際際滷

際際滷Share a Scribd company logo

Enterprise applications in the cloud: non-virtualized deployment

Leonid Grinshpan, Ph.D.
Leonid Grinshpan, Ph.D.

The cloud is a platform devised to support a number of concurrently working applications that share the clouds resources; being a platform of common use, the cloud features complex interdependencies among hosted applications, as well as among applications and the underlying hardware platform. The paper stydies non-virtualized deployment when a number of applications are hosted on the same physical server without logical borders among applications (no partitions, virtual machines or similar technologies in place).

1 of 11
Enterprise Applications in the Cloud: Non-virtualized Deployment Leonid Grinshpan, Oracle Corporation (www.oracle.com) Subject The cloud is a platform devised to support a number of concurrently working applications that share the clouds resources; being a platform of common use, the cloud features complex interdependencies among hosted applications, as well as among applications and the underlying hardware platform. Enterprise Applications (EAs) can be deployed in the cloud in two ways: 1. Non-virtualized setup hosts on the same physical servers different EAs without logical borders between them (no partitions, virtual machines or similar technologies in place). 2. Virtualized arrangement separates EAs logically from each other by employing the above-mentioned techniques. Both deployment models have advantages and disadvantages. The performance penalty introduced by virtualization (which we will analyze in the next article) prevents many EA vendors from recommending EA deployment in virtual environments. As an example, here is a policy of Thomson Reuters Elite business (http://www.elite.com/virtualization_servers/):
Elite generally recommends against using virtualization environments (e.g., Virtual Machines from VMware or Microsoft Virtual Server) for primary production servers hosting Elite products. Elite makes no performance warranties in relation to Elite applications hosted on Virtual Machines. In non-virtualized clouds an allocation of resources for different EAs is carried out by operating systems that provision software processes representing EAs. This environment makes all processing power of the physical servers available to the applications. Furthermore, it enables collection of reliable performance metrics by directly monitoring the servers counters. It is quite possible that for those reasons, Google applications are not embedded into virtual environments. Another example of a non-virtualized cloud is the popular project management and collaboration tool Basecamp [Is Virtualization a Cloud Prerequisite? http://gigaom.com/2009/08/30/is-virtualization-a-cloud-prerequisite/] One shortcoming of non-virtualized cloud is obvious: Instability of any EA resulting in hardware downtime affects availability of all EAs. But what happens much more often is that an EA might suffer performance degradation in response to the changes in workload and service demand experienced by any other EA. We analyze that phenomenon simulating happenings in the cloud using queuing models of EAs; methodological foundation for EA performance analysis based on queuing models can be found in the authors book [Leonid Grinshpan. Solving Enterprise Application Performance Puzzles: Queuing Models to the Rescue, Willey-IEEE Press; available in bookstores and from Web booksellers from January 2012]. Performance Impact of Workload Fluctuations The queuing model on Figure 1 represents simplified three-tiered Cloud with Web, Application, and Database servers. The cloud hosts three EAs (App A, App B, App C) serving three user groups, one group per EA. Each server corresponds to the models node with a number of processing units equal to the number of CPUs in a server. The users of each EA, as well as the network, are modeled by dedicated nodes. All servers are physical ones without any partitioning among applications. Web and Application servers have 8 CPUs each; Database server has 16 CPUs.
Figure 1 Model 1 of the cloud hosting three enterprise applications The models in this article were analyzed using TeamQuest solver [http://teamquest.com/products/model/index.htm]. Here is a description of Model 1 components in TeamQuest terms: Workload 1 for Model 1 is characterized in Table 1. For each application it is represented by transactions identified by application name. A user initiates transaction a number of times indicated in column Number of transaction executions per user per hour. We analyze the model for 200, 400, 600, and 800 users.
Table 1 Workload 1 for Model 1 Number of users Number of Transaction name Total Total Total Total Total transaction 3 200 400 600 800 executions per user per hour App A transaction 1 100 200 300 400 10 App B transaction 1 50 100 150 200 20 App C transaction 1 50 100 150 200 5 To solve the model we have to specify the profile of each transaction (Table 2). The Transaction Profile is a set of time intervals (service demands) a transaction has spent in all processing units it has visited while served by application. Table 2 Transaction Profiles (seconds) Time in Time in Web Time in App Time in Database Network node server node server node server node App A transaction 0.001 0.2 1.0 5.0 App B transaction 0.0015 0.1 0.5 2.5 App C transaction 0.003 0.2 5.0 5.0 Model 1 estimates that transaction times for all applications will start increasing when the number of users exceeds 400 (Figure 2). Figure 2 Transaction response times for three applications
To find a reason for transaction time degradation, lets look at utilization of the clouds servers (Figure 3). Figure 3 Utilization of the clouds servers When the number of users is close to 600, utilization of Database server exceeds 85% and causes noticeable increase in transaction times for all applications. Any further growth in the number of users maxes out the Database server and results in exponential explosion of transaction time. Non-virtualized cloud does not discriminateit punishes all applications by increasing their transaction times. Model 1 helps to determine the contribution of each application into Database server utilization (Figure 4). Figure 4 Breakdown of utilization of Database server by App A, B, and C (percentage)
Per Figure 4 the largest consumer of Database server capacity is App A. We will analyze how the decrease of App A number of users affects all other applications. Workload 2 in Table 3 shows that when we keep the number of users of App A limited to 100, for two other applications they remain the same as was specified in Table 1. Table 3 Workload 2 for Model 1 Number of users Number of Transaction name Total Total Total Total Total transaction 3 200 300 400 500 executions per user per hour App A transaction 1 100 100 100 100 10 App B transaction 1 50 100 150 200 20 App C transaction 1 50 100 150 200 5 Transaction times and utilization of the clouds server for Workload 2 are pictured on Figures 5 and 6; charts suggest that all applications now provide acceptable services for their users. Figure 5 Transaction response times for three applications for Workload 2
Figure 6 Utilization of the clouds servers for Workload 2 Performance Impact of Service Demand Fluctuations The transactions service demand for a particular hardware component characterizes the time this component has to spend to process the transaction. A transaction can be in two states: either waiting for resource or using resource. A service demand for a particular resource means that a time transaction is using this resource. Service demand in general depends on two parameters: - Resources processing speed - Volume of data to be processed by resource The first parameter characterizes hardware resource (for example, disk transfer rate is 1000 Mbit/second). For a resource such as CPU, a processing speed depends on clock speed (for example, 3 GHz), as well as on software algorithms. Resource processing speed is a constant for given hardware component and software release. To the contrary, the second parameters prevailing trend is an increase in data volume, because over time, business accumulates more data (for example, more sales executed in June than in January). If EA transaction represents a financial report on sales volume, then generation of the reports data for June will take more Database server time than
generation of the reports data for January. We analyze an impact of service demand fluctuations on EA performance using Model 2. Model 2 has the same topology and the same workload as Model 1 (Figure 1 and Table 1); the difference is that we analyze Model 2 for three values of service demands from App C transaction for processing in Database (5, 8, and 10 seconds; see Table 4). Table 4 Transaction Profiles (seconds) with Different Service Demands Time in Time in Web Time in App Time in Database server Network node server node server node node App A transaction 0.001 0.2 1.0 5.0 App B transaction 0.0015 0.1 0.5 2.5 App C transaction 0.003 0.2 5.0 Model analyzed for 5.0, 8.0, 10.0 seconds Model 2 predicts that times of all application transactions degrade faster with an increase of service demand from one application (Figure 7).
Figure 7 Transaction response times for different service demands (sd in the legend stands for service demand) Monitoring Resource Consumptions by Applications In a non-virtualized environment, each application is materialized by the operating system as the number associated with its processes. To find out resource consumptions by application, we have to monitor its processes. Figure 8 pictures Windows Task Manager reporting performance data for two processes belonging to the same application: Process beasvc.exe is an application server and process oracle.exe is a database. CPU column shows CPU usage as a percentage of time that process used the CPU since last update; Mem Usage column delivers a size of current working set of a process in kilobytes; Threads column counts a number of threads running in a process; remaining columns relate to I/O system.
More detailed information on processes can be collected using Windows Performance Monitor. For example, Performance Monitor reports for each process a speed of I/O operations, number of I/O operations per second, memory paging information, thread count and state of each thread, usage of virtual memory, etc. Performance Monitor can save collected data into log files; analyses of such log files discover trends in data collected in performance counters. Figure 8 Performance data reported by Windows Task Manager for different processes In UNIX environment process level data can be examined, for example, by using prstat a command (Figure 9): Process ID is reported in first column PID, columns SIZE and RSS specify memory usage and column CPU delivers CPU utilization. This command helps to associate process name (last column) and process ID. Figure 9 Process information in UNIX environment as reported by prstat a command If we have to log data on process resource utilization over some period of time, we can use the following command for a process named <process name>: while true; do ps -eo vsz,rss,pcpu,comm| grep <process name> >> log.txt; sleep 5; done This command collects the counters every five seconds: vsz - Total size of the process in virtual memory, in kilobytes.
rss - Resident set size of the process, in kilobytes. pcpu - Percentage of CPU utilization. We mentioned a few basic ways of providing insight into resource consumption by each hosted application; operating systems and third party monitoring products offer a broad spectrum of means to collect process-level data to make confident decisions on cloud capacity management. Take Away from the Article 1. In non-virtualized cloud performance, bottlenecks created by a shortage of hardware resources affect all applications by increasing their transaction times. 2. Cloud-wide bottleneck can be caused by increase in workload of any hosted application, as well as by fluctuations in its service demand while processing large or complex data. 3. A contribution of each application to resources utilization can be found by monitoring software processes created by the operating system for a particular application. Queuing models provide estimates of such contribution. About the Author During the last fifteen years as an Oracle consultant, the author was engaged in hands- on performance tuning and sizing of enterprise applications for various corporations (Dell, Citibank, Verizon, Clorox, Bank of America, AT&T, Best Buy, Aetna, Halliburton, Pfizer, Astra Zeneca, Starbucks, etc).

More Related Content

Enterprise applications in the cloud: non-virtualized deployment

  • 1. Enterprise Applications in the Cloud: Non-virtualized Deployment Leonid Grinshpan, Oracle Corporation (www.oracle.com) Subject The cloud is a platform devised to support a number of concurrently working applications that share the clouds resources; being a platform of common use, the cloud features complex interdependencies among hosted applications, as well as among applications and the underlying hardware platform. Enterprise Applications (EAs) can be deployed in the cloud in two ways: 1. Non-virtualized setup hosts on the same physical servers different EAs without logical borders between them (no partitions, virtual machines or similar technologies in place). 2. Virtualized arrangement separates EAs logically from each other by employing the above-mentioned techniques. Both deployment models have advantages and disadvantages. The performance penalty introduced by virtualization (which we will analyze in the next article) prevents many EA vendors from recommending EA deployment in virtual environments. As an example, here is a policy of Thomson Reuters Elite business (http://www.elite.com/virtualization_servers/):
  • 2. Elite generally recommends against using virtualization environments (e.g., Virtual Machines from VMware or Microsoft Virtual Server) for primary production servers hosting Elite products. Elite makes no performance warranties in relation to Elite applications hosted on Virtual Machines. In non-virtualized clouds an allocation of resources for different EAs is carried out by operating systems that provision software processes representing EAs. This environment makes all processing power of the physical servers available to the applications. Furthermore, it enables collection of reliable performance metrics by directly monitoring the servers counters. It is quite possible that for those reasons, Google applications are not embedded into virtual environments. Another example of a non-virtualized cloud is the popular project management and collaboration tool Basecamp [Is Virtualization a Cloud Prerequisite? http://gigaom.com/2009/08/30/is-virtualization-a-cloud-prerequisite/] One shortcoming of non-virtualized cloud is obvious: Instability of any EA resulting in hardware downtime affects availability of all EAs. But what happens much more often is that an EA might suffer performance degradation in response to the changes in workload and service demand experienced by any other EA. We analyze that phenomenon simulating happenings in the cloud using queuing models of EAs; methodological foundation for EA performance analysis based on queuing models can be found in the authors book [Leonid Grinshpan. Solving Enterprise Application Performance Puzzles: Queuing Models to the Rescue, Willey-IEEE Press; available in bookstores and from Web booksellers from January 2012]. Performance Impact of Workload Fluctuations The queuing model on Figure 1 represents simplified three-tiered Cloud with Web, Application, and Database servers. The cloud hosts three EAs (App A, App B, App C) serving three user groups, one group per EA. Each server corresponds to the models node with a number of processing units equal to the number of CPUs in a server. The users of each EA, as well as the network, are modeled by dedicated nodes. All servers are physical ones without any partitioning among applications. Web and Application servers have 8 CPUs each; Database server has 16 CPUs.
  • 3. Figure 1 Model 1 of the cloud hosting three enterprise applications The models in this article were analyzed using TeamQuest solver [http://teamquest.com/products/model/index.htm]. Here is a description of Model 1 components in TeamQuest terms: Workload 1 for Model 1 is characterized in Table 1. For each application it is represented by transactions identified by application name. A user initiates transaction a number of times indicated in column Number of transaction executions per user per hour. We analyze the model for 200, 400, 600, and 800 users.
  • 4. Table 1 Workload 1 for Model 1 Number of users Number of Transaction name Total Total Total Total Total transaction 3 200 400 600 800 executions per user per hour App A transaction 1 100 200 300 400 10 App B transaction 1 50 100 150 200 20 App C transaction 1 50 100 150 200 5 To solve the model we have to specify the profile of each transaction (Table 2). The Transaction Profile is a set of time intervals (service demands) a transaction has spent in all processing units it has visited while served by application. Table 2 Transaction Profiles (seconds) Time in Time in Web Time in App Time in Database Network node server node server node server node App A transaction 0.001 0.2 1.0 5.0 App B transaction 0.0015 0.1 0.5 2.5 App C transaction 0.003 0.2 5.0 5.0 Model 1 estimates that transaction times for all applications will start increasing when the number of users exceeds 400 (Figure 2). Figure 2 Transaction response times for three applications
  • 5. To find a reason for transaction time degradation, lets look at utilization of the clouds servers (Figure 3). Figure 3 Utilization of the clouds servers When the number of users is close to 600, utilization of Database server exceeds 85% and causes noticeable increase in transaction times for all applications. Any further growth in the number of users maxes out the Database server and results in exponential explosion of transaction time. Non-virtualized cloud does not discriminateit punishes all applications by increasing their transaction times. Model 1 helps to determine the contribution of each application into Database server utilization (Figure 4). Figure 4 Breakdown of utilization of Database server by App A, B, and C (percentage)
  • 6. Per Figure 4 the largest consumer of Database server capacity is App A. We will analyze how the decrease of App A number of users affects all other applications. Workload 2 in Table 3 shows that when we keep the number of users of App A limited to 100, for two other applications they remain the same as was specified in Table 1. Table 3 Workload 2 for Model 1 Number of users Number of Transaction name Total Total Total Total Total transaction 3 200 300 400 500 executions per user per hour App A transaction 1 100 100 100 100 10 App B transaction 1 50 100 150 200 20 App C transaction 1 50 100 150 200 5 Transaction times and utilization of the clouds server for Workload 2 are pictured on Figures 5 and 6; charts suggest that all applications now provide acceptable services for their users. Figure 5 Transaction response times for three applications for Workload 2
  • 7. Figure 6 Utilization of the clouds servers for Workload 2 Performance Impact of Service Demand Fluctuations The transactions service demand for a particular hardware component characterizes the time this component has to spend to process the transaction. A transaction can be in two states: either waiting for resource or using resource. A service demand for a particular resource means that a time transaction is using this resource. Service demand in general depends on two parameters: - Resources processing speed - Volume of data to be processed by resource The first parameter characterizes hardware resource (for example, disk transfer rate is 1000 Mbit/second). For a resource such as CPU, a processing speed depends on clock speed (for example, 3 GHz), as well as on software algorithms. Resource processing speed is a constant for given hardware component and software release. To the contrary, the second parameters prevailing trend is an increase in data volume, because over time, business accumulates more data (for example, more sales executed in June than in January). If EA transaction represents a financial report on sales volume, then generation of the reports data for June will take more Database server time than
  • 8. generation of the reports data for January. We analyze an impact of service demand fluctuations on EA performance using Model 2. Model 2 has the same topology and the same workload as Model 1 (Figure 1 and Table 1); the difference is that we analyze Model 2 for three values of service demands from App C transaction for processing in Database (5, 8, and 10 seconds; see Table 4). Table 4 Transaction Profiles (seconds) with Different Service Demands Time in Time in Web Time in App Time in Database server Network node server node server node node App A transaction 0.001 0.2 1.0 5.0 App B transaction 0.0015 0.1 0.5 2.5 App C transaction 0.003 0.2 5.0 Model analyzed for 5.0, 8.0, 10.0 seconds Model 2 predicts that times of all application transactions degrade faster with an increase of service demand from one application (Figure 7).
  • 9. Figure 7 Transaction response times for different service demands (sd in the legend stands for service demand) Monitoring Resource Consumptions by Applications In a non-virtualized environment, each application is materialized by the operating system as the number associated with its processes. To find out resource consumptions by application, we have to monitor its processes. Figure 8 pictures Windows Task Manager reporting performance data for two processes belonging to the same application: Process beasvc.exe is an application server and process oracle.exe is a database. CPU column shows CPU usage as a percentage of time that process used the CPU since last update; Mem Usage column delivers a size of current working set of a process in kilobytes; Threads column counts a number of threads running in a process; remaining columns relate to I/O system.
  • 10. More detailed information on processes can be collected using Windows Performance Monitor. For example, Performance Monitor reports for each process a speed of I/O operations, number of I/O operations per second, memory paging information, thread count and state of each thread, usage of virtual memory, etc. Performance Monitor can save collected data into log files; analyses of such log files discover trends in data collected in performance counters. Figure 8 Performance data reported by Windows Task Manager for different processes In UNIX environment process level data can be examined, for example, by using prstat a command (Figure 9): Process ID is reported in first column PID, columns SIZE and RSS specify memory usage and column CPU delivers CPU utilization. This command helps to associate process name (last column) and process ID. Figure 9 Process information in UNIX environment as reported by prstat a command If we have to log data on process resource utilization over some period of time, we can use the following command for a process named <process name>: while true; do ps -eo vsz,rss,pcpu,comm| grep <process name> >> log.txt; sleep 5; done This command collects the counters every five seconds: vsz - Total size of the process in virtual memory, in kilobytes.
  • 11. rss - Resident set size of the process, in kilobytes. pcpu - Percentage of CPU utilization. We mentioned a few basic ways of providing insight into resource consumption by each hosted application; operating systems and third party monitoring products offer a broad spectrum of means to collect process-level data to make confident decisions on cloud capacity management. Take Away from the Article 1. In non-virtualized cloud performance, bottlenecks created by a shortage of hardware resources affect all applications by increasing their transaction times. 2. Cloud-wide bottleneck can be caused by increase in workload of any hosted application, as well as by fluctuations in its service demand while processing large or complex data. 3. A contribution of each application to resources utilization can be found by monitoring software processes created by the operating system for a particular application. Queuing models provide estimates of such contribution. About the Author During the last fifteen years as an Oracle consultant, the author was engaged in hands- on performance tuning and sizing of enterprise applications for various corporations (Dell, Citibank, Verizon, Clorox, Bank of America, AT&T, Best Buy, Aetna, Halliburton, Pfizer, Astra Zeneca, Starbucks, etc).