The present invention generally relates to business intelligence or data warehouse systems, and more specifically to a method for characterizing the workload of a new business intelligence system that has not yet been run in a production environment. In particular, the workload is characterized in terms of resource usage and performance characteristics.
Database systems perform a vital role in the information technology infrastructure of a business or corporation. Specialized databases for data warehousing and data analysis are becoming an important segment of the total database market. Business intelligence systems transform raw data into useful information. Common applications of business intelligence systems are, for example, fraud detection, risk analysis, market segmentation, and profitability analysis.
In a business intelligence system, data is extracted from heterogeneous operational databases and external data sources, then cleansed, transformed, and loaded into a large data warehouse or data mart storage areas. Data warehouses are subject-oriented, integrated, and time-varied collections of data used primarily for making decisions. Data marts are departmentalized subsets of the data warehouse focusing on selected subjects, rather than the entire enterprise data.
Data is stored and managed by one or more data warehouse servers that provide data access to front end tools for querying, reporting, analysis, and mining. Specialized online analytical processing (OLAP) servers may also be used to construct multidimensional views of the data, where operations on data can be performed.
Business intelligence workloads have different characteristics than the traditional transaction processing workloads used in conventional capacity planning and sizing methods. Business intelligence workloads place a greater emphasis on summarized and consolidated data as opposed to individual records. Business intelligence workloads typically use a very large size database. Queries of business intelligence workloads are heterogeneous, complex, and ad-hoc in nature, varying greatly in the amount of time required to execute the queries. These queries often touch millions of records and may perform many table joins, sorts, or aggregations. Furthermore, queries of business intelligence workloads can produce very large results sets, requiring a large amount of concurrent I/O.
Computer capacity planning is the process of analyzing and projecting an existing workload to determine the type of hardware resources needed to meet future demand and to predict when system saturation occurs. The capacity planning process can be long and challenging, depending on the size and complexity of the application, the quality and quantity of information available, as well as the approaches and tools employed. A computer capacity-sizing expert often performs computer capacity planning manually with insufficient information and using an unstructured, informal approach.
Database system sizing attempts to arrive at an initial estimate of a hardware configuration that satisfies performance demands, cost constraints, and functional requirements of a new business intelligence system. Typically, detailed information about the system and its workload are not available during the sizing process. In conventional database system sizing, a sizing expert uses published performance results of a similar workload with similar performance requirements. The sizing expert extrapolates these results to the new business intelligence system. This extrapolation is performed using informal industry guidelines (“rules of thumb”) and published performance relationships between different types of hardware. The sizing expert thus obtains an initial estimate of the hardware configuration comprising the processor, disk, and memory required to meet resource demands of the expected workload and the expected size of the database.
Selecting the appropriate hardware resources can be a complicated task because of the wide variety of processor, disk, network, and memory technologies available. Further, determining the quantity of each resource needed and predicting how the different components interact under a specific workload are non-trivial tasks.
Many of the approaches and tools used for capacity planning were developed in the late-1970s and early-1980s when mainframe computers were the dominant computing platform. Mainframes were very expensive; therefore it was critical to perform detailed planning and analysis before a particular model was purchased. A variety of tools were created to help a planner with this task, including tools for performance monitoring, workload forecasting, performance simulation, and design/configuration advice.
As mainframe architectures slowly gave way to client-server, and more recently, n-tier architectures, the focus on planning was not as systematic. This may be partly attributed to financial factors; namely the declining cost and improving performance of computer hardware. The cost of a cluster of inexpensive server machines networked together became substantially less than that of a mainframe. Fixing configuration errors resulting from poor planning could cost in the thousands of dollars for n-tier architectures versus millions of dollars for mainframes. The additional cost to perform detailed planning analysis often exceeded the costs to correct configuration errors, thus complete planning studies were relatively unattractive.
The complexity of modeling performance in n-tier architectures also makes planning more difficult. Traditional methods used for mainframes are not directly transferable to n-tier architectures. In the mainframe domain, components such as processors, disks, and memory share similar designs and characteristics; however, this is not the case in n-tier environments. The proliferation of competing and sophisticated processor, disk, memory, and network technologies makes creating generic performance models very difficult.
The resource demands of modern applications are also more complex and demanding in nature than in the past, making their performance less predictable. The popularity and commercialization of the Internet and World Wide Web fostered the demand for newer and richer data such as graphics, audio, video, and XML. Whereas this data was once stored for archival purposes only, companies have now started analyzing it with specialized data analysis applications to discover new information about their business and customers. This places additional resource burdens on systems in addition to the traditional transaction processing workloads being handled.
Time and business pressures also make detailed capacity planning studies infeasible. In today's e-business on demand environment, customers demand and expect answers in a timely fashion. A day or week is often a critical amount of time for completing a hardware sale. This implies that any planning analysis needs to be performed quickly while maintaining a high degree of accuracy.
The result of the sizing process is an initial estimate of the hardware configuration (processor, disk, and memory) needed to meet the resource demands of the expected workload and size of database. Customers expect a cost-efficient and effective hardware solution that meets the performance requirements of their application while offering the maneuverability to accommodate future expansion. There is generally no opportunity for experts to validate their hardware recommendations because of financial and time constraints. The sizing process currently involves significant manual effort to complete.
Successful sizing of a business intelligence system requires a characterization of the anticipated workload. Workload characterization dates back to the 1970s when workloads largely comprised large transactions and batch jobs performed on mainframe computers. Techniques have evolved to accommodate some of the modern workloads encountered in current computing environments. The majority of conventional workload characterization approaches assume that detailed performance measurements from a production environment are available to build models of system performance.
One conventional workload characterization approach uses clustering analysis to construct a profile of a data warehousing workload to summarize the characteristics of the workload. The profile can be used to help a designer during logical and physical optimization. The profile can be further used to generate workloads useful for evaluating system performance in testing and benchmark settings.
Conventional techniques for workload characterization rely on a combination of structural query properties and statistical parameters to perform a clustering. The structural properties are based on the text of a query, such as the number of table joins, the number of predicates, and the type of predicates. Statistical parameters comprise quantitative values in the catalog tables of a database system, such as table size, size, the type of indexes on a table, and the skew of data values in the table. Although these techniques have proven to be useful, it would be desirable to present additional improvements. Currently, there is no known conventional technique for workload characterization that makes use of performance-oriented measurements, that are subsequently used to aid in the sizing of a new database system running a business intelligence workload.
Conventional techniques for selecting an initial size of a hardware configuration for a database system are manually performed by sizing experts. Furthermore, conventional techniques assume that little system environment information or performance measurements are available, thus a sizing expert relies on extrapolations from similar workloads, personal experience, industry benchmarks, informal industry guidelines, and hardware performance guidelines to determine the type and quantity of required resources. Currently, there is no available method for characterizing an anticipated workload based on performance-oriented characterization of a similar workload.
What is therefore needed is a system, a service, a computer program product, and an associated method for characterizing a business intelligence workload to aid in sizing the hardware configuration of a new database system. The need for such a solution has heretofore remained unsatisfied.
The present invention satisfies this need, and presents a system, a service, a computer program product, and an associated method (collectively referred to herein as “the system” or “the present system”) for characterizing the workload of a business intelligence system. This characterization is used to facilitate the process of sizing a hardware configuration required by a new database system. The present system is a workload characterization analysis that can be used as a basis for describing business intelligence workloads. The present system applies unsupervised data mining techniques to group individual business intelligence queries into general classes of queries based on system resource usage. The workload characterization generated by the present system provides insight into the resource demands of classes of queries typically found in a business intelligence workload. The workload characterization can be further exploited in, for example, workload management or meeting quality of service requirements.
The present system utilizes performance-oriented measurements to characterize a workload. The present system further employs clustering algorithms known for revealing underlying or hidden dimensions in data. The ability to reveal underlying dimensions is particularly useful because of the many different interactions occurring between software and hardware during query processing that affect system performance.
Business intelligence applications and their workloads vary depending on the type of application, the target industry, and the nature of business questions being asked. The present system uses an exemplary workload, specifically, the TPC-H benchmark, as a representative workload for characterizing a business intelligence system. The TPC-H benchmark comprises 22 ad-hoc queries that answer questions representative of any industry that manages, sells, or distributes a produce worldwide, such as a car rental business, a food distribution business, a parts business, a supplier, etc.
It is difficult to build an accurate model if a workload is considered as a single entity (i.e., an average of the heterogeneous queries comprising it). A workload model also becomes too complex if each individual query is considered, in effect, as an independent workload. The present system achieves a balanced and practical solution by partitioning a collection of queries into a few general classes of queries, based on the system resource usage by the queries. Each class comprises queries that are similar to each other based on resource usage and other relevant characteristics.
The present system identifies the basic components of a workload, chooses characterizing parameters for the workload, collects data for the workload, normalizes the collected data, partitions the workload into classes, and identifies interesting characteristics of the partitioned classes.
The present system groups or clusters the queries of the exemplary workload into broad categories, each with different characteristics. One group describes trivial types of queries, with short run-times, a small number of tables being joined, and exhibiting high CPU utilization. Another group represents simple queries that are I/O-bound and have a small number of tables being joined. A further group represents moderate-complexity queries with moderately high response times, and moderate CPU and I/O usage. Yet another group represents complex queries that are long-running, have a large number of tables being joined and exhibit high sequential and random I/O usage.
The present invention may be embodied in a utility program such as a workload characterization utility program. The present invention provides means for the user to provide the set of data to be characterized by the present system. The present invention further provides means to help the user to characterize a workload or set of data from descriptions of groups or clusters generated. A user specifies a benchmark, workload, or set of data and then invokes the workload characterization utility program. The user then describes the benchmark, workload, or set of data in terms of the clusters generated by the workload characterization utility program.
The various features of the present invention and the manner of attaining them will be described in greater detail with reference to the following description, claims, and drawings, wherein reference numerals are reused, where appropriate, to indicate a correspondence between the referenced items, and wherein:
Clients, such as remote Internet users, are represented by a variety of computers such as computers 20, 25, 30, and can access the host server 15 through a network 35. Computers 20, 25, 30 each comprise software that allows the user to interface securely with the host server 15. The host server 15 is connected to network 35 via a communications link 40 such as a telephone, cable, or satellite link. Computers 20, 25, 30, can be connected to network 35 via communications links 45, 50, 55, respectively. While system 10 is described in terms of network 35, computers 20, 25, 30 may also access system 10 locally rather than remotely. Computers 20, 25, 30 may access system 10 either manually, or automatically through the use of an application.
A client provides input to a workload characterization system 10 for a business intelligence system. System 10 facilitates the process of determining the broad characterizations that may be use to describe the workload composition of the new business intelligence system.
System 10 uses a workload, such as the 22 queries of the TPC-H benchmark, as a representative business intelligence workload in order to identify the basic components of a typical business intelligence workload and to simulate the business intelligence environment (step 305). While the present system is described in terms of a benchmark like the TPC-H benchmark, it should be clear that any standard workload may be used. For example, several workloads may be operating concurrently. The selected benchmark simulates the type of resource activity commonly found in business intelligence system implementations.
Care is taken to ensure that each benchmarked system is balanced. A system is considered balanced when all of the resources of the system are operating in concert to allow an optimal amount of workload through the system to meet specific objectives. System 10 assumes that the performance of a balanced system is relative to the available quantity of system resources. For example, if the quantity of system resources is increased in a balanced fashion, system performance increases. The difference in performance between benchmark configurations is typically relative. That is, the performance of a particular query in a benchmark run, is generally relative to the performance of other queries in the same benchmark run.
The parameter selection module 210 selects performance-oriented parameters to analyze for each individual query (step 310). Exemplary performance-oriented parameters comprise, for example, response time, average processor (CPU) utilization, sequential Input/Output (I/O) throughput rate, and the rate of random I/O operations per second. The response time is the amount of time (in seconds) that elapses from query submission to result set return. The average processor (CPU) utilization is the average utilization of the processor(s) over the duration of query execution.
The average processor utilization comprises the utilization of user processes, as opposed to the operating system kernel or privileged threads/processes. The sequential Input/Output (I/O) throughput rate is the average rate that data is sequentially read from disk over the duration of query execution, measured in megabytes per second (MB/second). The random I/O operations per second (IOPS) rate is the average rate of random I/O requests processed per time window over the duration of query execution. In one embodiment, system 10 monitors other types of parameters, such as memory and network utilization.
System 10 accepts workload data arising from different system configurations comprising varied computer models, hardware parts, operating systems, and database scales. The data collection module 215 accepts raw performance data obtained from the different systems. This data is typically obtained using standard operating system performance monitoring tools (step 315). These monitoring tools are configured to sample the desired parameters at a predetermined interval, such as five-second intervals. Representative parameter values for each query are determined by averaging the raw data samples collected over a predetermined elapsed time for each respective parameter. Table 1 illustrates an exemplary sample of collected performance data.
The choice of units of measurement can affect the characterization analysis of system 10. For instance, expressing temporal data in seconds versus hours can produce a different result, depending on the type of analysis technique used. To avoid dependence on the choice of units, the normalization module 220 standardizes the collected data through normalization (step 320). To normalize the data, the normalization module 220 calculates a z-score of each measured parameter variable. A z-score transforms the dataset of measured parameters into a dataset with a mean of 0 and standard deviation of 1. The z-score of a parameter value can be calculated as follows:
The normalized data resembles a format similar to that shown in Table 2.
Once the data for each benchmark is normalized, the normalization module 220 combines all the data into a single matrix (table), which is used by the partitioning module 225 to partition the workload into classes.
The partitioning module 225 employs clustering techniques to partition the workload (step 325). Clustering is the process of grouping data into classes or clusters so that objects within a cluster are similar to each other, but are dissimilar to objects in other clusters. The partitioning module 225 utilizes singular value decomposition (SVD) and semi-discrete decomposition (SDD) to partition the workload into classes. When used in combination with each other, SVD positions the dataset of system 10 in a graphical space while SDD provides further classification of the dataset within that graphical space. Singular value decomposition and semi-discrete decomposition are examples of unsupervised data mining techniques. Unsupervised data mining discovers structured information in a dataset without prior knowledge or user-provided hints as to what the structure might looks like.
Singular value decomposition and semi-discrete decomposition view the dataset as a matrix and decompose a dataset matrix into a product of three new matrices. However, the structure and meaning of each of the new matrices is different in each technique.
SVD decomposes a dataset matrix A into the product of matrices, U, S, and V such that:
A=U S VT
where U is n×m, S is a diagonal matrix of non-increasing non-negative values, and V is m×m. In effect, SVD transforms an m-dimensional space into a new m-dimensional space. The new m-dimensional space comprises axes that are orthonormal and ordered so that a maximum amount of variation is contained in the first m axes in the new space. The entries in the matrix S are scaling factors indicating the relative importance of each axis. Geometrically, the rows of U represent coordinates of the corresponding rows of A in a space spanned by the columns of V, while the rows of V represent the coordinates of the corresponding columns of A. A common practice in SVD is to truncate the representation to k dimensions, where k is some arbitrary constant, to make analysis more manageable. Since SVD concentrates as much variation as possible into the first few dimensions, truncating is feasible because the least possible information is discarded.
SDD is similar to SVD in that it decomposes a dataset matrix A into a product of three matrices, such that:
A=X D Y
However, the matrices of the SDD have a different form and meaning than SVD. X is an n×k matrix, D is a k×k diagonal matrix, and Y is a k×m matrix, where k is an arbitrary constant. The entries of X and Y are from the set {−1, 0, +1}. Objects are divided based on their value in an initial column of X (−1, 0, +1). Objects can be further subdivided according to their values in the subsequent columns of X. In effect, SDD discovers rectilinearly aligned regions of the matrix of similar (positive and negative) magnitude. These regions/partitions determine which objects are related.
SVD and SDD can be jointly applied to the dataset of system 10 by using both decompositions, truncating the SVD at k=3, plotting the points corresponding to queries, and labeling each point according to its location in the top few levels of the SDD decomposition.
In one embodiment, additional attributes are added to each row of the dataset of Table 1. One such attribute is, for example, the size of the largest n-way table join in each query. Table joins are a prominent characteristic of business intelligence queries and are processing-intensive. The addition of this attribute results in a tighter clustering of data points due to a closer relation between queries with the same label. For example, the data values for all the query 1s should appear closer together in the clustering, since the size of the largest n-way table join is the same for query 1, regardless of the system it is run on.
In the example of
As illustrated in
In general, cluster 2, 515, represents fairly simple queries that are IO-bound in nature and have a small number of tables being joined. Cluster 3, 520, represents long-running, large and complex queries with a large number of tables being joined (for example, greater than 5 joins). Queries in cluster 3, 520, further exhibit high sequential and random I/O usage. Cluster 4, 525, represents short-running trivial queries with a varying amount of tables being joined (for example, 3 to 8 table joins). Cluster 1, 510, represents medium-running queries with a smaller number of tables being joined (for example, 5 or fewer joins) and exhibiting high CPU utilization. Cluster 1, 510, is considered less interesting since the data points in cluster 1, 510, are those closest to the origin in the graph of data points shown in
Artificially-generated data points are used to further lend support to the meaning of the clusters 530.
To further illustrate the validity and appropriateness of the clustering performed by system 10, another set of artificially-generated query data points is created to represent extreme examples of each dimension. If the SVD transformation of these artificial query data points places them at extreme ends of one of the transformed dimensions, further evidence is provided to interpret the meaning of the new dimension.
It is to be understood that the specific embodiments of the invention that have been described are merely illustrative of certain applications of the principle of the present invention. Numerous modifications may be made to a system, service, and method for characterizing a business intelligence workload for sizing a new database system hardware configuration described herein without departing from the spirit and scope of the present invention. Moreover, while the present invention is described for illustration purpose only in relation to users connected through a network, it should be clear that the invention is applicable as well to, for example, to local users.
Furthermore, while the present invention is described for illustration purposes only in relation to a business intelligence workload, it should be clear that the invention is applicable as well to, for example, workloads for any type of database system or any other computational system using queries. Furthermore, while system 10 is described in terms of a benchmark for queries of a business intelligence workload, it should be clear that system 10 operate on any set of data requiring characterization.
The present application is related to co-pending U.S. patent application, titled “System, Method, And Service For Automatically Determining An Initial Sizing Of A Hardware Configuration For A Database System Running A Business Intelligence Workload,” Ser. No. _____, filed concurrently herewith, which is assigned to the same assignee as the present invention, and which is incorporated herein by reference.