11 is a chart of the response of an adjustment equation to the range of possible filtered loads.

[0031] System Overview

[0032] FIG. 2 is a block diagram of a multinode, multiprocessor computer system 10 in accordance with the invention. System 10 uses a computer architecture based on Distributed-Shared Memory (DSM) and is a non-uniform memory access (NUMA) machine. Four nodes 12, 14, 16, and 18 are shown connected by a system interconnect 20 (i.e., a network) that permits any node to communicate with any other node. Specifically, the purpose of system interconnect 20 is to allow processors in any node to access the memory resident in any other node. System interconnect 20 is a switch-based network that uses the Scalable Coherent Interface (SCI) interconnection mechanism. SCI is an IEEE-approved standard, or protocol (1596), well documented in a number of publications including IEEE Std 1596-1992 (Aug. 2, 1993) and Multiprocessor Interconnection Using SCI, a Master Thesis by Ivan Tving, DTH ID-E 579 (1994), both of which are hereby incorporated by reference.

[0033] The physical links of interconnect 20 provide high bandwidth and low latency and are scalable to allow for the addition of more nodes. Links that meet these requirements include point-to-point interconnects with a data throughput of one gigabyte/second or greater. The links can be configured in any number of suitable ways for connecting nodes 12, 14, 16, and 18 such as in a ring topology, in arbitrary topologies through switches, or in a combination of both. The links can be wired or wireless (optical, RF, etc.) depending upon system performance needs. Additional topologies are described in “Interconnect Topologies with Point-To-Point Rings,” Ross E. Johnson and James E. Goodman, December 1991, Computer Sciences Technical Report #1058, University of Wisconsin-Madison, which is hereby incorporated by reference.

[0034] Node Overview

[0035] A block diagram of node 12 on system 10 is shown in FIG. 3. The node includes a conventional symmetrical multiprocessor (SMP) node bus 22 for connecting multiple data processors 24 to local memory 26. For clarity, nodes 12, 14, 16, and 18 may be referred to as home nodes or remote nodes in certain circumstances. A home node is one whose local memory stores a memory block of interest (i.e., the address of the memory block falls within the address range supported by the local memory or cache); all of the other nodes are then remote nodes with respect to that memory block. Additionally, a node may be a requesting node or a responding node. A requesting node is one requesting data; a responding node is one furnishing such data. Input/output (I/O) 28, which is also connected to bus 22, connects the node to devices outside computer system 10 for communicating information between the computer system and the outside world. I/O 28 may be of conventional design and includes means for connecting the node (and hence system 10) to personal computers, local area networks, etc., that wish to utilize the power of the multinode computer system. The I/O 28 may also allow for connection to peripheral devices, such as floppy disks, hard disks, CD-ROMs etc.

[0036] To connect node 12 to the other nodes in the system, the node includes a system interconnect interface 30. The system interconnect interface forms part of interconnect 20 along with the physical links between nodes and the same devices on the other nodes of the computer system 10. In the present embodiment, interface 30 is constructed to implement the SCI standard for data communication between the nodes, allowing a processor on one node to directly access data stored on another node. The interface 30 also contains a remote cache in the present embodiment, although this combination is not required for the invention. The remote cache could also be separate from the system interconnect interface.

[0037] Operating System Overview

[0038] FIG. 4 shows an operating system 32 for computer system 10, which includes a kernel 34, a system call interface 36 and an I/O interface 38. The same operating system environment exists on the other nodes. The illustrated operating system is a UNIX-based operating system, although other operating systems may also be used. Kernel 34 (which is the heart of operating system 32) is a program stored in memory 26 on one or more nodes. Memory 26, processors 24, and other hardware shown in FIG. 3 are used for running the kernel 34 and are shown generally as computer hardware 40. Kernel 34 is responsible for controlling the computer system's resources and scheduling user requests so that each user request receives its fair share of the system resources. A system resource can be, for example, an I/O device (e.g., disk drive, tape drive, CD ROM, etc.), a shared memory segment, a file, a process, a CPU, etc.

[0039] Requests are generated by one or more processes, such as user process 42 and system process 44 that run on computer system 10. User process 42 includes a part of a program 48 (i.e., instructions) and an execution environment for running the program. For example, process 42 includes several components, such as a stack 50, private data 52 for storing local variables, and page tables 54 used to define physical memory associated with the process. The page table for a process may consist of a hierarchy of such tables, and both single and multi-level page tables are represented by page table 54. Process 44 includes a similar environment to that of process 42. Although two processes are shown, any number of processes may be run at one time on the node. The processes make requests to kernel 34 through system calls which are passed to the kernel by system call interface 36. Processes execute in parallel to the extent they can be run concurrently on the different processors of system 10.

[0040] Kernel 34 accesses I/O devices 56 through I/O interface 38. For example, if process 42 requests information from a disk drive (not shown) included within I/O devices 56, process 42 makes a system call to operating system 32. Operating system 32 uses I/O interface 38 to retrieve the information from the disk drive and returns the information to user process 42.

[0041] In a preferred embodiment of the invention, operating system 32 provides several functions (also known as primitives) that are used in system processes for scheduling the execution of other functions. These include primitives for deferred function execution (in the form of a timeout mechanism) and primitives for immediate function execution (in the form of a remote execution mechanism). These primitives work with unique data structures to schedule and reschedule functions efficiently in the manner described below.

[0042] The Timeout Mechanism

[0043] FIG. 5 is a block diagram of an overall data structure 60 used in a timeout mechanism in accordance with the invention for scheduling functions on a multiprocessor computer such as system 10. Although superficially similar in appearance to the data structure of FIG. 1, data structure 60 differs significantly in the makeup of its constituent data structures 64, 66, and 68 and their interaction in system 10. Unlike the global data structure of FIG. 1, there is no global data structure or global lock. Instead, data structure 60 is replicated by operating system 32 on a per-CPU basis, so that there is at least one such associated data structure and lock 62 for each CPU in computer system 10. And unlike the data structure of FIG. 1, data structures 60 are constructed to accept tocells (which include functions to be scheduled for execution) that identify individual CPUs and nodes in computer system 10, thereby permitting interaction between a data structure 60 and an unassociated CPU. Data structure 60 may be replicated by operating system 32 on a per-node basis in an alternative embodiment.

[0044] Each CPU in system 10 operates its data structure 60 largely independent of the other CPUs. If a CPU has no reason to concern itself with another CPU's tocells, then there is no interaction with another CPU's data structure. Contention for lock 62 thus rarely arises since only the CPU associated with the data structure typically accesses it, and the delay inherent with the prior timeout mechanism is avoided.

[0045] However, interaction is permitted (and encouraged) to occur in the following circumstances. One CPU may need to modify the scheduling of a function that has been scheduled by another CPU for execution. One CPU may need to schedule a function for execution by another CPU by placing a tocell 64 in the other CPU's data structure 60. Or a load imbalance may develop among the CPUs: one CPU is overloaded with functions to execute while another CPU is idle. Rectifying the load imbalance requires moving tocells from one CPU's data structure 60 to another CPU's data structure 60. This move is accomplished by changing the tocell's indication (such as field entry) of which CPU it is queued on. When the CPU indication is changed, the tocell's indication of its lock 62 must also change. To prevent data corruption, locks for both the source CPU (where the tocell was queued for execution) and the destination CPU (where the tocell is now queued for execution) must be held when updating the indication. A second way to prevent data corruption is to treat the updating of the indication as an implicit dropping of the lock.

[0046] Constituent Data Structures for the Timeout Mechanism

[0047] To enable each of the multiple CPUs in system 10 to access the data structures 60 of other CPUs, a number of novel constituent data structures within data structure 60 are provided in accordance with the invention. These are shown in FIG. 6 and include one or more timeout cells 64, a timeout_eng data structure 66, and a toc_qhead data structure 68.

[0048] The Timeout Cell (tocell)

[0049] Table 1 is a list of the significant fields in a preferred embodiment of a tocell data structure 64:

[0050] The to_stat field in Table 1, which gives the current state of the tocell, can assume one of a number of values:

[0051] The to_flags field in Table 1 can also assume a number of values:

[0052] The timeout_eng Data Structure

[0053] Data structure 60, being implemented on a per-CPU basis, is more complex than the prior system-wide structure of FIG. 1. Certain information about the state of the data structure must be tracked to efficiently use these data structures as a group, such as the number of tocells per data structure. The timeout_eng data structure 66 within data structure 60 tracks these states. Its fields are defined in Table 4 below:

[0054] The toc_qhead_t Data Structure

[0055] The toc_qhead_t structures 68 are t_now and t_buckets[ ] in Table 4. Each has the following fields (note the “tq_” prefixes on its field names):

[0056] FIG. 6 is a more detailed view of data structure 60, showing the pertinent fields described above for data structures 64, 66, and 68.

[0057] Selected Operations on the Timeout Data Structures

[0058] Step-by-step procedures for interaction between the CPUs of system 10 and their data structures 60 are described below.

[0059] Acquiring the Lock

[0060] In the global priority queue data structure of FIG. 1, there is a single, global lock. Acquiring that single lock gives a CPU free rein to modify any of the timeout-related data structures. However, the penalty for this simplicity is the noted severe performance limitations because only one CPU can perform timeout-related operations at one time. Furthermore, a single cache line containing the lock becomes a memory “hot spot”, and shuttling this single cache line back and forth between all of the CPUs of computer system 10 imposes substantial overhead. This overhead can result in “congestive collapse” under extreme timeout overload—all of the CPUs end up spending most of their time waiting to acquire the lock, and very little actual work is accomplished. The overhead becomes so great that in some instances the multiple CPUs of system 10 may accomplish less work than a single-CPU system.

[0061] This invention provides a solution to this problem by replicating data structure 60 on a per-CPU basis, as described above, and by including a separate lock 62 for each CPU as shown in FIGS. 5 and 6. Thus in the common case where a timeout is scheduled via a to_timeout( ) primitive and then fires normally, all processing for that timeout occurs on one CPU, thereby enjoying excellent cache locality and resulting high performance. Extreme overload situations that would lead to very poor performance with the prior timeout mechanism run very quickly on the new mechanism, consuming very little CPU time and memory bandwidth.

[0062] However, the previously simple operation of acquiring the lock is now more complex because there are multiple data structures 60 to consider. When operating on a tocell, a CPU must now look at the cell's to_eng pointer to determine the cell's current locking allegiance. Instead of acquiring a global lock when operating on a tocell pointed to by “tp”, the CPU must now acquire the lock given by “tp—>to_eng—>t_mutex”. The to_eng pointer determines a given tocell's locking allegiance.

[0063] Moreover, it is entirely possible that the tocell's locking allegiance may change while a CPU is attempting to acquire the tocell's lock. In this case, the CPU would find that the lock that it had acquired was no longer the correct one for operating on this tocell. Therefore, once the CPU has acquired the lock, it must check that the tocell's allegiance has not changed. It the tocell's allegiance has changed, the CPU must release the lock it just acquired and acquire the lock that the tocell now has allegiance to. Once this new lock has been acquired, the CPU must recheck the tocell's allegiance. In the preferred embodiment it does so in the following manner:

[0064] At the end of this process, the CPU is guaranteed to hold the correct lock, since it is forbidden to change a tocell's to_eng field without holding that tocell's lock. Since the CPU currently holds the correct lock, that lock will remain correct until this CPU changes it or until some time after this CPU releases the lock.

[0065] Scheduling a Function for Deferred Execution

[0066] Primitives for scheduling a function for execution on computer system 10 must be able to recognize individual CPUs and nodes. In the present embodiment, three such scheduling functions are provided: to_timeout( ), to_timeout_quad( ), and to_timeouteng( ) (they could, if desired, be combined into one). These primitives do the following. First, they select a CPU to queue a tocell on. The to_timeout( ) primitive selects by default the CPU that is actually doing/requesting the to_timeout( ) operation, while to_timeout_eng( ) and to_timeout_quad( ) allow the user to select a CPU other than the requesting processor or a specified node to execute the tocell handler. Second, they provide for selection of an alternative CPU to execute the tocell handler if the first CPU is unavailable. The to_timeout_quad( ) and to_timeout_eng( ) primitives allow the user to specify what happens if the selected CPU or node is unable to execute the timeout handler (due either to the CPU being placed offline or to the CPU becoming overloaded). The to_timeout( ) primitive defaults to TO_MIGRATE if the current CPU goes offline or becomes overloaded. Third, they respond to situations where this tocell's handler is currently running; in particular, they must handle cases where a to_untimeout( ) primitive has been invoked on a tocell since its handler started running, and they must avoid generating multiple concurrently executing handlers for a given tocell.

[0067] When invoked by a CPU, these primitives go through the following steps to accomplish the above functions:

[0068] As a result of these steps, a tocell becomes associated with a specified CPU and its data structure 60.

[0069] Modifying a Scheduled Specified Function

[0070] It is possible to move tocells among CPUs by removing them from a source CPU's timeout_eng list structure, changing the tocells' fields, and then adding them to a destination CPU's timeout_eng list structure. This might be done, for example, to balance loads among the CPUs of system 10 more evenly.

[0071] Prior timeout mechanisms such as in FIG. 1 do not balance loads on the multiple CPUs since these mechanisms have only a single, system-wide data structure. These prior mechanisms do not maintain a count of the number of tocells in a given bucket because the timeout mechanism is not CPU specific—the next available CPU executes the tocell handler. However, the present invention, in utilizing a data structure 60 for each CPU, can balance loads among the CPUs so that no CPU is overloaded where other CPUs are underutilized.

[0072] To determine which CPUs need relief, the present timeout mechanism must know how many tocells are in each CPU's t_now list.

[0073] This means that primitives for modifying functions (such as scheduling functions to_untimeout( ) and to_tryuntimeout( ) in the present embodiment) must be able to determine which list the tocell is in.

[0074] The to_untimeout( ) and to_tryuntimeout( ) primitives execute the following steps to remove a specified tocell from its current data structure 60:

[0075] Moving of Tocells from a Future Events Bucket to the Tonow List

[0076] In the prior-art timeout mechanism, the todclock( ) function moved tocells from the buckets to the tonow list. However, this is inefficient in the node-aware case, since the todclock( ) function executes only on one CPU per clock tick, and therefore, this single CPU would have to access and modify all of the other CPUs' timeout_eng structures. In doing this, it would incur expensive cache misses.

[0077] Therefore, this processing has been moved to the hardclock function in the present mechanism because hardclock( ) executes once per clock tick on each and every CPU. Therefore, it can operate on the local CPU's timeout_eng structure, and still be guaranteed that all such structures will be properly updated every tick.

[0078] The hardclock( ) timeout-related processing proceeds as follows:

[0079] When a tocell has been moved to the tonow list, its to_q pointer is changed to reflect the move. Compare the destination of the to_q pointer in FIG. 6, where it points to t_bucket[2], to its destination in FIG. 7, where it points to t_now.

[0080] Processing a Tocell

[0081] The softclock( ) primitive, well-known in the art, is used to process a tocell. The version of softclock( ) used in the prior timeout mechanism simply invokes a given tocell's handler and checks for deferred deletion. The present node-aware version, however, must check for deferred reposting and deferred untimeout of the tocell as well.

[0082] The softclock( ) function performs the following steps for each tocell on this CPU's t_now list:

[0083] Online/Offline Processing

[0084] The prior timeout mechanism takes no action when a CPU goes offline or comes back online, since the timeout data structures are global in scope and the next available CPU simply processes the tocell. However, the present node-aware version must be aware of CPUs going offline in order to avoid “stranding” tocells that were queued on a specific CPU's data structure 60. In addition, if there are any TO_NOMIGRATE tocells queued on a particular CPU's structure, then that CPU must be prevented from going offline (via EBUSY errno).

[0085] An attempt to offline a CPU goes through the following timeout-related steps:

[0086] When a CPU is brought online, it is taken through the following timeout-related steps:

[0087] Load Balancing

[0088] Prior-art timeout mechanisms do not require load balancing, since there was only one queue data structure for all tocells. However, load balancing is desirable in the node-aware case, since it is possible for a single CPU to generate more timeouts than it can process.

[0089] The softclock( ) function checks for overload at the end of each pass through its tocell-handling loop. If it detects overload, it invokes the timeout_balance_load( ) function. This function uses a weighted loading measurement as follows:

[0090] OVLD_—rruptload_weight*RRUPT_LOAD( )+OVLD_—cblen_weight*t_now.tq_count

[0091] (RRUPT_LOAD( ) is defined in a following section.) Currently, the static value of 10 is used for OVLD_rruptload_weight and 1 for OVLD_cblen_weight. This static scheme has proven sufficient thus far, but a dynamic scheme might prove necessary in the future.

[0092] The timeout_balance_load( ) function is described in Table 13:

[0093] Remote Execution

[0094] Remote execution is a request for a specified function to be invoked, as soon as possible, possibly on a particular CPU or node, with up to three specified arguments. The conceptual difference between timeout and remote execution is that remote execution can be immediate, while timeout implies a delay of at least until the next clock tick. In particular, specifying a delay of 0 to to_timeout( ) is defined by standards to be equivalent to specifying a delay of 1.

[0095] The concept of timeout predated that of remote execution by several decades. The need to maintain source-level compatibility with the old timeout mechanism, combined with the existing implementation of this mechanism, made a combined implementation timeout and remote execution unattractive. Therefore, the two mechanisms are described separately here, but only because the present embodiments are distinct. Conceptually, the two mechanisms are substantially similar. The remote-execution mechanism, for example, could simply place tocells directly on the t_now list for immediate execution.

[0096] The following table shows the analogous components of the timeout and remote-execution subsystems:

[0097] Overview of Remote-execution Data Structures

[0098] Referring to FIG. 8, an overall data structure 70 for remote execution includes a rmt_callback_t structure 72 (analogous to tocell 64) and the rmt_hdr_t structure 74 (analogous to timeout_engine structure 66). Just as for the timeouts, there is one data structure 70 per CPU. Also just as for the timeouts, the double-ended arrows indicate a circular doubly-linked list.

[0099] The rmt_callback_t structures are placed directly on the rmt_q list by the rmt_callback( ) primitive, unless that rmt_callback_t's handler is currently running, in which case the callbacks are instead placed on the rmt_incoming_q list. Any rmt_callback_t structures on the rmt_q list have their handlers invoked (and are removed from the rmt_q) by rmt_intr( ). When a handler completes executing, it may return CALLBACK_DEFER, in which case rmt_intr( ) will place its rmt_callback_t into RCB_DEFERRED state and queue it onto an appropriate rmt_q list. Otherwise, if the rmt_callback_t has been placed in RCB_DEFCB by an rmt_callback( ) that executed while the handler was running, rmt_intr( ) removes it from its rmt_incoming_q and places it on the appropriate rmt_q list for later execution.

[0100] The Remote-execution Data Structures

[0101] The rmt_callback_t structure represents one request for a function to be invoked as soon as possible, possibly on some other CPU or node. FIG. 8 shows the most pertinent field, rcb_q. This and the other fields are defined as follows (note that they all have the “rcb_” prefix):

[0102] The rcb_flags field in Table 15 has the following values (note the “RCB_” prefixes, except for the “CB_” prefixed values that are used by other subsystems):

[0103] The rmt_hdr_t structure 74 in FIG. 8 is analogous to the per-CPU timeout_eng structure 66 in FIG. 5. FIG. 8 shows its most pertinent field, rmt_q. This and the structure's other fields begin with the “rmt_” prefix and are as follows:

[0104] Selected Operations on Remote-execution Data Structures

[0105] Although there are many similarities between the posting and dispatching of timeouts and remote executions, there are some important differences. The following sections give step-by-step procedures for the rmt_callback( ), rmt_intr( ), rmt_callback_done( ), online/offline, and load-balancing operations.

[0106] Scheduling a Function

[0107] The rmt_callback( ) primitive is similar to the to_timeout( ) primitive, but there is no future events buckets array. The rmt_callback_t structures are placed directly on the rmt_q list corresponding to the selected CPU (or onto the rmt_incoming_q list when the rmt_callback( ) is invoked on an rmt_callback_t whose handler is executing). The procedure rmt_callback( ) goes through is as follows:

[0108] Invoking a Handler

[0109] The rmt_intr( ) primitive invokes handlers for previously-posted rmt_callbacks. It is invoked by software interrupt, and runs at SPL1.

[0110] The rmt_callback done( ) Operation

[0111] This function allows code outside of a given rmt_call_t structure's handler to repost the handler (using rmt_callback( ) while the current instantiation of the handler is still executing. Among other things, it substitutes RCB_DEFDONE into the rcb_flags field.

[0112] Online/Offline Processing

[0113] As with the timeout mechanisms, remote execution must handle offline and online events to avoid stranding rmt_callback_t structures. However, since there is no concept of required time delay, rmt_callback_t structures are disposed of by invoking their handlers on the CPU before it goes offline.

[0114] Offlining a CPU is handled as follows:

[0115] Onlining a CPU is handled as follows:

[0116] Load Balancing

[0117] Just as with node-aware timeouts, load balancing is required for node-aware remote execution, since it is possible for a single CPU to post more rmt_callback_t structures than it can process.

[0118] The rmt_intr( ) function checks for overload at the end of each pass through its rmt_callback_t-handling loop. If it detects overload, it invokes the rmt_load_balance( ) function. This function uses a weighted loading measurement as follows:

[0119] OVLD_—rruptload_weight*RRUPT_LOAD( )+OVLD_—cblen_weight*rmt_callbacks_waiting

[0120] Currently, the static value of 10 is used for OVLD_rruptload_weight and 1 for OVLD_cblen_weight. This static scheme has proven sufficient thus far, but a dynamic scheme might prove necessary in the future. The RRUPT_LOAD( ) primitive is described in the following section on interrupt load.

[0121] The rmt_load_balance( ) function goes through the following steps:

[0122] Load Measurement and Throttling

[0123] Traditionally, interrupts have been executed at absolute higher priority than have processes. In the larger, more complex systems found at the high end, this absolute priority causes problems such as process starvation. Some components, though implemented as processes or as lower-priority interrupts, are nonetheless critical to overall system health. It is therefore unacceptable to allow these components to be starved by a flood of “high-priority” interrupts.

[0124] Instead, interrupt priority is treated as a hint that the corresponding handler needs low latency under light load. Under heavy load, interrupts are throttled as needed to allow all components of the system to get their fair share of CPU time.

[0125] This throttling is implemented by measuring the interrupt load and by providing primitives that allow an interrupt handler to determine when it has exhausted its fair share of CPU time. The data flow through this process is shown in FIG. 9.

[0126] The portion of the diagram above the dotted line describes interrupt load measurement, and the portion below the dotted line describes interrupt throttling.

[0127] Interrupt Load Measurement

[0128] All data structures described in this section are replicated on a per-CPU basis, so that the notion of interrupt load is strictly per-CPU. There is no explicit notion of overall system interrupt load.

[0129] The current interrupt nesting level is maintained in I.intnesting. The handler start times are maintained in an array named I.intstartusclk[ ], and the cumulative handler times are maintained in an array named I.intcumusclk[ ]. These values are maintained by interrupt dispatch functions.

[0130] The rrupt_measureload( ) function takes the cumulative time spent in interrupt along with the time so far in the current interrupt and computes the maximum allowable times for each possible priority of interrupt handler. These maximum times are placed in the I.rrupt_spl_curlimit[ ] array.

[0131] The rrupt_measure_load( ) function uses the following steps to compute the values in this array:

[0132] The step response of the digital filter against time in seconds is as shown in FIG. 10.

[0133] The measured load is the fraction of time spent in an interrupt handler during the preceding clock tick, expressed as parts of 255 (so that a value of 255 indicates that 100% of the time was spent in interrupt). The filtered load is the value computed in step 3 above, again expressed as parts of 255.

[0134] The response of the adjustment equation to the range of possible filtered loads (taking the clamping action of rrupt_time_left( ) into account) is shown in FIG. 11.

[0135] Note that an interrupt handler is required to do at least one unit of work even if rrupt_time_left( ) immediately returns zero. This is necessary to ensure forward progress.

[0136] The overall effect is that as the interrupt load rises, the amount of time that an interrupt handler is permitted to run decreases. If the handler still has work to do once its time expires, it may use the rmt_callback( ) remote-execution primitives to reschedule that work to some other CPU or to a later point in time, depending on the relative loadings of the various CPUs.

[0137] Interrupt Throttling

[0138] The rrupt_time_left( ) primitive returns the amount of additional time that the caller may spend executing on the current CPU. If the filtered interrupt load is above a threshold, then rrupt_time_left( ) returns zero. Otherwise, it computes the amount of time spent in the current interrupt handler, subtracts this from the limit taken from I.rrupt_spl_curlimit[ ] array, and returns the result.

[0139] In view of the many possible embodiments to which the principles of our invention may be applied, it should be recognized that the illustrated embodiment is only a preferred example of the invention and should not be taken as a limitation on the scope of the invention. Rather, the scope of the invention is defined by the following claims. We therefore claim as our invention all that comes within the scope and spirit of these claims.