CONTROLLING PERFORMANCE/POWER BY FREQUENCY CONTROL OF THE RESPONDING NODE

Abstract

A processing node tracks probe activity level associated with its internal caching or memory system. If the probe activity level increases above a threshold probe activity level, the performance state of the processing node is increased above its current performance state to provide enhanced performance capability in responding to the probe requests. After entering the higher performance state in response to the probe activity level being above the threshold probe activity level, the processing nodes returns to a lower performance state in response to a reduction in probe activity. There may be multiple threshold probe activity levels and associated performance states.

Description

CROSS-REFERENCE TO RELATED APPLICATION(S)

This application claims the benefit of U.S. provisional application No. 61/258,798, filed Nov. 6, 2009, entitled “Controlling Performance/Power by Frequency Control of the Responding Node,” naming inventors Alexander Branover et al., which application is hereby incorporated by reference.

BACKGROUND

1. Field of the Invention

This invention relates to performance of computer systems and more particularly to performance associated with cache probes.

2. Description of the Related Art

A processing node in a computer system may be placed in any of multiple performance states (or operational states) Pn, where the particular performance state (or P-state) is characterized by an associated voltage and frequency. One factor for determining the appropriate performance state of a node is its utilization. Utilization is the ratio of the time spent by the processing node in the active (execution) state to the overall time interval over which the execution time was tracked or measured. For example, if the overall time interval was 10 milliseconds (ms) and the processor node spent 6 ms in the active (C0) state, then the utilization of the processor node is 6/10=60%. The processor node spends the remaining 4 ms in the idle (non-C0) state where code execution is suspended. A higher node utilization triggers the selection of a higher performance state P higher voltage and frequency to better address performance/watt requirements. Normally the decision to transition the processing node between performance states is made by either the operating system (OS), or high-level software, a driver, or some hardware controller. For example, if the processing node runs at a low performance state resulting in longer code execution time, the system perceives the need for a higher utilization and triggers software or hardware to transition the processing node to a higher performance state where it can complete code execution faster and spend more time in the idle state. That allows increased power savings from an overall better performance per watt. While using utilization as a trigger can provide increased performance per watt in some situations, it fails to address some issues associated with better performance per watt or preventing its degradation.

SUMMARY OF EMBODIMENTS OF THE INVENTION

Accordingly, in one embodiment a method is provided that includes tracking probe activity level in a processing node. The probe activity level is compared to a threshold probe activity level. In an embodiment, if the probe activity level is above the threshold probe activity level, the performance state of the processing node is increased above its current performance level. In an embodiment, if the probe activity level is above the first threshold probe activity level threshold and a predicted idle duration of the processing node is greater than an idle threshold, the cache memory in the processing node is flushed. In an embodiment, after entering the first performance state in response to the probe activity level being above the threshold probe activity level, the processing nodes returns to the lower performance state from which it started in response to a sufficient reduction in probe activity. In an embodiment, the sufficient reduction is to a level that is the first threshold less a hysteresis factor. In embodiments there may be multiple threshold probe activity levels and associated performance states.

In another embodiment, an apparatus includes a probe tracker to track probe activity level in a processing node. The apparatus responds to the probe activity level increasing above a first threshold probe activity level to increase a performance state of the processing node from a current performance state to a first performance state. In an embodiment, the apparatus responds to the probe activity level falling a predetermined level below the first threshold probe activity level to cause the processing node to enter a second performance state lower than the first performance state.

In an embodiment, the probe tracker includes a queue into which probe request is entered and from which a probe request in the queue is retired after the processing node responds to the probe request with at least one of a data movement and a response. In another embodiment, the probe tracker includes a counter having a count value representing probe activity level. The counter increments a count value by a predetermined amount in response to probe activity and decrements the count value by another predetermined amount in response to a passage of a predetermined period of time.

BRIEF DESCRIPTION OF THE DRAWINGS

The present invention may be better understood, and its numerous objects, features, and advantages made apparent to those skilled in the art by referencing the accompanying drawings.

FIG. 1 illustrates a multi-core processor according to an embodiment of the present invention.

FIG. 2 illustrates a flow diagram of an embodiment of the invention having a single threshold.

FIG. 3A illustrates a state diagram of an embodiment of the invention having multiple thresholds.

FIG. 3B illustrates a state diagram of an embodiment of the invention having multiple thresholds.

FIG. 4 illustrates an embodiment of the invention in which the node's caches are flushed to save power.

FIG. 5 illustrates an embodiment for tracking probe activity using an In-Flight Queue (IFQ) structure having a single threshold.

FIG. 6 illustrates an embodiment for tracking probe activity using an IFQ having multiple thresholds.

FIG. 7 illustrates another embodiment for tracking probe activity using a counter having different increment and decrement criteria.

Note that the use of the same reference symbols in different drawings indicates similar or identical items.

DESCRIPTION OF THE PREFERRED EMBODIMENT(S)

Referring to FIG. 1, a high-level block diagram illustrates a multi-core processor embodiment where each core or node includes cache memory 102 and probe control 103, which is described further herein. In the caching system of FIG. 1, each processing node in the system needs to maintain coherency in the memory by responding to probing requests (providing dirty data from cache, cache line invalidation, etc.) coming from other nodes or the input/output (I/O) domain, even if the processing node is in a low performance state or idle state. Thus, even though local copies of memory locations may be maintained in the various caches, coherency is maintained in the memory system. However, while the performance state of requesting nodes of a probe operation may be effectively controlled by evaluating utilization, that approach does not increase the performance state P of responding nodes in a direct way. The utilization-based performance control applicable to the requesting nodes leaves overall system performance vulnerable in cases where the responding node is a bottleneck.

The coherent activity in the responding node does not contribute to increased utilization of the node itself (based on the node's execution stream) since the node can be in the idle state but still respond to probe requests. Additionally, a node's execution stream can be totally independent of probe responses, and therefore coherent activity in a responding node does not lead to a higher execution utilization that normally triggers the increase of the performance state. If a responding node is in a low performance state and is probed by numerous requesting nodes, its probe responding ability (probing bandwidth), which is dependent on the clock frequency of the responding node, may turn into a performance bottleneck and start causing performance loss with respect to application threads running on requesting processing nodes. Accordingly, it is useful to identify scenarios where the probing bandwidth of a responding processing node is insufficient and to address the lack of bandwidth by prompt and controllable transition of the responding node to a higher performance state. Once the burst of probing activity is finished and extra bandwidth is no longer needed, the responding node may be transitioned back to its previous performance state dictated by its execution utilization.

One approach to addressing potential probing response bottlenecks is a software-based solution in those systems where the operating system (OS) or high-level software handling of system devices can properly tune the processor P-state. One software-based solution requires the OS or higher-level software to re-evaluate the processor P-state more frequently (in order to properly respond to a burst of activity) and thus wakes up the processor more frequently for this re-evaluation with any application. This approach likely leads to higher power consumption with an application where such frequent re-evaluation is unnecessary. Making the OS or higher-level behavior more sophisticated and not application-invariant leads to additional overhead in the idle handlers or routines (where P-state re-evaluation happens as a rule) and therefore to a higher power consumption as well. Generally speaking, the granularity of the software-based solution provides no match with a hardware-based approach and is not able to promptly identify both the start of probing activity and the end of probing activity. The latter (end of probing activity) is equally important to identify for power savings, as a processor should not be left in a higher performance state for extra time since that also leads to extra power consumption, which degrades the performance/watt.

Another solution is a hardware-based solution that provides shared voltage/clock planes for all requesting and responding nodes. Such a hardware configuration increases the frequency of the responding node (core) when the requesting node (core) increases its frequency. Slow responses of the responding node will contribute to the increased utilization of the requesting node (core). Thus, software controlling the performance state of the requesting node will increase the performance state of the requesting node, and the responding node performance state will be increased as well (due to the shared frequency and voltage planes), thus eventually increasing the probing bandwidth of the responding core. However, this approach consumes extra power in multi-core processors in situations where applications are running on only a single or a few of the nodes (cores), which is the most typical type of workload in the mobile or ultra-mobile market segments. Further, the software usually fails to respond immediately to the need for a higher clock frequency due to a utilization increase of the requesting node (core), with the time interval typically ranging from a few hundreds of microseconds to milliseconds, which can lead to performance loss over this interval.

Thus, in an embodiment of the invention, each processing node tracks its probing activity. If the level of probing activity exceeds a threshold, the performance state of the processing node is elevated to a minimal performance floor—MinPstateLimit, to address the increased requirement for probing activity bandwidth. After probing activity goes below the threshold minus associated hysteresis, the processing node transitions back to its previous performance state (P-state) in situations where its previous P-state is lower (from the performance standpoint) than the MinPstateLimit. Note that in some embodiments, the hysteresis value may be zero and in other embodiment may be fixed or programmable.

The flowchart of FIG. 2 illustrates an exemplary decision process that may operate in probe control logic 103 (see FIG. 1) according to an embodiment of the invention. In 201, the node determines if the processing unit is in a performance state that is lower than the MinPstateLimit. If it is not in a lower state, then the current performance state is sufficient to handle probing activity and the flow remains in 201. If the current performance state is lower, then in 203 the node tracks probing activity. In 205, if the probing activity is greater than the threshold, then the node elevates the performance state to MinPstateLimit in 207 and continues to track the probing activity in 208. Note that the control logic to adjust performance states is assumed, for ease of illustration, to be part of the probe control logic 103. In some embodiments, it may be separate from the probe control logic. Controlling performance states of processing nodes using voltage and frequency is well known in the art and will not be described in detail herein. If the probing activity remains above the threshold minus a hysteresis factor, the node stays in the MinPstateLimit to address the probing activity. If, however, the probing activity goes back down to a level below the threshold minus a hysteresis factor in 209, the node determines in 211 whether the prior performance state (in steps 201 and 203) was less than the MinPstateLimit. If so, the node transitions to the previous lower performance state in 213 and then returns to 201 to determine whether the current performance state is adequate to address a probe activity increase above the threshold level. Note that transition to the lower performance state does not happen in 211 if the current performance state of the processing node has been increased to MinPstateLimit or higher by the normal flow managed by software (or hardware) based on the processing node utilization factor.

The embodiment illustrated in FIG. 2 includes only one probe performance threshold addressed by the performance state MinPstateLimit. Any performance state (P-state) higher than MinPstateLimit is assumed to satisfy worst case probe bandwidth requirements. However, other embodiments can have more than one threshold associated with probe bandwidth. A higher probe bandwidth requirement requires a higher operational P-state to address the probe bandwidth limitation. Table 1 illustrates an embodiment having three performance states (P-states) corresponding to different requirements for probing bandwidth:

	TABLE 1
	P-state	Probing Activity Threshold	Hysteresis
	Pm	ProbActM	HystM
	Pn	ProbActN	HystN
	Pk	ProbActK	HystK

For the P-states, Pm>Pn>Pk. From the performance standpoint, PrbActM>PrbActN>PrbActK. The hysteresis values, HystM, HystN, and HystK may be identical, or may be different for each threshold. The hysteresis values may be configurable, along with the thresholds.

The processing node remains in P-state Pm as long as probing activity remains above (ProbeActivityM-HysteresisM). Once the probing activity drops below (ProbeActivityM-HysteresisM) and if the earlier performance state (before the increase in probe activity) is lower than Pm, the processing node transitions to a lower performance state. Note that the transition to a lower performance state does not happen if the current performance state of the processing node has been increased to Pm or higher by normal flow managed by software (or hardware) based on the processing node utilization factor.

FIGS. 3A and 3B illustrate the inter-state transitions for embodiments with more than one probe performance threshold, with each performance threshold corresponding to a different level of probing activity. The state transitions may be implemented in probe control logic 103 (FIG. 1). Once the probe activity exceeds one of the thresholds, the responding node is transitioned to the P-state corresponding to the level of probing activity. That helps to ensure that a responding node in idle state will reside in the minimal performance state (or even in the retention state) for all the time except for periods of increased probing activity where a higher performance state (P-state) is required. Referring to FIG. 3A, assume Pm (301)>Pn (303)>Pk (305)>Current P-state (307) from the frequency standpoint. Then, if an increase in probe activity occurs while in P-state 307 to a higher probe activity (Prob_Act) level, then the node may enter one of the P-states Pk, Pm, or Pn depending on the level of probe activity as described below. The following describes the transition-up of the processing node assuming the node is currently in a low power state 307.

If (Prob_Act > PrbActM), then P-state = Pm
Else If (Prob_Act > PrbActN), then P-state = Pn
Else If (Prob_Act > PrbActK), then P-state = Pk

In addition, in an embodiment, the node may transition up to a next higher-level P-state when in P-state Pn 303 or Pk 305 as shown in FIG. 3B. If the node detects increase probe activity while in P-state Pn 303, (Prob_Act>PrbActM), the node transitions via 306 to P-state Pm 301. If the node detects an increase in probe activity while in P-state Pk 305 (PrbActM>Prob_Act>PrbActN), the node transitions to P-state Pn 303 via transition 308. If the node detects an increase in probe activity while in P-state Pk 303 (Prob_Act>PrbActM), the node transitions to P-state Pm 301 via transition 310.

An additional aspect in an embodiment is to lower the P-state of the idle node to the minimal P-state if the probing activity is below the threshold. If software or hardware, responsible for utilization-based setting of the P-state of the processing node, has left it in sub-optimally high P-state (higher than MinPstateLimit), the probing P-state control function can lower the node P-state to Pmin (minimal operational P-state) or even to the retention power state so that the node can still respond to the non-bursty or lower level probing activity while saving power. The following describes the transitioning down shown in FIG. 3A based on decreased levels of probe activity (Prob_Act):

If (Prob_Act < (PrbActM-HystM) AND Prob_Act > PrbActN AND
Current P-state <Pm), then P-state = Pn
Else If (Prob_Act < (PrbActN-HystN) AND Prob_Act > PrbActK AND
Current P-state <Pn),
then P-state = Pk
Else If (Prob_Act < PrbActK-HystK AND Current P-state <Pk), then
P-state = Current P-state

Similarly, as shown in FIG. 3B, in an embodiment, the node may transition down from one P-state 303 or 305 to the appropriate P-state to reflect a decrease in probe activity. For example, while in P-state Pn 303, the node may transition to either P-state Pk 305 or the current P-state 307 depending upon probe activity. If the probe activity decreases such that (Prob_Act<PrbActN-HystN AND Prob_Act>PrbActK), the node transitions to P-state Pk 305. If the probe activity decreases while in P-state Pn 303, such that Prob_Act<PrbActK-HystK, then the node transitions to the current P-state 307. Similarly, if probe activity decreases while in P-state Pk 305, such that Prob_Act<PrbActK-HystK, then the node transitions to the P-state 307.

Thus, the control logic will transition the power state up or down based on the current probe activity level to try to match the current power state the probe activity needs. That can help avoid bottlenecks in the responding nodes while still striving to achieve power savings where available.

In another embodiment, the probing activity can trigger the flushing (write-back invalidate and disabling) of the node's caching system when the node is idle and its probing activity exceeds a threshold. That approach may be useful for multi-node systems or for nodes with relatively short cache flushing time. The decision to flush may be based on factors such as probing activity exceeding the threshold (meaning that power consumed by the responding node on the cache probing is getting higher than power associated with flushing the caching system) and node is predicted to remain idle for sufficient time. Approaches to predicting idleness include making predictions based on internal trackers and activity trackers typically in the North-Bridge (or more generally in those parts of the processor integrated circuit (the Uncore), that are not the processor cores, which typically includes such functionality as the memory controller and power management). Additionally, I/O subsystem activity predictions, e.g., interrupts, incoming or outgoing transfers, and timer-ticks, may also be utilized in the prediction of idleness and based in a separate integrated circuit (e.g., the South-Bridge). Additional details on approaches to predicting node idleness have been described, e.g., in NORTH-BRIDGE TO SOUTH-BRIDGE PROTOCOL FOR PLACING PROCESSOR IN LOW POWER STATE, naming Alexander Branover et al. as inventors, application Ser. No. 12/436,439, filed May 6, 2009, which application is incorporated herein by reference in its entirety.

FIG. 4 illustrates an exemplary flow diagram of an embodiment for cache flushing based on probe activity and node idleness prediction. In 401, if the processing node is in the idle state, then the processing node tracks probing activity in 402, and in 403, the node checks for probe activity being greater than the probe threshold. If it is, then in 405, the flow checks if the processing node idleness is predicted to be greater than an idle threshold. If so, then in 407 the processing node flushes its cache, disables its caching system, applies a retention voltage or other appropriate power savings voltage, and the system stops probing the node. If, however, the predicted duration of the node idleness is below the threshold, thus making cache flushing unattractive since it does not save power or save sufficient power, the P-state control algorithm (described above) may be applied in 409 and the node continues to track probing activity and adjust the P-state, if necessary, according to the level of probing activity.

One embodiment for tracking probe activity utilizes a queue structure referred to herein as an In-Flight Queue (IFQ) as shown in FIG. 5. The IFQ structure 500 is a multi-entry array that logically reflects the level of the probing activity. Any transaction (coherent or non-coherent) 501 is placed into an available entry of the IFQ and resides there until the eviction point. The transaction is de-allocated (evicted) from the IFQ at 503 after a response by the responding node. The response may be either the data phase (i.e. data movement from the processing node to shared memory or from shared memory to the processing node) for transactions involving data movement or after the response phase for transactions with no data movement (i.e. a request to invalidate cache entry in the local cache or memory of the processing node). The IFQ structure can be shared between processing nodes or be instantiated per processing node. The level of probing activity is represented by the number of active IFQ entries (entries which are populated with outstanding coherent requests pending completion).

In one embodiment, the node (or wherever the control functionality resides) compares the number of active IFQ entries with a single threshold 502. Note that the control functionality can reside internal or external to the node. If external to the node, it may still reside on the same die in the Uncore portion of the die as described above. If the number of entries exceeds the threshold, the transition to a higher P-state (MinPstateLimit) occurs. After the number of active IFQ entries drops to the level lower than the threshold minus hysteresis, the MinPstateLimit performance floor is cancelled and the processing node is transitioned back to the current P-state where the lower probing bandwidth can be addressed while running at lower power.

Other embodiments may utilize a multi-level IFQ-based approach shown in FIG. 6 wherein each level has an associated minimal performance level (P-state threshold) associated with a different probing bandwidth. For example, 16-entry IFQ structure 600, may have two thresholds 602 and 604 corresponding to P-states Pm and Pk, respectively, representing an increased need in probing bandwidth. The inter-state transition may be accomplished as shown in FIGS. 3A and 3B.

In other embodiments, different approaches to tracking probing activity may be used. For example, in systems having hidden, unavailable or a difficult-to-track completion phase for the probing requests, the tracking approach can be predicated on a probe-count mechanism with different increment and decrement rates. For example, referring to FIG. 7, counter 701 incremented (CNT=CNT+w_inc) every time a new probing request 703 dispatched to the processing node is identified. The count value is decremented (CNT=CNT−w_dec) every configurable time interval (IntervalTolerated) that matches the probing rate (bandwidth) associated with the specific P-state of the processing node. In an embodiment, the configurable time interval matches the maximum probing bandwidth associated with the specific P-state. Thus, it is assumed that the probing requests are serviced at a particular rate even if the actual response (data movement, response phase for transactions with no data movement) is not tracked.

Any new probing request causes counter to increment (CNT=CNT+w_inc) where w_inc is a configurable weight added to the current value of the counter. In some embodiments, the increment/decrement values may be configurable and their settings dependent on customer or higher-level software preference (performance biased, balanced or power biased). For a performance biased setting, w_inc (increment weight) is set to higher value and w_dec (decrement weight) is set to lower value. For a power savings biased setting, these parameters may be set in the opposite way. Also, the IntervalTolerated value may be configurable depending on performance/power preference of the customer or high-level software. The counter value represents the level of probing activity and is compared with ProbeActivity thresholds to figure out an optimal P-state. A higher counter value requires a higher operational P-state in order to match an increased probing bandwidth that the current P-state cannot satisfy.

A low pass filter (LPF) 705 may be used to filter out bursts of probing activity, not properly representing the workload uniformity and leading to the over-increments of the counter and choice of the performance state (P-state) that may be sub-optimal from the performance/watt standpoint. Depending on the particular embodiment, a configurable (from 1 to N) number of probe requests is tracked over configurable interval T. The low pass filter may be designed in different ways to avoid over-counting of the probing requests in case the frequency of their appearance exceeds some configurable limit over time-interval. For example, the low pass filter may be implemented to track no more than n (where 1≦n≦N) probing events over interval T. Thus, if the number of probing events >n, the counter only counts n. The low pass filter supplies the filtered probing requests to the counter.

Alternatively, the low pass filter 705 may be implemented to average the number of probing events over multiple intervals T so that if a particular interval T happens to have a high burst of activity, that high burst is limited by the average over multiple intervals. The average may be implemented, e.g., as a moving average. In one implementation, probe requests are not supplied to the counter at a higher rate than the moving average.

The implementation of the low pass filter may of course influence how the weight w_inc is determined. Thus, for example, if the average over a number of time intervals is utilized, the weight may be scaled to reflect the time interval. In other embodiments, the counter may be supplied directly with probing requests with no filtering.

Aspects of the embodiments herein may be partially implemented in software stored in volatile or non-volatile memory associated with the processor shown in FIG. 1. Software may be stored in non-volatile portions of a computer system, loaded into volatile memory and executed. Thus, embodiments of the present invention may include features or processes embodied within machine-executable instructions provided by a machine-readable medium such as nonvolatile memory. Such a medium may include any mechanism which stores data in a form accessible by a machine, such as a microprocessor or, more generally, a computer system. A machine readable medium may include volatile and/or non-volatile memory, such as read only memory (ROM); random access memory (RAM); magnetic disk storage media; optical storage media; flash memory devices; tape, or other magnetic, optical or electronic storage medium. Such stored instructions can be used to cause a general or special purpose processor, programmed with the instructions, to perform processes of the present invention.

Note that some of the processes of the present invention may include hardware operating in response to programmed instructions. Alternatively, processes of the present invention may be performed by specific hardware components containing hard-wired logic such as state machines to perform operations or by any combination of programmed data processing components and hardware components. Thus, embodiments of the present invention may include software, data processing hardware, data processing system-implemented methods, and various processing operations, as described herein.

Thus, various embodiments have been described. Note that the description of the invention set forth herein is illustrative, and is not intended to limit the scope of the invention as set forth in the following claims. Variations and modifications of the embodiments disclosed herein may be made based on the description set forth herein, without departing from the scope of the invention as set forth in the following claims.

Claims

1. A method comprising:

tracking probe activity level in a processing node; and

comparing the probe activity level to a first threshold probe activity level.

2. The method as recited in claim 1 further comprising:

increasing a performance state of the processing node to a first performance state higher than a current performance state if the probe activity level is above the first threshold probe activity level.

3. The method as recited in claim 2 further comprising after entering the first performance state in response to the probe activity level being above the threshold probe activity level, entering a second performance state lower than the first performance state when the probe activity level falls below a predetermined level below the first threshold probe activity.

4. The method as recited in claim 3 further comprising entering the second performance state when the probe activity level is less than the first threshold less a hysteresis factor.

5. The method as recited in claim 3 wherein the second performance state is a performance state from which the processing node entered the first performance state.

6. The method as recited in claim 3 wherein the first and second performance states are defined by at least one of voltage and frequency.

7. The method as recited in claim 2 further comprising increasing the performance state of the processing node to a third performance state in response to the probe activity level increasing above a second threshold probe activity level, the second threshold probe activity level being higher than the first threshold probe activity level, and wherein the third performance state is higher than the first performance state.

8. The method as recited in claim 7 further comprising after increasing the performance state of the processing node to the third performance state in response to the probe activity level increasing above the second threshold probe activity level, reducing the performance state.

9. The method as recited in claim 1 further comprising beginning the tracking of the probe activity level when the processing node is in a performance state lower than the first performance state.

10. The method as recited in claim 1 wherein tracking the probe activity further comprises entering each probe request into a queue and retiring a probe request from the queue after the processing node responds to the probe request with at least one of a data movement and a response.

11. The method as recited in claim 10 further comprising comparing a number of entries in the queue to the first threshold probe activity level to determine if the probe activity level is above the first threshold probe activity level.

12. The method as recited in claim 1 wherein tracking the probe activity further comprises incrementing a count value indicative of a level of probe activity in response to occurrence of probe activity and decrementing the count value based on passage of a predetermined amount of time.

13. An apparatus comprising:

a probe tracker to track probe activity level in a processing node;

wherein the apparatus is responsive to increase a performance state of the processing node from a current performance state to a first performance state if the probe activity level increases above a first threshold probe activity level.

14. The apparatus as recited in claim 13 wherein the apparatus is responsive to the probe activity level falling a predetermined level below the first threshold probe activity level to cause the processing node to enter a second performance state lower than the first performance state.

15. The apparatus as recited in claim 14 wherein the first and second performance states are defined by at least one of voltage and frequency.

16. The apparatus as recited in claim 13 wherein the apparatus is further operable to increase the performance state of the processing node to a third performance state in response to the probe activity level increasing above a second threshold probe activity level, the second threshold probe activity level being higher than the first threshold probe activity level, and wherein the third performance state is higher than the first performance state.

17. The apparatus as recited in claim 13 wherein the probe tracker is responsive to the node being in a performance state below the first performance state to begin the tracking of the probe activity level.

18. The apparatus as recited in claim 13 wherein the probe tracker further comprises a queue into which probe request is entered and from which a probe request in the queue is retired after the processing node responds to the probe request with at least one of a data movement and a response.

19. The apparatus recited in 18 further wherein the apparatus is operable to compare a number of entries in the queue to the first threshold probe activity level to determine if the probe activity level is above a first threshold probe activity level.

20. The apparatus as recited in claim 13 wherein the probe tracker comprises a counter responsive to probe activity to increment a count value indicative of a level of probe activity and responsive to a passage of a predetermined period of time to decrement the count value.