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

Abstract

The invention relates to a method for configuring a thermal sensor of a microprocessor. The method includes the following steps: running multiple working conditions on the microprocessor, a power consumption calculation unit calculating the power consumption of the microprocessor under different working conditions, and a temperature distribution calculation unit Calculate the temperature distribution of the microprocessor under the power consumption; perform data processing on the temperature distribution diagrams obtained under different working conditions, and obtain a superimposed map of hot spot distribution of each module of the microprocessor; set the upper limit of the hot spot monitoring error, and based on the above The limit value is to optimize the double clustering of the data points in the superposition map of the hot spot distribution considering the singular point, and obtain the clustering results of these hot spots; according to the above clustering results, each cluster is equipped with a thermal sensor, and the thermal sensor is set in the The centroid of all hotspots included in the cluster, that is, the weighted average center of temperature values, ensures that all hotspot temperature values are monitored within the set maximum error. Compared with the prior art, the invention has the advantages of high measurement precision, low cost and the like.

Description

A microprocessor thermal sensor configuration method

technical field

The invention relates to a sensor configuration method, in particular to a microprocessor thermal sensor configuration method.

Background technique

In modern high-performance circuits, due to the rapid development of miniaturization and large-scale integration process technology, the power density of microprocessors is becoming higher and higher. These power losses are converted into heat to increase the temperature of the chip, resulting in a decrease in the reliability of the microprocessor, an increase in leakage power consumption, and an increase in cooling costs. In order to better understand the operation of the microprocessor, many microprocessor manufacturers use the readings of the thermal sensor implanted on the chip to evaluate the current operating temperature of the chip in thermal monitoring. For example, the Intel Pentium 4 processor is equipped with a thermal sensor, which will trigger an alarm when the temperature exceeds the set operating temperature. After the processor receives these alarms, it will reduce power consumption by temporarily turning off the clock. Therefore, how to allocate the number of thermal sensors and their implantation locations has become the focus of microprocessor thermal management design.

By searching the existing technical literature, it was found that Ryan Cochran et al. proposed in 2009 a method of using spectrum technology in signal processing to reconstruct the thermal data sampled by uniformly distributed sensors and restore the thermal distribution of the 16-core microprocessor. To achieve the purpose of thermal monitoring of the entire microprocessor (Spectral Techniques for High-Resolution Thermal Characterization with Limited Sensor Data), this document provides a high-precision microprocessor thermal monitoring method, but is subject to the requirements of uniform distribution of sensors, There are certain limitations. Mukherjee et al. proposed a microprocessor thermal sensor allocation and placement strategy (Systematic temperature sensor allocation and placement for microprocessors) based on the k-means clustering algorithm for the hot spot monitoring problem of each microprocessor module. The entire heat distribution focuses on the hot spots of each module of the microprocessor, providing a feasible solution for emergency mechanism design and thermal management.

Through the search of the prior art, it is found that although these methods can provide a hot spot monitoring solution, under the requirement of high precision, the number of sensors used is still large, which is not conducive to saving design cost and manufacturing cost. Especially for some high-performance and high-integration microprocessors, due to process requirements, it is impossible to implant a large number of sensors inside them. How to reduce the number of sensors while keeping the accuracy basically unchanged is still an urgent problem to be solved.

Contents of the invention

The object of the present invention is to provide a microprocessor thermal sensor configuration method that effectively reduces the number of sensors and reduces the cost under the requirement of ensuring high precision in order to overcome the above-mentioned defects in the prior art.

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

A microprocessor thermal sensor configuration method, the method comprises the following steps:

1) Running multiple working conditions on the microprocessor, the power consumption calculation unit calculates the power consumption of the microprocessor under different working conditions, and the temperature distribution calculation unit calculates the temperature distribution of the microprocessor under the power consumption;

2) Perform data processing on the temperature distribution diagrams obtained under different working conditions, and obtain superimposed diagrams of hot spot distribution of each module of the microprocessor;

3) Set the upper limit value of the hotspot monitoring error, and according to the upper limit value, consider the optimized double clustering of the data points in the hotspot distribution overlay graph to obtain the clustering results of these hotspots;

4) According to the above clustering results, each cluster is configured with a thermal sensor, and the thermal sensor is set at the centroid of all hot spots contained in the cluster, that is, at the average center considering the weighted temperature value, to ensure that it is within the set maximum error Monitor all hot spot temperature values.

The temperature distribution of the microprocessor in the step 1) refers to the two-dimensional temperature matrix in the working state of the microprocessor.

The hotspot distribution overlay map in the described step 2) refers to: according to the division of microprocessor architecture modules, pick out the hotspots of each module, and then overlap all hotspot distributions under different working conditions on an architecture diagram, and eliminate duplicates. Hotspot distribution obtained after hotspot.

The upper limit of the error of the hot spot in step 3) refers to the maximum difference between the allowed sampling value of the thermal sensor and the actual temperature of the monitored hot spot in the thermal management design of the microprocessor.

The dual clustering in the step 3) refers to a clustering algorithm performed on the space domain and the attribute domain at the same time.

The optimized dual clustering considering singular points in the step 3) refers to: when the dual clustering algorithm initializes the cluster centers, select the data point whose temperature value is the largest difference from the average temperature value.

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

1) According to the characteristics of heat distribution, the solution strategy of self-organizing cluster fusion is adopted, which improves the clustering efficiency and reduces the time cost. In the case of multiple hotspots, the analysis speed is fast;

2) Using an improved double clustering algorithm that considers singular points, under the same accuracy requirements, the number of thermal sensors used is greatly reduced compared with other existing solutions, and the cost is lower.

Description of drawings

Fig. 1 is a flowchart of steps of the present invention.

Detailed ways

The embodiments of the present invention will be described in detail below in conjunction with the accompanying drawings. This embodiment is implemented on the premise of the technical solution of the present invention, and detailed implementation methods and specific operating procedures are provided, but the scope of protection of the present invention is not limited to the following Described embodiment.

Example 1

As shown in Figure 1, a microprocessor thermal sensor configuration method, the method includes the following steps:

The first step is to adopt the Alpha EV6 microprocessor architecture, run multiple working conditions on the microprocessor, use the SimpleScalar software to simulate the SPEC2000 (system performance evaluation test) standard test program to simulate different working conditions, and calculate the module through the power consumption calculation unit power consumption, the temperature distribution calculation unit calculates the temperature distribution of the microprocessor under the power consumption;

The above power consumption calculation unit is Wattch infrastructure software, and the temperature distribution calculation unit is Hotspot software.

In the second step, the temperature distribution diagrams obtained by simulation under different test programs are merged and eliminated through MATLAB to obtain a superimposed map of the hotspot distribution of each module of the microprocessor, with a total of 20 hotspots;

The hotspot merging and elimination refers to placing all the hotspot data in the hotspot distribution diagram on the same architecture diagram, and if there are hotspots that completely overlap, then only keep one and remove other overlapping points.

The third step is to set the maximum hotspot monitoring error as 5%, and use the improved dual clustering algorithm considering singular points to perform self-organizing cluster fusion clustering operation of all hotspots. The algorithm framework is as follows:

1) Initialize a class center, which is the point with the largest difference from the mean value among all the data points to be clustered, scan the adjacent data points based on this center, and calculate whether the attribute distance between these data points and the center is Exceeding the threshold, if the attribute distance is less than the threshold, it will be merged into one cluster, if the attribute distance is greater than the threshold, it will not be merged;

2) When merging occurs, immediately take the newly merged data point set as a new class, and scan all the surrounding adjacent data points again until no merging occurs;

3) If a cluster can no longer be merged with the surrounding, then this cluster is a complete cluster we are looking for;

4) Repeat steps 1, 2, and 3 in the remaining data points until all data points belong to a certain cluster.

The advantage of considering singular points is that in the process of clustering, hotspots whose temperature values differ greatly from the average temperature value are often difficult to be clustered with most hotspots, and are isolated separately, resulting in the need to assign them separately In the case of a sensor, it is preferred to use it as the cluster center to assign as many surrounding points as possible into the cluster to reduce the number of sensors.

The above-mentioned adjacent data points refer to: superimpose data according to the distribution of hotspots, and make VORONOI diagrams of these hotspot lattices. If the VORONOI polygons where two data points are located meet the Queen criterion, that is, there are common vertices in the two polygons, then the two The data points are judged as adjacent data points.

The attribute distance refers to: the Euclidean distance of non-spatial attributes between two data points. Assume N d-dimensional data points P _n = {g _n ¹ , ... g _n ^d , a _n ¹ , ... a _n ^m }, 1≤n≤N; where g _n ¹ ,...g _n ^d are the spatial position coordinates of the point, a _n ¹ ,...a _n ^m are the non-spatial attribute values of the point. Then there are any two points P ₁ , the attribute distance of P ₂ is:

$\mathrm{<span\; class="notranslate">\; D.</span>}<span\; class="notranslate">\; =</span>\sqrt{\underset{\mathrm{<span\; class="notranslate">\; m</span>}}{\mathrm{<span\; class="notranslate">\; \&Sigma;</span>}}{<span\; class="notranslate">\; (</span>{{\mathrm{<span\; class="notranslate">\; a</span>}}_{\mathrm{<span\; class="notranslate">\; 1</span>}}}^{\mathrm{<span\; class="notranslate">\; i</span>}}<span\; class="notranslate">\; -</span>{{\mathrm{<span\; class="notranslate">\; a</span>}}_{\mathrm{<span\; class="notranslate">\; 2</span>}}}^{\mathrm{<span\; class="notranslate">\; i</span>}}<span\; class="notranslate">\; )</span>}^{\mathrm{<span\; class="notranslate">\; 2</span>}}}$

In this embodiment, since the only non-spatial attribute value is temperature, the attribute distance is the temperature difference between two points.

The threshold refers to: a variable that judges whether two data points can be clustered into one class, and is the key to ensure that the hotspot monitoring error does not exceed the set maximum value when the sensors are allocated according to the clustering results. value-related. Suppose ε _max is the maximum error ratio allowed by the project, n is the number of data points contained in the cluster, and a _i is the non-spatial attribute value of all data points in the cluster. Then there are:

Here α is a correction factor, which is used to balance the number of sensors and errors. It should be flexibly selected according to different maximum error setting values. In this embodiment, the maximum error setting value is 5%, and the α value is selected as 2.0.

In the fourth step, the number of clusters finally obtained through the above steps is 4, and each cluster is assigned a thermal sensor, that is, all hot spots are monitored according to a maximum error of 5%, and the required number of thermal sensors is 4. The position of the sensor is placed at the centroid of all hotspots contained in the cluster, that is, at the center of the weighted average considering the temperature value.

Example 2

Referring to Fig. 1, the specific steps of the method are the same as those in Example 1, the difference is that the maximum monitoring error is set to 2%, 3% and 4% respectively, and obtained by using the method of the present invention and the existing k-means clustering technology The number of sensors used in the thermal monitoring scheme is compared, and the results are shown in Table 1, which proves that the method of the present invention greatly reduces the number of sensors under the same accuracy requirements. Especially when the precision requirement is high, the efficiency improvement of the method of the present invention is very obvious. (The 2% accuracy requirement is already close to the error limit, so the improvement is not obvious.)

Table 1

Claims (5)

1. A microprocessor thermal sensor configuration method is characterized in that the method may further comprise the steps: 1) Running multiple working conditions on the microprocessor, the power consumption calculation unit calculates the power consumption of the microprocessor under different working conditions, and the temperature distribution calculation unit calculates the temperature distribution of the microprocessor under the power consumption; 2) Perform data processing on the temperature distribution diagrams obtained under different working conditions, and obtain superimposed diagrams of hot spot distribution of each module of the microprocessor; 3) Set the upper limit value of the hotspot monitoring error, and according to the upper limit value, consider the optimized double clustering of the data points in the hotspot distribution overlay graph to obtain the clustering results of these hotspots; 4) According to the above clustering results, each cluster is configured with a thermal sensor, and the thermal sensor is set at the centroid of all hot spots contained in the cluster, that is, at the average center considering the weighted temperature value, to ensure that it is within the set maximum error Monitor all hotspot temperature values; The optimized dual clustering considering the singular point in the step 3) refers to: when the dual clustering algorithm initializes the cluster center, selects the data point whose temperature value differs the most from the average temperature value, specifically: 301) Initialize a class center, which is the point with the largest difference from the mean value among all the data points to be clustered, scan the adjacent data points based on this center, and calculate whether the attribute distance between these data points and the center is Exceeding the threshold, if the attribute distance is less than the threshold, it will be merged into one cluster, if the attribute distance is greater than the threshold, it will not be merged; 302) When merging occurs, use the newly merged data point set as a new class, and scan all adjacent data points around again until no merging occurs; 303) If a cluster can no longer be merged with surroundings, then this cluster is a complete cluster; 304) Re-execute steps 301, 302, and 303 in the remaining data points until all data points belong to a certain cluster; The formula for calculating the threshold is: ${\mathrm{<span\; class="notranslate">\; D.</span>}}_{\mathrm{<span\; class="notranslate">\; max</span>}}<span\; class="notranslate">\; =</span>{\mathrm{<span\; class="notranslate">\; \&alpha;</span>}<span\; class="notranslate">\; \&times;</span>\mathrm{<span\; class="notranslate">\; \&epsiv;</span>}}_{\mathrm{<span\; class="notranslate">\; max</span>}}<span\; class="notranslate">\; \&times;</span>\frac{\mathrm{<span\; class="notranslate">\; 1</span>}}{\mathrm{<span\; class="notranslate">\; no</span>}}\underset{\mathrm{<span\; class="notranslate">\; no</span>}}{\mathrm{<span\; class="notranslate">\; \&Sigma;</span>}}{\mathrm{<span\; class="notranslate">\; a</span>}}_{\mathrm{<span\; class="notranslate">\; i</span>}}$ In the formula, α is the correction factor, ε _max is the upper limit of the hotspot monitoring error, n is the number of data points contained in the cluster, a _i is the non-spatial attribute value of all data points in the cluster, that is, the temperature of each data point . 2. a kind of microprocessor thermal sensor configuration method according to claim 1, is characterized in that, the temperature distribution of the microprocessor in described step 1) refers to the two-dimensional temperature matrix when microprocessor working state . 3. a kind of microprocessor thermal sensor configuration method according to claim 1, it is characterized in that, described step 2) in the hotspot distribution overlay map refers to: according to the division of microprocessor architecture modules, pick out each module hotspots, and then superimpose the distribution of all hotspots under different working conditions on an architecture map, and remove the hotspot distribution map obtained after repeated hotspots. 4. A kind of microprocessor thermal sensor configuration method according to claim 1, is characterized in that, the hot spot monitoring error upper limit value in described step 3) refers to: in microprocessor thermal management design, allowable The maximum difference between the sampled value of the thermal sensor and the actual temperature of the monitored hot spot. 5. A kind of microprocessor thermal sensor configuration method according to claim 1, is characterized in that, the double clustering in described step 3) refers to the clustering algorithm that carries out on space domain and attribute domain simultaneously.