The invention relates to a system and methods for monitoring a set of metrics. More particularly, the invention provides a system and methods for dynamically computing thresholds, and for signaling threshold violations.
Transactions are at the heart of web-based enterprises. Without fast, efficient transactions, orders dwindle and profits diminish. Today's web-based enterprise technology, for example, is providing businesses of all types with the ability to redefine transactions. There is a need, though, to optimize transaction performance and this requires the monitoring, careful analysis and management of transactions and other system performance metrics that may affect web-based enterprises.
Due to the complexity of modern web-based enterprise systems, it may be necessary to monitor thousands of performance metrics, ranging from relatively high-level metrics, such as transaction response time, throughput and availability, to low-level metrics, such as the amount of physical memory in use on each computer on a network, the amount of disk space available, or the number of threads executing on each processor on each computer. Metrics relating to the operation of database systems and application servers, operating systems, physical hardware, network performance, etc. all must be monitored, across networks that may include many computers, each executing numerous processes, so that problems can be detected and corrected when (or preferably before) they arise.
Due to the number of metrics involved, it is useful to be able to call attention to only those metrics that indicate that there may be abnormalities in system operation, so that an operator of the system does not become overwhelmed with the amount of information that is presented. To achieve this, it is generally necessary determine which metrics are outside of the bounds of their normal behavior. This is typically done by checking the values of the metrics against threshold values. If the metric is within the range defined by the threshold values, then the metric is behaving normally. If, however, the metric is outside the range of values defined by the thresholds, an alarm is typically raised, and the metric may be brought to the attention of an operator.
Many monitoring systems allow an operator to set the thresholds beyond which an alarm should be triggered for each metric. In complex systems that monitor thousands of metrics, this may not be practical, since setting such thresholds may be labor intensive and error prone. Additionally, such user-specified fixed thresholds are inappropriate for many metrics. For example, it may be difficult to find a useful fixed threshold for metrics from systems with time varying loads. If a threshold is set too high, significant events may fail to trigger an alarm. If a threshold is set too low, many false alarms may be generated.
In an attempt to mitigate such problems, some systems provide a form of dynamically-computed thresholds using simple statistical techniques, such as standard statistical process control (SPC) techniques. Such SPC techniques typically assume that metric values fit a Gaussian, or “normal” distribution. Unfortunately, many metrics do not fit such a distribution, making the thresholds that are set using typical SPC techniques inappropriate for certain systems.
For example, the values of many performance metrics fit (approximately) a Gamma distribution. Since a Gamma distribution is asymmetric, typical SPC techniques, which rely on a Gaussian or normal distribution, which is symmetric, are unable to set optimal thresholds. Such SPC thresholds are symmetric about the mean, and when applied to metric data that fits an asymmetric distribution, if the lower threshold is set correctly, the upper limit will generally be set too low. If the upper limit is set correctly, then the lower limit will generally be set too low.
Additionally, typical SPC techniques are based on the standard deviation of a Gaussian or normal distribution. There are many performance metrics that exhibit self-similar or fractal statistics. For such metrics, standard deviation is not a useful statistic, and typical SPC techniques will generally fail to produce optimal thresholds.
Many performance metrics exhibit periodic patterns, varying significantly according to time-of-day, day-of-week, or other (possibly longer) activity cycles. Thus, for example, a metric may have one range of typical values during part of the day, and a substantially different set of typical values during another part of the day. Current dynamic threshold systems typically fail to address this issue.
Additionally, current dynamic threshold systems typically ignore data during alarm conditions for the purpose of threshold adjustment. Such systems are generally unable to distinguish between a short alarm burst and a persistent shift in the underlying data. Because of this, such systems may have difficulty adjusting their threshold values to account for persistent shifts in the values of a metric. This may cause numerous false alarms to be generated until the thresholds are reset (possibly requiring operator intervention) to take the shift in the underlying data into account.
In view of the foregoing, there is a need for a system and methods for dynamically generating alarm thresholds for performance metrics, wherein the metrics may not fit a Gaussian or normal distribution, or may exhibit cyclic behavior or persistent shifts in the values of the metrics. The present invention uses a variety of statistical methods, including statistical process control (SPC) methods, normalization methods, and heuristics to generate such thresholds.
In general, in one aspect, the system establishes one or more default alarm thresholds associated with a metric, repeatedly receives data associated with the metric, statistically analyzes the received data to establish one or more updated alarm thresholds, and triggers an alarm on receipt of data that violates one or more updated thresholds. By basing the updated alarm thresholds on a statistical analysis of the metric data, the system is able to update one or more thresholds dynamically, based on the values of the metric.
In one embodiment the statistical analysis determines whether the received data fits a normal distribution (i.e., the data is normal). This may be done in embodiments of the invention by applying a chi-square test to the received data, by applying an Anderson-Darling test to the received data, or by applying both these tests and combining the results. If the data is determined to fit a normal distribution, it is categorized as “normal.” One embodiment uses SPC techniques to compute the thresholds for data that is categorized as normal.
In one embodiment of the invention, if the data is not normal, the system determines whether the data is normalizable. One embodiment makes this determination by operating on the received data with a function representing the estimated cumulative distribution of the received data, and then using the quantile function of a normal distribution to attempt to normalize the data. If this is successful, the data is categorized as “normalizable.” If the data is normalizable, one embodiment normalizes the data, and then uses SPC techniques to calculate one or more thresholds. When these thresholds are later applied, it may be necessary to first normalize the data.
In one embodiment, if the data is not normal, and is not normalizable, then the data is categorized as “non-normal,” and heuristic techniques may be used to determine one or more thresholds. As part of these heuristic techniques, embodiments of the invention may use combinations of statistical techniques, including weighted linear regression techniques, and techniques based on a quantile function.
By categorizing the data as normal, normalizable, or non-normal, and applying different techniques to compute one or more thresholds, the dynamic threshold calculation of the present invention is able to compute one or more thresholds for data that does not fit a Gaussian or normal distribution.
In one embodiment, the statistical analysis is repeated, and one or more updated alarm thresholds are updated based on previous values of the alarm thresholds. This permits the system to handle metrics that are cyclic in nature, or that have persistent shifts in the values of their data. In one embodiment, this may be achieved by applying a filter in the computation of one or more updated thresholds. In one embodiment this filter uses a weighted sum of data that may include historical data, a statistical summarization of metric data, metric data associated with a predetermined time period, or any combination thereof. In one embodiment, after the filter is applied, one or more thresholds may be computed using SPC techniques or heuristic techniques, depending on the category of the data, as discussed above.
Embodiments of the invention may test data against one or more thresholds to trigger alarms using a variety of methods. For some metrics, use of fixed alarms may be appropriate, and for these metrics the received data is compared with one or more fixed thresholds to determine whether an alarm should be triggered. In some embodiments, if the metric was determined to be normal, then the mean and standard deviation of the received data may be checked against one or more alarm thresholds to determine if an alarm should be triggered.
In some embodiments, when the metric was determined to be normalizable, then the received data is normalized. The mean and standard deviation of the normalized data are then compared against one or more thresholds to determine whether an alarm should be triggered.
In some embodiments, when the metric was determined to be non-normal, then the mean of the received data is compared to one ore more thresholds determined by heuristic techniques. If the mean falls outside of the range of values defined by the threshold(s), then an alarm is triggered.
In some embodiments, the methods of dynamically computing and applying one or more thresholds can be implemented in software. This software may be made available to developers and end users online and through download vehicles. It may also be embodied in an article of manufacture that includes a program storage medium such as a computer disk or diskette, a CD, DVD, or computer memory device. The methods may also be carried out by an apparatus that may include a general-purpose computing device, or other electronic components.
Other aspects, embodiments, and advantages of the present invention will become apparent from the following detailed description which, taken in conjunction with the accompanying drawings, illustrating the principles of the invention by way of example only.
The foregoing and other objects, features, and advantages of the present invention, as well as the invention itself, will be more fully understood from the following description of various embodiments, when read together with the accompanying drawings, in which:
In the drawings, like reference characters generally refer to the same parts throughout the different views. The drawings are not necessarily to scale, emphasis instead being placed on illustrating the principles of the invention.
As shown in the drawings for the purposes of illustration, the invention may be embodied in a system that collects, analyzes, and reports performance metrics for systems such as, for example, complex transaction-based structures typified by (but not limited to) e-commerce systems. A system according to the invention provides the capability to discern, group, and highlight performance information that facilitates the efficient operation and control of the monitored system. A system manager presented with information so organized is relieved from the difficulties associated with visualizing and interpreting what appears to be a large amount of unrelated data.
In brief overview, embodiments of the present invention provide a system and methods for collecting, analyzing and reporting on significant irregularities with respect to a set of system performance metrics. These metrics are collected from the various sub-systems that make up, for example, an e-commerce transaction processing system. Typical metrics include measures of CPU and memory utilization, disk transfer rates, network performance, process queue depths and application module throughput. Key performance indicators at the business level, such as transaction rates and round-trip response times are also monitored. Statistical analysis is used to detect irregularities in individual metrics. Correlation methods are used to discover the relationship between metrics. The status of the system is presented via a graphical user interface that highlights significant events and provides drill-down and other visual tools to aid in system diagnosis.
The system and methods of the present invention are described herein as applying to software for use by a system manager, such as an web-based enterprise system manager, to assist, for example, in the achievement and maintenance of Service Level Agreements in terms of system performance. It will be understood that the system and methods of the present invention are not limited to this application, and can be applied to the control and maintenance of most any system whose operation can be described through use of a set of system metrics.
Referring to
Metric collection module 102 includes one or more data adapters 108, installed in the systems to be monitored. Each data adapter 108 collects information relating to the metrics that are being monitored from a particular sub-system, such as an operating system, web server, or database server. The data adapters 108 transmit their information to a dynamic sampling agent 110, which collects the metrics, performs fixed and statistical threshold checks, and sends both the metric data and threshold alarm events to metric analysis module 104.
Metric Analysis module 104 performs progressive information refinement on the metrics that are being monitored. Analysis module 104 includes dynamic threshold testing component 114, and metric correlation component 116, as well as optional components, such as event correlation component 118, root cause analysis component 120, and action management component 122. In one embodiment, dynamic threshold testing component 114 is part of dynamic sampling agent 110, for reasons of scalability and efficiency, while the other components of metric analysis module 104 act as plug-in components of a central service management platform.
Dynamic threshold testing component 114 detects when individual metrics are in abnormal condition, producing threshold alarm events. It uses both fixed, user-established thresholds and thresholds derived from a statistical analysis of the metric itself. Dynamic threshold testing component 114 includes a fixed threshold check module, a dynamic threshold check module, and a dynamic threshold computation module, as will be discussed in detail below in the section entitled “Adaptive Threshold Determination.”
Metric correlation component 116 analyzes pairs of metric values collected from one or more dynamic sampling agent 110. It applies various correlation and regression techniques to determine the statistical relationship between pairs of metrics, and to form groups of closely related metrics. It also tries to determine temporal relationship between metrics. Metric correlation component 116 will be described in detail below, in the section entitled “Adaptive Metric Grouping.”
Event correlation component 118 receives threshold alarms events from the synamic sampling agent(s) 110. It uses a set of rules to convert groups of related events into a single Significant Event. The rules include use of techniques such as event counting, temporal analysis, pattern recognition, and event promotion.
Root cause analysis component 120 applies threshold alarm events and the results of metric correlation component 116 to build and maintain a Bayesian belief network that is used to determine the most likely root cause of Significant Events. When a set of likely root causes is determined, this component generates a root cause event.
Action management component 122 uses rule-based reasoning to take action on root cause events. Such action may include suppressing the event, notifying the system manager, sending e-mail, or executing a script.
Metric reporting module 106 provides a user, such as a system manager, with detailed reporting on the metrics, alarms, and analysis performed by the system. In one embodiment, reporting module 106 uses a 3D graphical user interface that provides for drill-down exploration of metric data and analysis, permitting, for example, exploration of the activity that caused an abnormal or suspicious condition in the system being monitored. Additionally, graphical tools may be used in reporting module 106 to display a time history of sets of related metrics. Other interfaces and reports, such as 2D graphical interfaces or printed activity reports and charts may also be provided by reporting module 106. In one embodiment, reporting module 106 permits an operator to monitor the system remotely, over a web connection.
Adaptive Threshold Determination
Referring now to
Data manager 200 also provides small sets (or subgroups) of consecutive samples to dynamic threshold check module 208, which compares the statistics of each such subgroup against previously computed dynamic thresholds, which are stored in distribution information store 210 and dynamic limit store 212. The individual samples are also accumulated and temporarily stored in accumulated samples store 214.
Periodically, data manager 200 signals dynamic threshold computation module 216 to compute new thresholds. As will be described in greater detail below, dynamic threshold computation module 216 first classifies the data based on the statistics of the accumulated samples into one of three types: normal, normalizable or non-normal. For normal data, provisional thresholds are computed using standard statistical process control techniques. For normalizable data, the probability distribution of the metric is estimated from the accumulated samples. The estimated distribution is stored in distribution information store 210, and is used to transform the samples into normally distributed values. Standard statistical process control (SPC) techniques are applied to the transformed samples to compute provisional thresholds. For non-normal data, the thresholds are computed using a heuristic technique, which combines linear regression with the quantile function of the stored samples.
These provisional thresholds are modified by a baseline filtering process. This process records the history of the threshold statistics and uses it to filter new threshold estimates. In this way, the process can adapt to cyclic patterns of activity, producing thresholds that cause far fewer false alarms, and yet remain sensitive to unexpected shifts in the metric statistics. The new threshold levels are stored in dynamic limit store 212, for use in subsequent sample checks.
It will be understood by one skilled in the relevant arts that the various components and modules of dynamic sampling agent 110 may be implemented as programmed code that causes a general-purpose computing device, such as a digital computer, to perform the functions described, effectively transforming the general-purpose computing device into each of the components or modules as the programmed code is executed. Alternatively, the components and modules of dynamic sampling agent 110 may be embodied in a combination of electronic components, such as logic gates, transistors, registers, memory devices, programmable logic devices, or other combinations of devices that perform the described functions.
At a predetermined interval (for example, hourly) samples for a metric from accumulated samples store 214 are sent to dynamic threshold computation module 216. These samples are typically divided into a number of subgroups, in which each subgroup typically contains a predetermined number of samples. For example, a subgroup may consist of ten values of a metric, sampled at a rate of one value per second, over ten seconds. In this example there would be 360 subgroups per hour collected for each metric. In this example, every half-hour, a list containing the values for a metric over the last hour (typically 3600 values) may be sent to dynamic threshold computation module 216.
At step 300 a normal test is applied to the samples, to determine whether the samples of metric data provided to dynamic threshold computation module 216 fit a normal distribution, and the metric may be classified as “normal.” To make this determination, the normal test that is executed in step 300 uses a combination of the chi-square test and the Anderson-Darling test on a specially formed histogram of the metric data. The chi-square (x2) test is a well-known distribution test that is utilized here to compare accumulated samples to the probability density function (PDF) of normal data. The Anderson-Darling test is a more sensitive test that is used to compare the cumulative distribution function (CDF) of accumulated data to the CDF of normal data. If either of these tests gives a strong indication of non-normal distribution, the normal test 300 fails. If both give a weak indication of a non-normal distribution, the normal test 300 also fails, otherwise the data is assumed to be normal.
If the data fit a normal distribution (step 302), then at step 304, the system sets a flag indicating that the data for the metric in question are normal, and stores the flag in distribution information store 210 for use by dynamic threshold check module 208.
Next, in step 306, mean and standard deviation values are computed for each subgroup as well as the estimated mean and standard deviation over all of the subgroups. The following formulas are used to compute these values:
In step 308 a baseline filter (baseline filter A) uses the recent history of the statistics of the metrics to filter the estimated mean and standard deviation over all of the subgroups. This computation uses, for example, a record of the parameters needed to compute thresholds for each hour of each day of the week for up to a user-specified number of weeks. As described below, it produces a decaying weighted sum of the values for a particular time slot for each of the weeks available. This acts to filter the values and give memory of recent history of the metrics, so that the thresholds are more stable, and more sensitive to longer-term changes. Using this type of filtering permits the thresholds to adapt to regular (e.g., weekly) patterns of activity.
Using this baseline filer, the value to be calculated, V', may be computed using the following formula:
Where:
Next, in step 310, upper and lower control limits are computed using the filtered values of the mean and standard deviation. These limits are referred to as SPC limits, because they are computed using standard statistical process control (SPC) techniques. To compute the SPC limits, the following formulas are used:
c4=E(s)/σ=√{square root over (2/(N−1))}*Γ(N/2)/Γ((N−1)/2) (Eq. 6)
var(s)=σ2/(1−c42) (Eq. 7)
A3=3.0/(c4*√{square root over (N)}) (Eq. 8)
B3=1.0−(3.0/c4)√{square root over (1−c42)} (Eq. 9)
B4=1.0+(3.0/c4)√{square root over (1−c42)} (Eq. 10)
LCL—X=μ−A3*σ (Eq. 11)
UCL—X=μ+A3*σ (Eq. 12)
LCL—S=B3*σ (Eq. 13)
UCL—S=B4*σ (Eq. 14)
Where:
Once the SPC limits are computed, they are stored in SPC limits store 350, which, in one embodiment, is a portion of dynamic limit store 212. The SPC limits may be used by dynamic threshold check module 208.
Referring again to
For example, for a gamma distribution, the system computes the mean and standard deviation, and then uses known statistical methods to estimate the parameters α, β, and γ, which determine the shape, scale, and origin of the gamma distribution, respectively. A chi-square test, similar to the test used in the normal distribution test, described below, but with histogram bin counts that are computed against the probability density function of the estimated distribution, is used to assess the fit between the data and the estimated distribution. The resulting chi-square value is checked against critical values, which can be found in a standard chi-square table to determine whether the fit is good.
If the fit is good (step 314), in step 316, the distribution is saved and a flag is set indicating that the data is normalizable. The estimated distribution parameters (such as the mean, standard deviation, shape, scale, and origin for a gamma distribution) and the flag are stored in distribution information store 210.
In step 318, the data is transformed into a normally distributed dataset. This is done by (1) passing data through the function representing its estimated cumulative distribution, and then (2) passing that result through the quantile function of a normal distribution.
For example, for the gamma distribution, the cumulative distribution function (CDF) may be derived using known statistical techniques, based the estimated shape, scale, and origin parameters from distribution information store 210. Once the CDF has been derived, the probability Pi can be found for each sample xi that represents the portion of values expected to be less than xi. Pi may be computed using:
Pi=CDF(xi) (Eq. 16)
Since the CDF represents the distribution of the xi, the values of Pi will be uniformly distributed in the range (0,1).
Next, the quantile function of the normal distribution QN will be computed using known statistical techniques from the estimated mean and standard deviation values that were stored in distribution information store 210. For each probability Pi, a new sample value Zi is computed using:
Zi=QN(Pi) (Eq. 17)
The non-normal data samples xi have been transformed into normal samples Zi by the path:
xi→Pi→Zi
Once the data have been normalized in step 318, the system proceeds to step 306, to compute subgroup statistics, apply a baseline filter, and compute SPC limits, as described above. As with data that fit the normal distribution, the computed limits are stored in SPC limits store 350.
In step 320, if a good fit was not found for the data, upper and lower threshold limits for the mean are calculated using a statistical method based on the sample or empirical quantile function of the means. This process of heuristic limit calculation will be described in greater detail below.
In step 322, a baseline filter (baseline filter B) is used, along with the recent history of the limits, to filter the limit values so the thresholds adapt to regular patterns of activity. Baseline filter B, used in step 322, is similar to baseline filter A, used in step 308, and may be expressed using the following formula:
Where:
The adjusted upper and lower threshold limits for the mean are stored in heuristic limits store 352, which, in one embodiment, is a portion of dynamic limits store 212. These threshold limits can then be accessed by dynamic threshold check module 208.
As an example of the baseline filter (baseline filter B) used in step 322 (a similar filter is also used in step 308), suppose that the system starts up on Wednesday Jul. 4, 2001, at 12:00:00. At 13:00:00 dynamic threshold computation module 216 calculates an upper threshold value, value0, on data for Metric X. This upper limit is based on metric data from 12:00:01–13:00:00. In the baseline filter step, valuec is not multiplied by a weight factor because there is no historical data. So, the upper threshold limit for Metric X at 13:00:00 is computed by:
(Note that when the system first starts up, there are no histories saved so wc=1.0.)
At the next hour, at 14:00:00, an upper threshold value, value 1.U, is calculated based on data from 13:00:01–14:00:00. Now there is historical data for the last hour that has weighted influence on current data. The threshold value for last hour, value h U, is equal to valuec. Thus, the value of the upper threshold limit for Metric X at 14:00:00 is:
The varied weights are based on baseline history, such that there is maximum use of available information. As more history is gathered, weekly data is weighted more heavily and dependence on current information is reduced. Optimal settings are application dependent.
For any time period from the second day following system start-up at onward, the upper threshold value from the previous day (value d U) can be used in the baseline filter, as the following equation shows:
The next time the upper threshold value for Wednesday at 14:00:00 is updated is during the next week, on Wednesday, Jul. 14, 2001, at 14:00:00. Since there is now more than a week of historical data, the weighted values from the previous day (Tuesday at 14:00:00) and the previous week (Wednesday, Jul. 4, 2001 at 14:00:00) can be added in. The equation would then be:
All of the weighted factors (wc, wh, wd, . . . ) must add up to 1.0. As time passes, and more history becomes available, the weighted factors will have values that look like those in the following chart. Weighted factors for different applications can be set at different values. After collecting data for N weeks the weighted factors would no longer change.
The following chart illustrates an example of possible weight values for up to 8 weeks
Where:
Referring now to
In step 504, the first step of performing the chi-square test, the mean [μ] and standard deviation [σ] of the data are calculated using known statistical methods. The mean and standard deviation are used to determine the shape of a theoretical “bell curve” of the histogram of the data, assuming that the distribution of the data is normal.
Referring again to
For example, when these rules are applied when the number of samples is 3600, the number of bins will be 30, since 10% of 3600 is 360, which is greater than 30. Similarly, if the number of samples is 90, the number of bins will be 9 (10% of the number of samples). If the number of samples is 40, there will be 6 bins (10% of 40 is less than 6, so 6 bins will be used). If the number of samples is 25, the chi-square test will return a result indicating that the data are not normal.
An expected number of samples for each bin is computed by dividing the number of samples by the number of bins:
Where:
The upper limits of the bins are computed using QN, the quantile function of the normal distribution, with mean μ and standard deviation σ. The quantile function is the inverse of the cumulative distribution function (CDF). Thus, given that i/k is the portion of samples that should be in or below bin i, the upper bin limit (UBL) for bin i is computed according the following formula:
CDF(UBLi)=i/k UBLi=QN(i/k) (Eq. 20)
Where:
For example, for 30 bins (i.e., k=30), the following upper bin limits will be computed:
Next, in step 508, the number of samples that belong in each bin are counted. This may be accomplished by sorting the samples into ascending order, and then segmenting the samples into the bin so that each bin gets all those samples that have values greater than the UBL of the previous bin, and less than or equal to the UBL for the current bin. The count, Ci of the samples in bin i is used to compute the chi-square value. In one embodiment, after the samples have been sorted in ascending order, the counts Ci for each bin can be computed using the method described in the following pseudocode:
Next, in step 510, the system computes the chi-square value to measure the “goodness of fit” between the histogram and the normal curve. The chi-square value is computed using the following formula:
Where:
In step 512, the system tests the chi-square value against two critical values, the lower critical value and upper critical value, taken from a standard chi-square table, to determine whether the data fit a normal distribution. If the chi-square value is below the lower critical value, then the data is probably normal (no indication of non-normality). If the chi-square value is between the lower critical value and the upper critical value, then the data is possibly normal (weak indication of non normality). If the chi-square value is above the upper critical value, then the data is not likely to be normal (strong indication of non-normality).
Next, at step 514, the system starts performing the Anderson-Darling test. The Anderson-Darling test is a “distance test” based on the probability integral transform, which uses the CDF of the hypothesized distribution to transform the data to a uniformly distributed variant. In step 514, the data is sorted into ascending order. This may have already been performed, during step 508, in which case the data need not be re-sorted.
In step 516, the system uses the mean and standard deviation of the data (that were already computed above) to shift and scale the data to transform the data into a standard normal variable, given that the data itself is normal. This transformation is performed using the following formula:
Where:
Next, in step 518, the system computes the corresponding probabilities (Pi) from the standard normal CDF (FN). The probabilities are computed according to the following formula:
Pi=FN(Wi) (Eq. 23)
In step 520, a statistic called A2 is computed, using the following formula:
Next, in step 522, the A2 value is compared to upper and lower critical values that can be found in a standard table of critical values for the Anderson-Darling test. If the A2 value is below the lower critical value, then the data is probably normal (no indication of non-normality). If the A2 value is between the lower critical value and the upper critical value, then the data is possibly normal (weak indication of non-normality). If the A2 value is above the upper critical value, then the data is not likely to be normal (strong indication of non-normality).
Finally, in step 530, the results of the chi-square test and the Anderson-Darling test are combined to reach a conclusion as to whether the data are normal. The results are combined according to the following table:
As can be seen in Table 1, if either the chi-square or the Anderson-Darling tests (or both) indicate that there is no indication of non-normality, and neither test indicates that there is a strong indication of non-normality, then the normal test will conclude that the data fit a normal distribution. Otherwise, the normal test will conclude that the data do not fit a normal distribution.
Referring now to
At step 700, the heuristic limit calculation computes a mean and standard deviation for each subgroup. This is done by applying the formulas:
μj=1/kΣxi (Eq. 25)
σj=√{square root over ([Σ(xi−μj)2]/(k−1))}{square root over ([Σ(xi−μj)2]/(k−1))} (Eq. 26)
Where:
Next, in step 702, the μj are sorted into ascending order, forming the quantile function of the means.
Referring again to
Where:
Referring again to
Where:
Next, in step 708, known statistical techniques are used to perform linear regression on the original time-ordered μj to estimate a slope (b) and a confidence interval (c). An illustration of linear regression on the subgroup means is shown in
Referring again to
kb=b/(b+R×c) (Eq. 29)
Where:
Finally, in step 712, the upper and lower threshold limits are computed, using the following formula:
UTL=max+ks*σ+kb*b*Δt (Eq. 30)
LTL=min−ks*σ+kb*b*Δt (Eq. 31)
Where:
Referring now to
Next, in step 1104, dynamic threshold check module 208 determines whether the distribution for the metric that is being checked is normal. This is typically done by looking up a flag that indicates whether the distribution for the metric is normal from distribution information store 210.
If the distribution for the metric is not normal, in step 1108, dynamic threshold module 208 determines whether the distribution for the metric is normalizable. This is typically done by retrieving a flag indicating whether the distribution for the metric is normalizable from distribution information store 210. Generally, the estimated distribution for a metric is normalizable if dynamic threshold computation module 216 determined that it fit one of several standard distribution (e.g., a gamma distribution, a beta distribution, a Weibull distribution, etc.).
If the estimated distribution for the metric is normalizable, then in step 1110 the dynamic threshold check module 208 will normalize the sample data in the subgroup. This is done using methods similar to those discussed above in step 318, by, for example, (1) passing data through the function representing its estimated cumulative distribution, and then (2) passing that result through the quantile function of a normal distribution.
In step 1112, the mean and standard deviation of the subgroup data are computed.
Step 1114 performs an SPC limit test on the data to determine whether there is a threshold violation. The SPC limit test is performed by obtaining the SPC limits from SPC limits store 350. The mean and standard deviation are compared to the upper and lower limits for the mean and standard deviation for each subgroup. If the mean or standard deviation of any of the subgroups falls outside of the limits, then notification of a threshold violation is sent to alarm manager 206.
Referring again to
Overall, use of an adaptive threshold process, as described hereinabove, permits computation of a baseline of the statistics of the data of a metric to uncover daily, weekly, monthly, yearly, or any combination thereof, cyclic patterns. A beneficial, technical effect of this feature is that the patterns in the metrics are revealed or emphasized when they might have been obscured by, for example, the volume of data collected. A large number of metrics and increased system complexity also contribute to the problem of detecting the patterns—a problem that this feature solves. These patterns are used to filter the statistics used to compute thresholds such that they can use history to predict the next threshold setting, thereby reducing false alarms. In addition short-term future expected behavior is predicted, and used to adjust the thresholds, further reducing false alarms. As shown in an example of
Adaptive Metric Grouping
Referring now to
Generally, metric sample values are monitored by dynamic sampling agent 110 or a network of computers monitored by dynamic sampling agents 110 within a system. As discussed in detail hereinabove, dynamic sampling agent(s) 110 provide threshold alarms based on dynamic, statistical threshold generation. Generally, dynamic sampling agent(s) 110 monitor all metrics for threshold violations. If a metric causes a threshold violation, alarm manager 206 within dynamic sampling agent(s) 110 sends out an alarm, indicating an out-of-tolerance metric. Additionally, dynamic sampling agent(s) 110 provides metric data to recent metric data collector 1406, where the information is stored for a predetermined time, for example, fifteen minutes.
Threshold alarm events (i.e., notification of out-of-tolerance metrics) are received by alarm event controller 1404. Alarm event controller 1404 collects: 1) the metric name that caused the threshold alarm and, 2) the time the threshold alarm occurred. Alarm event controller 1404 continuously collects alarms generated by dynamic sampling agent(s) 110. Action is taken when there are a particular number of alarm groups within a predetermined time period, (i.e., when a predetermined frequency of alarms is reached). This alarm group frequency parameter is configurable. For example, the alarm group parameter can be configured to cause alarm event controller 1404 to trigger further action if it receives ten alarms within thirty seconds. When the alarm frequency specified in the alarm frequency parameter is reached, alarm event controller 1404 is activated, and sends a list of these out-of-tolerance metrics to recent metric data collector 1406. The set of out-of-tolerance metrics in the list that is sent to recent metric data collector 1406 is referred to as an “alarm group.”
Alarm event controller 1404 only counts the threshold alarms at the time the threshold is crossed, and does not count an alarm repeatedly if the alarm remains active over a period of time. An alarm will generally only be counted twice if it is triggered, reset, and then triggered again. In cases where there are alarms that are repeatedly triggered in this manner, one embodiment of alarm event controller 1404 will count all such triggers, but report only the most recent one to data collector 1406.
For example, suppose that twenty-six metrics are monitored, having ID's A–Z. Alarm event controller 1404 receives occasional alarms on these metrics. Suppose that alarm event controller 1404 receives the following alarms:
At time t=09:45:30 alarm event controller 1404 is activated because the #alarms/Δt (in this example, ten alarms within thirty seconds) parameter is reached, since there have been a total of twelve alarms within the timeframe 09:45:21–09:45:30.
As noted above, alarm event controller 1404 only counts the threshold alarms at the time the threshold is crossed, not each second of threshold violation. For example, the alarm for metric D could be active for five minutes from when it was tripped at 09:45:21, but the alarm is recorded by the Alarm Event Controller only once at the time it was activated at 09:45:21. The metric alarm is counted again only if it is reset and tripped again, such as metrics C and P. Metric C was tripped at 09:45:21 and reset some time within the next seven seconds and tripped again at 09:45:29. Metric P was tripped at 09:45:26 and then reset some time within the next three seconds and tripped again at 09:45:30.
In the example, the list of metrics that activated the alarm event controller is:
C D F M AP R W H GC P
Since (the underlined) metrics C and P appear twice, alarm event controller 1404 dismisses the first metric C and metric P in the Metrics List and retains the most recent ones. Thus, alarm event controller 1404 sends the following metric name list (alarm group) to recent metric data collector 1406:
D F M A R W H G C P
Recent metric data collector 1406 collects historic metric data from metric data collector 1402 for metrics that are in the alarm group provided by alarm event controller 1404. This historic metric data is collected over a predetermined period of time (e.g., ten minutes), and is sent to correlation module 1408. The historic data for the metrics is synchronized, so that collection of the data starts from the activation time of alarm event controller 1404 (e.g., 09:45:30 in the example) and goes back in time for a predetermined period at predetermined intervals. For example, if the system was configured to look at historical data for the previous ten minutes, at one second intervals, there would be 600 samples for each alarmed metric.
Sometimes there are gaps in the data collected by recent metric data collector 1406 where metric samples were not recorded. In one embodiment, recent data collector 1406 includes an align and filter process (not shown) that aligns data in the correct timeslots and filters out entire timeslots containing incomplete data. An example of such a process, using the metrics of the previous example, is shown in
In an alternative embodiment, instead of deleting the entire dataset, “pairwise deletion” may be used. Using pairwise deletion, only data for certain pairs of metrics within a given time slot are deleted. Typically, such pairwise deletion occurs during the correlation process (i.e., within correlation module 1408), when data for one of the metrics to be correlated is missing. Data for the other metric in the pair is deleted, and the correlation for that pair uses only samples from the time slots that have data for both metrics. While more data is retained using this technique, because the correlation is performed on incomplete datasets, a higher correlation coefficient is required to signify statistically significant correlation.
Referring again to
The equivalent data for metrics M, A, R, W, H, G, C, and P is also received by correlation module 1408.
Next, correlation module 1408 creates a list of all metric pairs in this alarm group. The number of pairs follows the formula:
Where N is the number of metrics in the alarm group.
In our example, N=10, so applying this formula yields (10*9)/2=45 pairs of metrics to be correlated. If all metric pairs in the example system had to be correlated, instead of having 45 pairs, there would be (26*25)/2=325 pairs to be correlated. By correlating only the metrics in an alarm group, a system according to the invention makes metric grouping possible for systems with thousands of metrics.
Next, correlation module 1408 correlates the metric values for each pair. For example, for the pair D and F, the following values would be correlated (note that the datasets at t=09:45:30 and t=09:45:28 are missing, because they are incomplete):
In one embodiment of the invention, instead of using conventional linear correlation, correlation module 1408 uses nonparametric, or rank correlation to correlate pairs of metrics in an alarm group. Generally, in rank correlation, the value of a particular data element is replaced by the value of its rank among other data elements in the sample (i.e., 1, 2, 3, . . . , N). The resulting lists of numbers will be drawn from a uniform distribution of the integers between 1 and N (assuming N samples). Advantageously in rank correlation, outliers have little effect on the correlation test and many non-linear relationships can be detected. In one embodiment, Spearman rank-order correlation is used to correlate the metric values for each pair of metrics.
A pair is considered correlated by correlation module 1408 if a significant value for the correlation coefficient is reached. Significance is determined using known statistical techniques, such as the “Students' t test” technique, with t computed as follows:
Where
For a number of samples (N) of 600, a correlation coefficient (rs) of 0.08, t would equal 1.96, which is a 95% confidence indicator. So, 0.1 is a reasonable minimum threshold for the correlation coefficient in our example.
Sometimes when one metric affects another metric, the effect is delayed, causing the pair of metrics not to correlate due to the delay. To detect this, the first list in the metric pair is shifted one second in relation to the second list and re-correlated to see if the result is better. In one embodiment, this process is repeated 5 times for +Δt and 5 times for −Δt. This permits delayed correlations to be detected within a predetermined period. For example, the following shows metric D values shifted by a Δt of +2 relative to the values of metric F:
VF-2, VD-599, and VD-600 are not used in this correlation because the time shift leaves them unpaired.
The benefit of this time shifting can be seen in
Referring again to
Correlated pair graph module 1410 handles all correlated metrics as nodes in a graph with the edges, or links in the graph indicating a correlation link between a pair of nodes. Whenever correlated pairs of metrics enter correlated pair graph module 1410 from correlation module 1408, the graph is updated. Correlation connecting links are added, the correlation coefficients on existing links are updated, or links are deleted if not reinforced. For example, if a particular pair of correlated metrics has not been entered for a preset time, such as a month, the correlation link between the pair in the correlation pair graph may be deleted.
For the example in
As noted above, the correlation pair graph may be updated as further correlation data for the metric pairs is added. For example, the correlation coefficient for the A-M link in the graph of
The correlation pair graphs in accordance with the invention are dynamic, changing as additional correlation data is added, updated, or deleted. For example, if correlation data indicating a correlation between a pair is entered, or there is other evidence of correlation, the correlation coefficient for that pair may be updated or increased to represent a strengthened correlation. This can be done, for example, by computing a weighted sum of the old correlation coefficient and the new correlation coefficient. Similarly, if no correlations for a pair are entered over a predetermined time, or there is a lack of supporting evidence for a pair being correlated, or both, the correlation coefficient for the pair may be decreased. This may be achieved by causing the correlation coefficient for a pair to decay over time, for example, by periodically multiplying the correlation coefficient by a decay factor that is slightly less than one (e.g., 0.95, or other values, depending on the desired rate of decay), when there is a lack of supporting evidence for correlation. Correlation coefficients that fall below a predetermined threshold may be deleted.
As shown in
As shown in
As the system continues to run, correlation links between metrics of disjointed groups can be added as seen in the example of
Referring back to
If a significant event occurs, an operator may want to identify all metrics that may have an effect on a key metric so that the root cause of the event can be discerned. By using a UI the operator can identify these associated metrics by query to on-demand group generation module 1412, which, in turn accesses the correlation pair graph through correlation pair graph module 1410. This query can either be operator initiated or reports can be automatically generated. Additionally, other components of the system or analysis plug-ins, such as a root-cause analysis plug-in may submit queries through on-demand group generation module 1412.
To determine which metrics are associated with a key metric, on-demand group generation module 1412 applies two conditions to the metrics. A metric is considered “associated” with the key metric if either of the two conditions are met.
The first condition that a metric can meet to be considered “associated” with a key metric is that the metric must be correlated to the key metric, and to at least P % of the other metrics that are correlated to the key metric, where P is a predetermined threshold. Typically, P will be relatively low, for example 25 or 30 (though many other values in the range of 0 to 100 are possible), since the associated metric must also be directly correlated with the key metric.
For example, referring to the correlation pair graph shown in
The second condition that a metric can meet to be considered associated with the key metric is that the metric must be correlated with at least X % of the metrics that are correlated with the key metric, where X is a predetermined value. Typically X will be relatively high, such as 80 or 90 (though other values in the range of 0 to 100 are possible), since the metric meeting this second condition need not be correlated to the key metric.
Because
This application claims priority to and the benefit of, and incorporates herein by reference, in their entirety, the following provisional U.S. patent applications: Ser. No. 60/307,055, filed Jul. 20, 2001, andSer. No. 60/322,021, filed Sep. 13, 2001. Further, this application incorporates herein by reference, in its entirety, U.S. provisional application Ser. No. 60/307,730, filed Jul. 3, 2001.
Number | Name | Date | Kind |
---|---|---|---|
4821220 | Duisberg | Apr 1989 | A |
5062055 | Chinnaswamy et al. | Oct 1991 | A |
5101348 | Arrowood et al. | Mar 1992 | A |
5161116 | Schneider et al. | Nov 1992 | A |
5307499 | Yin | Apr 1994 | A |
5367473 | Chu et al. | Nov 1994 | A |
5375070 | Hershey et al. | Dec 1994 | A |
5432932 | Chen et al. | Jul 1995 | A |
5450406 | Esaki et al. | Sep 1995 | A |
5459837 | Caccavale | Oct 1995 | A |
5459868 | Fong | Oct 1995 | A |
5475625 | Glaschick | Dec 1995 | A |
5475843 | Halviatti et al. | Dec 1995 | A |
5483468 | Chen et al. | Jan 1996 | A |
5491791 | Glowny et al. | Feb 1996 | A |
5504921 | Dev et al. | Apr 1996 | A |
5506955 | Chen et al. | Apr 1996 | A |
5513126 | Harkins et al. | Apr 1996 | A |
5553235 | Chen et al. | Sep 1996 | A |
5555191 | Hripcsak | Sep 1996 | A |
5572672 | Dewitt et al. | Nov 1996 | A |
5581482 | Wiedenman et al. | Dec 1996 | A |
5581696 | Kolawa et al. | Dec 1996 | A |
5586254 | Kondo et al. | Dec 1996 | A |
5598429 | Marshall | Jan 1997 | A |
5615135 | Waclawsky et al. | Mar 1997 | A |
5615323 | Engel et al. | Mar 1997 | A |
5616904 | Fernadez | Apr 1997 | A |
5619656 | Graf | Apr 1997 | A |
5621663 | Skagerling | Apr 1997 | A |
5634009 | Iddon et al. | May 1997 | A |
5636344 | Lewis | Jun 1997 | A |
5655081 | Bonnell et al. | Aug 1997 | A |
5661668 | Yemini et al. | Aug 1997 | A |
5668944 | Berry | Sep 1997 | A |
5675510 | Coffey et al. | Oct 1997 | A |
5684945 | Chen et al. | Nov 1997 | A |
5696486 | Poliquin et al. | Dec 1997 | A |
5696701 | Burgess et al. | Dec 1997 | A |
5699403 | Ronnen | Dec 1997 | A |
5732273 | Srivastava et al. | Mar 1998 | A |
5740441 | Yellin et al. | Apr 1998 | A |
5748098 | Grace | May 1998 | A |
5752062 | Gover et al. | May 1998 | A |
5758071 | Burgess et al. | May 1998 | A |
5778230 | Wimble et al. | Jul 1998 | A |
5796663 | Park et al. | Aug 1998 | A |
5799154 | Kuriyan | Aug 1998 | A |
5802256 | Heckerman et al. | Sep 1998 | A |
5815718 | Tock | Sep 1998 | A |
5819028 | Manghirmalani et al. | Oct 1998 | A |
5845125 | Nishimura et al. | Dec 1998 | A |
5850388 | Anderson et al. | Dec 1998 | A |
5870540 | Wang et al. | Feb 1999 | A |
5896536 | Lindsey | Apr 1999 | A |
5901315 | Edwards et al. | May 1999 | A |
5923834 | Thieret et al. | Jul 1999 | A |
5948113 | Johnson et al. | Sep 1999 | A |
5956662 | Hemker et al. | Sep 1999 | A |
5974237 | Shurmer et al. | Oct 1999 | A |
5974457 | Waclawsky et al. | Oct 1999 | A |
5978594 | Bonnell et al. | Nov 1999 | A |
5987249 | Grossman et al. | Nov 1999 | A |
5991881 | Conklin et al. | Nov 1999 | A |
6021437 | Chen et al. | Feb 2000 | A |
6026234 | Hanson et al. | Feb 2000 | A |
6026237 | Berry et al. | Feb 2000 | A |
6028842 | Chapman et al. | Feb 2000 | A |
6042614 | Davidson et al. | Mar 2000 | A |
6044335 | Ksendzov | Mar 2000 | A |
6058102 | Drysdale et al. | May 2000 | A |
6058393 | Meier et al. | May 2000 | A |
6061518 | Hoffman | May 2000 | A |
6067412 | Blake et al. | May 2000 | A |
6072777 | Bencheck et al. | Jun 2000 | A |
6073089 | Baker et al. | Jun 2000 | A |
6076083 | Baker | Jun 2000 | A |
6081797 | Hittt | Jun 2000 | A |
6085029 | Kolawa et al. | Jul 2000 | A |
6106572 | Halpern | Aug 2000 | A |
6108800 | Asawa | Aug 2000 | A |
6115393 | Engel et al. | Sep 2000 | A |
6118936 | Lauer et al. | Sep 2000 | A |
6118940 | Alexander, III et al. | Sep 2000 | A |
6119074 | Sarangapani | Sep 2000 | A |
6119247 | House et al. | Sep 2000 | A |
6138122 | Smith et al. | Oct 2000 | A |
6139198 | Danforth et al. | Oct 2000 | A |
6145121 | Levy et al. | Nov 2000 | A |
6149318 | Chase et al. | Nov 2000 | A |
6151701 | Humphreys et al. | Nov 2000 | A |
6154736 | Chickering et al. | Nov 2000 | A |
6161200 | Rees et al. | Dec 2000 | A |
6167398 | Wyard et al. | Dec 2000 | A |
6182022 | Mayle et al. | Jan 2001 | B1 |
6182157 | Schlener et al. | Jan 2001 | B1 |
6216119 | Jannarone | Apr 2001 | B1 |
6222652 | Roberts | Apr 2001 | B1 |
6249755 | Yemini et al. | Jun 2001 | B1 |
6263298 | Kerman et al. | Jul 2001 | B1 |
6269401 | Fletcher et al. | Jul 2001 | B1 |
6320585 | Engel et al. | Nov 2001 | B1 |
6327550 | Vinberg et al. | Dec 2001 | B1 |
6327677 | Garg et al. | Dec 2001 | B1 |
6359976 | Kalyanpur et al. | Mar 2002 | B1 |
6377907 | Waclawski | Apr 2002 | B1 |
6381306 | Lawson et al. | Apr 2002 | B1 |
6405250 | Lin et al. | Jun 2002 | B1 |
6446058 | Brown | Sep 2002 | B1 |
6453346 | Garg et al. | Sep 2002 | B1 |
6466929 | Brown et al. | Oct 2002 | B1 |
6470464 | Bertram et al. | Oct 2002 | B1 |
6484129 | Klein et al. | Nov 2002 | B1 |
6505246 | Land et al. | Jan 2003 | B1 |
6513065 | Hafez et al. | Jan 2003 | B1 |
6557035 | McKnight | Apr 2003 | B1 |
6615259 | Nguyen et al. | Sep 2003 | B1 |
6658367 | Conrad | Dec 2003 | B1 |
6792456 | Hellerstein et al. | Sep 2004 | B1 |
6810356 | Garcia-Franco et al. | Oct 2004 | B1 |
6836800 | Sweet et al. | Dec 2004 | B1 |
20010052087 | Garg et al. | Dec 2001 | A1 |
20020012011 | Roytman et al. | Jan 2002 | A1 |
20020019870 | Chirashnya et al. | Feb 2002 | A1 |
20020049838 | Sylor et al. | Apr 2002 | A1 |
20020054169 | Richardson | May 2002 | A1 |
20020090134 | Van Zon | Jul 2002 | A1 |
20020095661 | Angel et al. | Jul 2002 | A1 |
20020111755 | Valadarsky et al. | Aug 2002 | A1 |
20020123865 | Whitney et al. | Sep 2002 | A1 |
20020133757 | Bertram et al | Sep 2002 | A1 |
20020152185 | Jamadagni | Oct 2002 | A1 |
20020152305 | Jackson et al. | Oct 2002 | A1 |
20020158918 | Feibush et al. | Oct 2002 | A1 |
20020159386 | Grosdidier et al. | Oct 2002 | A1 |
20020165892 | Grumann et al. | Nov 2002 | A1 |
20020169870 | Vosseler et al. | Nov 2002 | A1 |
20020173997 | Menard et al. | Nov 2002 | A1 |
20020174174 | Ramraj et al. | Nov 2002 | A1 |
20020183972 | Enck et al. | Dec 2002 | A1 |
20020184065 | Menard et al. | Dec 2002 | A1 |
20020184615 | Sumner et al. | Dec 2002 | A1 |
20020198879 | Schwarcz | Dec 2002 | A1 |
20020198985 | Fraenkel et al. | Dec 2002 | A1 |
20030005362 | Miller et al. | Jan 2003 | A1 |
20030009507 | Shum | Jan 2003 | A1 |
20030014692 | James et al. | Jan 2003 | A1 |
20030018241 | Mannheimer | Jan 2003 | A1 |
20030018494 | Bronstein et al. | Jan 2003 | A1 |
20030023712 | Zhao et al. | Jan 2003 | A1 |
20030028631 | Rhodes | Feb 2003 | A1 |
20030033404 | Richardson | Feb 2003 | A1 |
Number | Date | Country |
---|---|---|
296 00 609 | Mar 1997 | DE |
0 508 386 | Oct 1992 | EP |
2 783 620 | Mar 2000 | FR |
WO 9612224 | Apr 1996 | WO |
WO 9913427 | Mar 1999 | WO |
Number | Date | Country | |
---|---|---|---|
20030079160 A1 | Apr 2003 | US |
Number | Date | Country | |
---|---|---|---|
60322021 | Sep 2001 | US | |
60307055 | Jul 2001 | US |