Arbitration technique based on processor task priority

[00002] Not applicable.

[00003] Not applicable.

[00004] 1. Field of the Invention

[00005] The present invention generally relates to a technique for arbitration (or message routing) among multiple pending requests for system resources in a computer system. Still more particularly, the invention relates to an arbitration technique based on processor task priority.

[00006] 2. Background of the Invention

[00007] Modem computer system generally include a plurality of devices interconnected through a system of buses which are linked by way of one or more bridge logic units. For example, a conventional computer system typically contains one or more central processing unit (“CPUs”) coupled through a host bridge to a main memory unit. A CPU bus usually couples the CPU(s) to the host bridge, and a memory bus connects the bridge to the main memory. The bridge logic typically incorporates a memory controller which receives memory access requests (such as from the CPUs) and responds by generating standard control signals necessary to access the main memory. The bridge logic may also include an interface to another bus, such as the Peripheral Component Interconnect (“PCI”) bus. Examples of devices which link to such a bus include network interface cards, video accelerators, audio cards, SCSI adapters, and telephony cards, to name a few.

[00008] Because a conventional computer system includes multiple interconnected devices that function independently of each other, they often attempt to concurrently access common resources. For example, in a system having multiple CPUs, more than one CPU may need to access main memory at a given time. By way of additional example, a device coupled to the PCI bus may need to extract data from main memory at the same time that the CPU is requesting instructions stored in the main memory. Since, main memory generally can only respond to a single memory request at a time, it is generally the function of the memory controller to choose which device to service first. Such conflicts necessitate “arbitration,” in which the various pending memory requests are ranked with the highest ranking requests generally being serviced first.

[00009] There are many well-known arbitration techniques. For instance, according to a fixed priority scheme, each type of cycle request (e.g., CPU to memory write, PCI to memory write, CPU read from memory, etc.) is assigned a predetermined ranking. Although some cycle requests may have the same ranking, in general some cycle requests will have rankings that are higher than other types of requests. Using such a fixed priority scheme, a memory controller, if faced with multiple pending memory access requests, simply grants memory access to the device with the highest ranking. Although simple to implement, the deficiency of this type of arbitration scheme is that a low ranking pending request may not be permitted to complete because numerous higher ranking requests are pending. This condition is called “starvation.”

[00010] Another arbitration technique is the Least-Recently-Used (“LRU”) algorithm. In the LRU algorithm a memory arbiter grants the request which has least recently been granted (i.e., the “oldest” request). This type of arbitration technique ensures that no one device or cycle request is starved from completing in favor of higher ranking requests. The downside of this technique is that it essentially equalizes, or fixes, the priority of all devices in the computer system, since the arbitration scheme does not take into account the urgency associated with memory transactions from certain devices. That is, the newest request may be far more critical and time sensitive than the older requests, but will not be permitted to run until all older requests have run. Further, the devices which use memory infrequently actually tend to experience shorter waits for memory access, since these devices are less likely to have recently accessed memory than are devices which access memory more frequently. As a consequence, real-time applications and devices, which need frequent and quick access to memory, may consistently lose memory arbitration to other devices under an LRU scheme. Hence, an LRU scheme, while more equitable that a fixed scheme, lacks the flexibility to allow the computer system designer to directly set the memory request priorities.

[00011] It would thus be advantageous to have an arbitration scheme that addresses the problems noted above. Despite the apparent advantages that such a system would provide, to date no such system has been developed that addresses the foregoing problems.

[00012] The problems noted above are solved in large part by having multiple CPUs inform other logic in a computer system as to the priority level (e.g., task priority) associated with the CPU or software executing thereon. The logic makes arbitration decisions regarding CPU transactions based, at least in part, on the task priorities of the various CPUs.

[00013] In one embodiment, the logic comprises a host bridge to which a plurality of CPUs couple, as well as memory and one or more peripheral devices. Each CPU may issue a special cycle to the host bridge to inform the host bridge of that CPU's task priority. The host bridge stores the task priorities of the various CPUs and uses that information when making arbitration decisions regarding how to allocate system resources (e.g., memory, peripheral devices) to multiple pending CPU transactions.

[00014] In accordance with another embodiment, the logic comprises a switch to which multiple nodes couple. Each node may include one or more CPUs and the various nodes send messages back and forth to other nodes through the switch. Each message from a node may include a priority value pertaining to that message. The switch makes decisions as to how to route message between nodes based on the priority values.

[00015] Numerous variations on how to route the node messages or arbitrate the CPU cycles are possible. For example, the routing/arbitration decision may be made solely based on the task priorities. Alternatively, the routing/arbitration decision may be made based on the task priorities in conjunction with an anti-starvation algorithm, a tie breaking algorithm, and other criteria that must be met. Further still, criteria unrelated to task priorities can be used as the primary criteria and the task priorities can be used as tie breaking criteria.

[00016] These and other advantages will become apparent upon reviewing the following disclosures.

[00021] Referring now to FIG. 1, system 100 is shown constructed in accordance with a preferred embodiment of the invention. As shown, system 100 includes one or more central processor units (“CPUs”) 102 (also labeled as CPU0-CPU3), a host bridge 110, main memory 120, one or more peripheral devices 130, a south bridge 140 and various input and output devices 142, 144. The host bridge couples to the CPUs 102, memory 120 and the various peripheral device 130 via the south bridge 140. Other architectures are possible also; the architecture in FIG. 1 is merely exemplary of one suitable embodiment.

[00022] The peripheral device 130 may be whatever devices are appropriate for a given computer 100 such as a modem, network interface card (“NIC”), etc. The peripheral devices 130 couple to the host bridge 110 via any suitable interconnect bus 132. Of course, devices 130 are compliant with whatever interconnect bus 132 is used.

[00023] In general, one or more of the CPUs 102 can read data or instructions from and write data to memory 120 via the host bridge 110. Similarly, the peripheral devices 130 also can read/write memory 120. Further still, the CPUs 102 can run cycles through the host bridge 110 that target one or more of the peripheral devices 130. Additionally, signals to/from the input and output devices 142, 144 may propagate through the south bridge 140 and the host bridge 110 as desired. The input device 142 may be any type of device such as a mouse or keyboard and the output device may be a disk drive or other type of output device. These input/output devices may connect directly to the south bridge 140 or couple to the bridge 140 via other intervening logic (not shown) or couple to the system 100 via other architectures.

[00024] Referring still to FIG. 1, in accordance with the preferred embodiment of the invention, each CPU 102 can be assigned a “task priority.” The task priority is assigned to each CPU or groups of CPUs preferably by software and may be changed as desired. The task priority may take many forms. One suitable form is a value in the range of 0 to 15. As such, there are 16 different task priorities with task priority 15 being the highest priority and 0 the lowest, or vice versa. Task priorities 0-15 represent various gradations of priority between highest and lowest. The task for which a task priority is assigned to a CPU 102 may be a program, part of a program, or low level functions such as a single memory read/write. Software can program a CPU 102 to a certain task priority by writing a task priority value to a register 106 within each CPU 102. Each CPU 102 is thus independently configurable to a desired task priority. Multiple CPUs 102 can be assigned to the same task priority if desired.

[00025] Each CPU 102 can run a cycle on bus 104 to the host bridge 110 by which the CPU informs the bridge of that CPU's current task priority. The cycle through which a task priority is transmitted to the host bridge 110 may be a cycle separate from a CPU request for a system resource such as a memory access. Alternatively, a CPU may transmit a task priority in the same cycle as the CPU request for a system resource. As shown, the host bridge 110 includes a task priority table 112, or other form of storage, in which the bridge stores the task priorities received from the CPUs 102. The task priority table 112 may include an entry 114 for each CPU 102. In the example of FIG. 1, the system 100 includes four CPUs (CPU0-CPU1) and accordingly, task priority table 112 in the host bridge 110 includes four entries, one entry corresponding to each of the four CPUs. Each entry 114 includes the ability to store the task priority for the corresponding CPU. Thus, the first entry 114 stores the CPU 0 task priority as shown, the second entry stores the CPU 1 task priority, and so on.

[00026] At any time during power on self test (“POST”) or during run-time, the CPUs 102 may inform the host bridge of their task priorities and may update the task priorities at subsequent times. In accordance with the preferred embodiment, the host bridge 110 uses the CPU task priorities to make decisions about granting individual CPUs access to memory or other resources within the computer system. This technique preferably selects only between competing CPUs for system resources, and not for non-CPU related cycles such as peripheral device 130 writes to memory. However, the concept explained herein can easily be extended to devices other than CPUs and the claims which follow should be interpreted broadly enough, unless otherwise indicated by the language of the claim itself, not to be limited to just CPU-based cycles.

[00027] A non-exhaustive list of the use of CPU task priorities in making the arbitration decision with respect to CPU cycles includes:

• [00028] 1. Use task priority as the sole arbitration criterion
• [00029] 2. Use task priority as the primary arbitration criterion coupled with an anti-starvation algorithm
• [00030] 3. Use task priority as the primary arbitration criterion coupled with a tie-breaking algorithm
• [00031] 4. Use criteria unrelated to task priorities as the primary criteria, but use the CPU task priorities as way to break a tie between two or more pending CPU cycles
• [00032] 5. Use task priority coupled with other arbitration criteria that must be met

[00033] The first algorithm is self-explanatory and states that, among multiple CPU cycles pending at the host bridge 110, the host bridge preferably selects the cycle to run that corresponds to the CPU having the highest task priority. This algorithm may have limited use, however, in that starvation may occur and that two or more CPUs may have pending cycles to run that are at an equal task priority.

[00034] The second algorithm listed above states that an anti-starvation technique is used in conjunction with the first algorithm. Any suitable anti-starvation algorithm can be used such as that described in U.S. Pat. No. 6,286,083, incorporated herein by reference. Accordingly, with this approach the CPU having the highest task priority is always selected, but other CPUs may be selected to avoid starvation should the condition so warrant.

[00035] The third algorithm says that the CPU with the highest task priority is selected by the host bridge, and that a tie breaking algorithm is used in the event that two more CPUs have the highest, yet equal, task priority. A suitable tie breaking algorithm could be a fixed priority technique such as CPU writes to memory are always selected over CPU reads.

[00036] In accordance with the fourth algorithm, other conventional arbitration techniques that are not based on CPU task priorities are used as the primary arbitration decision making technique. The host bridge 110 uses the CPU task priorities as a mechanism to break a deadlock in the event such conventional arbitration techniques are unable to select between two or more CPU cycles. That is, this technique embodies the algorithm that, all else being equal, the CPU with the highest task priority gets selected.

[00037] The fifth algorithm listed above is that CPU task priorities are used in conjunction equally with other arbitration criteria that must be met. For example, the system designer may want certain cycles to always happen in a predetermined order regardless of task priority. An example of this is when one CPU 102 having a relatively low task priority wants to write a block of data to memory 120 and another, higher task priority CPU wants to read that same data. Although the reading CPU has a higher task priority, in this situation it is generally desirable that the higher task priority reading CPU not be allowed to read the data until all of the data is written to memory by the lower priority writing CPU. The fifth algorithm thus takes into account that CPU cycles corresponding to higher task priorities should generally be permitted to run before lower task priority cycles, but certain types of predetermined activities should be permitted to occur in a different order.

[00038] The computer system 100 shown in FIG. 1 can represent a “node” in a system in which multiple such nodes are coupled together. Further, the aforementioned use of CPU task priorities can be extended to the embodiment shown in FIG. 2 in which four nodes N0-N3 are coupled together via a switch 150. Each node represents a collection of electronic components such as the combination shown in FIG. 1, although difference components can be implemented for each node N0-N3. The switch 150 is conceptually similar to the host bridge 110 in FIG. 1 in that the switch permits each node 100 to communicate with any of the other nodes.

[00039] The switch 150 can use task priorities from each node when deciding how to route the messages between nodes. Accordingly, each node creates a message to send to another node (e.g., a write message, a read message, or a control message) and includes with the message a task priority. This type of task priority may pertain to the message itself or may pertain to a particular CPU within the node that sends the message. The switch 150 stores the messages from each node N0-N3 along with their task priorities and makes decisions on which messages to route through the switch based on the associated task priorities. The same or similar decision algorithms as explained above with regard to FIG. 1 can be implemented in switch 150.

[00040] Thus, the preferred embodiments make use of task priorities when deciding when to route a message through a system. Broadly stated, a plurality of CPUs (or nodes) couple through logic (e.g., a host bridge, a switch, etc.) to one or more system resources (e.g., memory, NICs, modems, other CPUs or nodes, etc.) and the logic uses information associated with each CPU that indicates the priority level of each CPU or software executing on each CPU. Allocation of system resources is accordingly weighted in favor of the most critical or important activities. Because the priority associated with a given transaction is taken into account, overall system performance should be improved as the more critical activities are given heightened priority when competing for common system resources.

[00041] The above discussion is meant to be illustrative of the principles and various embodiments of the present invention. Numerous variations and modifications will become apparent to those skilled in the art once the above disclosure is fully appreciated. It is intended that the following claims be interpreted to embrace all such variations and modifications.