In the last post I covered several architectural techniques you can use to build a highly scalable, failure resistant application in the cloud. However, these architectural changes – along with the inherent unreliability of the cloud – introduce some new problems for application performance management. Many organizations rely on logging, profilers, and legacy application performance monitoring (APM) solutions to monitor and manage performance in the data center, but these strategies and solutions simply aren’t enough when you move into the cloud. Here are a few important considerations for choosing an APM solution that works in the cloud.
Many monitoring solutions check for server availability and alert users when a server goes down. In the cloud, however, servers can come and go all the time, so alerting on availability will result in a lot of false positives. In addition, many of the server-level metrics that APM tools and server monitoring tools report are no longer as relevant as they were on a vertically scaled system. For example, what does 90% CPU utilization mean to the behavior of your cloud application? Does it mean there is an impending performance problem that needs to be addressed? Or does it mean that more servers need to be added into that tier? This goes for other metrics, too, like physical memory usage, JVM memory usage, thread usage, database connection pool usage, and so on. These are all good indicators of the performance of a single server, but when servers can come and go they’re no longer the best approximation of the performance of your application as a whole.
Instead, it’s best to understand performance in terms of Business Transactions. A business transaction is essentially a user request – for an eCommerce application, “Check out” or “Add to Cart” may be two important business transactions. Each business transaction includes all of the downstream activities until the end user receives a response (and perhaps more, if your application uses asynchronous communication). For example, an application may define a service that performs request validation, stores data in a database, and then publishes a request to a topic. A JMS listener might receive that message from the topic, make a call to an external service, and then store the data in a Hadoop cluster. All of these activities need to be grouped together into a single Business Transaction so that you can understand how every part of your system affects your end users.
With these various tiers tracked at the Business Transaction level, the next step is to measure performance at the tier level. While it is important to know when a Business Transaction is behaving abnormally, it is equally as important to detect performance anomalies at the tier level. If the response time of a Business Transaction, as a whole, is slow by one standard deviation (which is acceptable) but one of its tiers is slower by a factor of three standard deviations, you may have a problem developing, even though it hasn’t affected your end users yet. Chances are the tier’s problem will evolve into a systemic problem that causes multiple Business Transactions to suffer.
Returning to our example from figure 3, let’s say the web service behaves well, but the topic listener is significantly slower than usual. The topic listener has not caused a problem in the Business Transaction itself, but it has slowed down enough to cause concern, so there might be an issue that needs to be addressed. Business Transactions, therefore, need to be evaluated both as a whole and at the tier level in order to identify performance issues before they arise. The only way to effectively monitor the performance of an application in a dynamic environment is to capture metrics at the Business Transaction level and the tier level.
One of the most important reasons that many organizations move to the cloud is to be able to scale applications up and down rapidly as load changes. If the load on your application fluctuates dramatically over the day, week or year, the cloud will allow you to scale your application infrastructure efficiently to meet that load. However, most application monitoring tools are not equipped to handle such dramatic shifts in load or performance. Application monitoring tools that rely on static thresholds for alerting and data collection will create alert storms when load increases and miss potential problems when it decreases. You need to be able to understand what normal performance is for a given time of day, day of the week or time of the year, which is best done by baselining the performance of your application over time.
Baselining your application essentially means collecting data around how your application performs (or how a specific Business Transactions performs) at any given time. Having this data will allow you (or your APM solution) to determine if how your application is performing now is normal or if it might indicate a problem. Baselines can be defined on a per-hour basis over a period of time – for example, for the past 30 days, how has Checkout performed from 9:00am to 10:00am? In this configuration, the response time of a specific Business Transaction will be compared to the average response time for that Business Transaction over the past 30 days, between the hours of 9:00am and 10:00am. If the response time is greater than some measurable value, such as two standard deviations, then the monitoring system should raise an alert. Figure 4 attempts to show this graphically.
The average response time for this Business Transaction is about 1.75 seconds, with two standard deviations being between 1.5 seconds and 2 seconds, captured over the past 30 days. All incoming occurrences of this Business Transaction during this hour (9:00am to 10:00am in this example) will be compared to the average of 1.75 seconds, and if the response time exceeds two standard deviations from this normal (2 seconds), then an alert will be raised.
What happens if the behavior of your users differs from day to day or month to month? Your monitoring solution should be configurable enough to handle this. Banking applications probably have spikes in load twice a month when most people get paid, and eCommerce applications are inundated on Black Friday. By baselining the performance of these applications over the year, an APM tool could anticipate this load and expect slower performance during these times. Make sure your APM tool is configurable or intelligent enough that it can understand what’s “normal” behavior for your app.
Dynamic Application Mapping
Many monitoring solutions today require manual configuration to instrument and monitor a new server. If new servers are disappearing and appearing all the time, however, this will result in blind spots as you update the tool to reflect the new environment. This will quickly become untenable as your application scales. A cloud-ready monitoring tool must automatically detect and map the application in real time, so you always have an up-to-date idea of what your application looks like. For agent-based monitoring solutions, this can be accomplished by deploying your agent along with your application so that new nodes are automatically instrumented by your APM solution of choice.
Take five minutes to get complete visibility into the performance of your cloud applications with AppDynamics today.