CN102662914B - Method for configuring heat sensor of microprocessor - Google Patents

Abstract

The invention relates to a method for configuring a heat sensor of a microprocessor, and the method comprises the steps of running a plurality of operating modes on the microprocessor, calculating the power consumption of the microprocessor under different operating modes by utilizing a power consumption calculating unit, and calculating the temperature distribution under the power consumption by utilizing a temperature distribution calculating unit; implementing data processing for the temperature distribution obtained under different operating modes, and acquiring a hot-spot distribution stacking chart of each module of the microprocessor; setting an upper limit value for the hot-spot monitoring error, performing the optimized double clustering considering singular points for data points in the hot-spot stacking chart according to the upper limit value, and obtaining a clustered result of the hot spot; and configuring one heat sensor for each clustering according to the clustered result, placing the heat sensor on a mass center of all hot spots contained by the clustering, namely, an average point of weight of the temperature value, and ensuring that the temperature of all hot spots are monitored within the set maximal error. Compared with the prior art, the method has advantages of high measurement precision and low cost.

Description

A kind of microprocessor heat Way of Sensor Deployment

Technical field

The present invention relates to a kind of Way of Sensor Deployment, especially relate to a kind of microprocessor heat Way of Sensor Deployment.

Background technology

In modern high performance circuit, due to developing rapidly of miniaturization and large-scale integrated micronizing technology, the power density of microprocessor becomes more and more higher.These power attenuations change into heat makes chip temperature rise, and causes the increase of the reliability decrease of microprocessor, electricity leakage power dissipation, cooling cost rising.In order to better understand microprocessor ruuning situation, a lot of microprocessor manufacturers all adopts the thermal sensor reading that sheet is implanted to assess the temperature conditions of the current operation of chip in heat monitoring.Such as intel pentium 4 processor, it is just equipped with thermal sensor, just can cause warning when temperature exceedes setting working temperature, and processor, after obtaining these and reporting to the police, just reduces power consumption by closing clock temporarily.Therefore, the emphasis that thermal sensor quantity and implantation position become microprocessor heat management design how is distributed.

By finding prior art literature search, Ryan Cochran etc. proposed in 2009 a kind of utilize the spectrum technology in signal transacting by equally distributed sensor sample to dsc data be reconstructed, recover 16 core microprocessor heat distributions to reach the object (Spectral Techniques for High-Resolution Thermal Characterization with Limited Sensor Data) to whole microprocessor heat monitoring, the document gives a kind of high-precision microprocessor heat method for supervising, but be limited by the equally distributed requirement of sensor, there is certain limitation.Mukherjee etc. are for microprocessor each module focus monitoring problem, propose the microprocessor heat sensor assignment based on k-means clustering algorithm and Placement Strategy (Systematic temperature sensor allocation and placement for microprocessors), the document has cast aside the whole heat distribution of microprocessor, focal point is placed on the focus of microprocessor modules, provides Feasibility Solution for emergency mechanism designs with heat management.

By finding the retrieval of prior art, although these methods can provide a kind of focus to monitor solution, under high-precision requirement, the number of sensors adopted is still more, is unfavorable for saving design cost and manufacturing cost.Especially for the microprocessor of some high-performance high integration, due to technological requirement, a large amount of sensors cannot be implanted therein.How can ensure that reducing number of sensors when precision is substantially constant is still a problem demanding prompt solution.

Summary of the invention

Object of the present invention is exactly provide a kind of under the high-precision requirement of guarantee to overcome defect that above-mentioned prior art exists, effectively reduces number of sensors, the microprocessor heat Way of Sensor Deployment reduced costs.

Object of the present invention can be achieved through the following technical solutions:

A kind of microprocessor heat Way of Sensor Deployment, the method comprises the following steps:

1) run multiple operating mode on the microprocessor, power consumption calculation unit calculates the power consumption of microprocessor under different operating mode, and Temperature Distribution computing unit calculates the Temperature Distribution of microprocessor under this power consumption;

2) data processing is carried out to the temperature profile obtained under different operating mode, obtain microprocessor each module hotspot's distribution stacking diagram;

3) set focus monitoring error higher limit, and according to this higher limit, the dual cluster of optimization considering singular point is carried out to the data point in hotspot's distribution stacking diagram, obtain the cluster result of these focuses;

4) according to above-mentioned cluster result, each cluster configures one piece of thermal sensor, thermal sensor is arranged on the barycenter place of all focuses contained by this cluster, namely considers the mean center place of temperature value weighting, ensures to monitor all hot(test)-spot temperature values in setting maximum error.

Described step 1) in the Temperature Distribution of microprocessor two-dimension temperature matrix when referring to microprocessor work state.

Described step 2) in hotspot's distribution stacking diagram refer to: according to microprocessor architecture design Module Division, choose the focus of each module, again all hotspot's distributions under different operating mode are overlapped on a width Organization Chart, reject the hotspot's distribution obtained after repeating focus.

Described step 3) in focus error higher limit refer to: in microprocessor heat management design, the thermal sensor sample magnitude allowed with monitor the maximum difference of focus actual temperature.

Described step 3) in dual cluster refer to the clustering algorithm simultaneously carried out in spatial domain and Attribute domain.

Described step 3) in the dual cluster of optimization of consideration singular point refer to: dual clustering algorithm, when initialization cluster centre, is chosen temperature value and is differed maximum data point compared with average temperature value.

Compared with prior art, the present invention has the following advantages:

1) for heat distribution feature, adopt the self-organization bunch resolution policy merged, improve cluster efficiency, reduce time cost, when many focuses, analysis speed is fast;

2) adopt the dual clustering algorithm of improvement considering singular point, under equal accuracy requires, the thermal sensor quantity of use is compared with other schemes existing and is significantly reduced, and cost is lower.

Accompanying drawing explanation

Fig. 1 is flow chart of steps of the present invention.

Embodiment

Elaborate to embodiments of the invention below in conjunction with accompanying drawing, the present embodiment is implemented under premised on invention technical scheme herein, give detailed embodiment and concrete operating process, but protection scope of the present invention is not limited to following embodiment.

Embodiment 1

As shown in Figure 1, a kind of microprocessor heat Way of Sensor Deployment, the method comprises the following steps:

The first step, take Alpha EV6 microprocessor architecture design, run multiple operating mode on the microprocessor, SimpleScalar software emulation SPEC2000 (System Performance Evaluation Corporation) Standard test programme is used to simulate different operating mode, and by the power consumption of power consumption calculation unit computing module, Temperature Distribution computing unit calculates the Temperature Distribution of microprocessor under this power consumption;

Above-mentioned power consumption calculation unit is Wattch infrastructure software, and Temperature Distribution computing unit is Hotspot software.

Second step, carrying out focus merging and rejecting by emulating the temperature profile obtained under different test procedure by MATLAB, obtaining microprocessor each module hotspot's distribution stacking diagram, amounts to 20 focuses;

Described focus merges and rejecting refers to: being placed on obtaining hot spot data in all hotspot's distribution figure on same Organization Chart, if there is hotspot location to overlap completely, so only retains one, other coincide points being rejected.

3rd step, is set as 5% by maximum focus monitoring error, and utilize and consider that the dual clustering algorithm of improvement of singular point carries out the self-organization bunch Fusion of Clustering operation of all focuses, its algorithm frame is as follows:

1) initialization one Ge Lei center, this center is a point maximum with average difference in all data points to be clustered, with this center for the adjacent with it data point of reference scan, whether the attributive distance calculating these data points and center exceedes threshold value, if attributive distance is less than threshold value, then merge into one bunch, if attributive distance is greater than threshold value, then nonjoinder;

2) when occurring to merge, immediately using the data point set be newly merged into as new class, until there is not any merging in scanning all adjacent data points around again;

3) if one bunch cannot merge with periphery again, then this bunch is exactly the complete cluster of required by us one;

4) again circulate execution 1,2,3 step in remaining data point, until all data points all belong to certain cluster.

Consider that the benefit of singular point is, in the process of cluster, it is a class that the focus that temperature value difference compared with average temperature value is larger is often difficult to gather with most of focus, and coverlet Du Gu stands, cause the situation needing to distribute a sensor for it is independent, preferential can as much as possible can the permeate point of class of periphery be distributed into this cluster, to reduce number of sensors using it as cluster centre.

Above-mentioned adjacent data point refers to: according to hotspot's distribution superposition of data, make the VORONOI figure of these focus dot matrix, if the VORONOI polygon at certain two data point place meets Queen criterion, namely there is public vertex in two polygons, so, these two data points are then judged as consecutive number strong point.

Described attributive distance refers to: between two data points, the Euclidean distance of non-spatial attributes.Suppose N number of d dimension strong point P _n={ g _n ¹... g _n ^d, a _n ¹... a _n ^m, 1≤n≤N; Wherein g _n ¹... g _n ^dfor the locus coordinate of this point, a _n ¹... a _n ^mfor the non-spatial attributes value of this point.Then there is any two points P ₁, P ₂attributive distance be:

$D=\sqrt{\underset{m}{\mathrm{\Σ}}{({a_1}^i-{a_2}^i)}^2}$

In the present embodiment, because non-spatial attributes value only has temperature, therefore attributive distance and point-to-point transmission temperature difference.

Described threshold value refers to: one judges whether two data points can gather is the variable of a class, is the key that focus monitoring error is no more than setting maximal value when ensureing to distribute sensor according to cluster result, so its threshold value itself is relevant to maximum heat point tolerance setting value.Suppose ε _maxbe the maximum error ratio that engineering allows, n is the number of data points that intra-cluster has contained, a _iit is the non-spatial attributes value of all data points in cluster.Then have:

Here α is improvement factor, is used for balance sensor quantity and error, and should choose flexibly according to different maximum error setting value, the present embodiment maximum error setting value is that 5%, α value chooses 2.0.

4th step, the number of clusters finally obtained by above-mentioned steps is 4, and each cluster distributes one piece of thermal sensor, namely monitors all focuses according to the maximum error of 5%, and required thermal sensor quantity is 4.The position of sensor is just placed on the barycenter place of all focuses contained by this cluster, namely considers the mean center place of temperature value weighting.

Embodiment 2

Shown in figure 1, the concrete steps of method are with embodiment 1, difference is, setting maximum monitoring error is respectively 2%, 3% and 4%, and the number of sensors that the hot monitoring scheme adopting the inventive method and existing k-means clustering technique to obtain adopts compares, result is as shown in table 1, and this illness that has not attacked the vital organs of the human body understands that the inventive method significantly reduces in equal accuracy requirement lower sensor quantity.When particularly accuracy requirement is higher, adopt the inventive method improved efficiency clearly.(accuracy requirement of 2% due to close to the limit of error, therefore promotes not too obvious.)

Table 1

Claims (5)

1. a microprocessor heat Way of Sensor Deployment, is characterized in that, the method comprises the following steps:

1) run multiple operating mode on the microprocessor, power consumption calculation unit calculates the power consumption of microprocessor under different operating mode, and Temperature Distribution computing unit calculates the Temperature Distribution of microprocessor under this power consumption;

2) data processing is carried out to the temperature profile obtained under different operating mode, obtain microprocessor each module hotspot's distribution stacking diagram;

3) set focus monitoring error higher limit, and according to this higher limit, the dual cluster of optimization considering singular point is carried out to the data point in hotspot's distribution stacking diagram, obtain the cluster result of these focuses;

4) according to above-mentioned cluster result, each cluster configures one piece of thermal sensor, thermal sensor is arranged on the barycenter place of all focuses contained by this cluster, namely considers the mean center place of temperature value weighting, ensures to monitor all hot(test)-spot temperature values in setting maximum error;

Described step 3) in the dual cluster of optimization of consideration singular point refer to: dual clustering algorithm, when initialization cluster centre, is chosen temperature value and is differed maximum data point compared with average temperature value, be specially:

301) initialization one Ge Lei center, this center is a point maximum with average difference in all data points to be clustered, with this center for the adjacent with it data point of reference scan, whether the attributive distance calculating these data points and center exceedes threshold value, if attributive distance is less than threshold value, then merge into one bunch, if attributive distance is greater than threshold value, then nonjoinder;

302) when occurring to merge, using the data point set be newly merged into as new class, until there is not any merging in scanning all adjacent data points around again;

303) if one bunch cannot merge with periphery again, then this bunch of individual complete cluster;

304) again circulate execution 301,302,303 step in remaining data point, until all data points all belong to certain cluster;

The computing formula of described threshold value is:

$D_{\max }={\mathrm{\α}\×\mathrm{\ϵ}}_{\max }\×\frac{1}{n}\underset{n}{\mathrm{\Σ}}{a}_i$

In formula, α is improvement factor, ε _maxbe focus monitoring error higher limit, n is the number of data points that intra-cluster has contained, a _ithe non-spatial attributes value of all data points in cluster, the i.e. temperature of each data point.

2. a kind of microprocessor heat Way of Sensor Deployment according to claim 1, is characterized in that, described step 1) in the Temperature Distribution of microprocessor two-dimension temperature matrix when referring to microprocessor work state.

3. a kind of microprocessor heat Way of Sensor Deployment according to claim 1, it is characterized in that, described step 2) in hotspot's distribution stacking diagram refer to: according to microprocessor architecture design Module Division, choose the focus of each module, again all hotspot's distributions under different operating mode are overlapped on a width Organization Chart, reject the hotspot's distribution figure obtained after repeating focus.

4. a kind of microprocessor heat Way of Sensor Deployment according to claim 1, it is characterized in that, described step 3) in focus monitoring error higher limit refer to: in microprocessor heat management design, the thermal sensor sample magnitude allowed with monitor the maximum difference of focus actual temperature.

5. a kind of microprocessor heat Way of Sensor Deployment according to claim 1, is characterized in that, described step 3) in dual cluster refer to the clustering algorithm simultaneously carried out in spatial domain and Attribute domain.