Performance Metrics
Before we begin our journey into the world of .NET performance, we must understand the metrics and goals involved in performance testing and optimization. In Chapter 2, we explore more than a dozen profilers and monitoring tools; however, to use these tools, you need to know which performance metrics you are interested in.
Different types of applications have a multitude of varying performance goals, driven by business and operational needs. At times, the application’s architecture dictates the important performance metrics: for example, knowing that your Web server has to serve millions of concurrent users dictates a multi-server distributed system with caching and load balancing. At other times, performance measurement results may warrant changes in the application’s architecture: we have seen countless systems redesigned from the ground up after stress tests were run—or worse, the system failed in the production environment.
In our experience, knowing the system’s performance goals and the limits of its environment often guides you more than halfway through the process of improving its performance. Here are some examples we have been able to diagnose and fix over the last few years:
- We discovered a serious performance problem with a powerful Web server in a hosted data center caused by a shared low-latency 4Mbps link used by the test engineers. Not understanding the critical performance metric, the engineers wasted dozens of days tweaking the performance of the Web server, which was actually functioning perfectly.
- We were able to improve scrolling performance in a rich UI application by tuning the behavior of the CLR garbage collector—an apparently unrelated component. Precisely timing allocations and tweaking the GC flavor removed noticeable UI lags that annoyed users.
- We were able to improve compilation times ten-fold by moving hard disks to SATA ports to work around a bug in the Microsoft SCSI disk driver.
- We reduced the size of messages exchanged by a WCF service by 90 %, considerably improving its scalability and CPU utilization, by tuning WCF’s serialization mechanism.
- We reduced startup times from 35 seconds to 12 seconds for a large application with 300 assemblies on outdated hardware by compressing the application’s code and carefully disentangling some of its dependencies so that they were not required at load time.
These examples serve to illustrate that every kind of system, from low-power touch devices, high-end consumer workstations with powerful graphics, all the way through multi-server data centers, exhibits unique performance characteristics as countless subtle factors interact. In this chapter, we briefly explore the variety of performance metrics and goals in typical modern software. In the next chapter, we illustrate how these metrics can be measured accurately; the remainder of the book shows how they can be improved systematically.
Performance Goals
Performance goals depend on your application’s realm and architecture more than anything else. When you have finished gathering requirements, you should determine general performance goals. Depending on your software development process, you might need to adjust these goals as requirements change and new business and operation needs arise. We review some examples of performance goals and guidelines for several archetypal applications, but, as with anything performance-related, these guidelines need to be adapted to your software’s domain.
First, here are some examples of statements that are not good performance goals:
- The application will remain responsive when many users access the Shopping Cart screen simultaneously.
- The application will not use an unreasonable amount of memory as long as the number of users is reasonable.
- A single database server will serve queries quickly even when there are multiple, fully-loaded application servers.
The main problem with these statements is that they are overly general and subjective. If these are your performance goals, then you are bound to discover they are subject to interpretation and disagreements on their frame-of-reference. A business analyst may consider 100,000 concurrent users a “reasonable” number, whereas a technical team member may know the available hardware cannot support this number of users on a single machine. Conversely, a developer might consider 500 ms response times “responsive,” but a user interface expert may consider it laggy and unpolished.
A performance goal, then, is expressed in terms of quantifiable performance metrics that can be measured by some means of performance testing. The performance goal should also contain some information about its environment—general or specific to that performance goal. Some examples of well-specified performance goals include:
- The application will serve every page in the “Important” category within less than 300 ms (not including network roundtrip time), as long as not more than 5,000 users access the Shopping Cart screen concurrently.
- The application will use not more than 4 KB of memory for each idle user session.
- The database server’s CPU and disk utilization should not exceed 70%, and it should return responses to queries in the “Common” category within less than 75ms, as long as there are no more than 10 application servers accessing it.
Note These examples assume that the “Important” page category and “Common” query category are well-known terms defined by business analysts or application architects. Guaranteeing performance goals for every nook and cranny in the application is often unreasonable and is not worth the investment in development, hardware, and operational costs.
We now consider some examples of performance goals for typical applications (see Table 1-1 ). This list is by no means exhaustive and is not intended to be used as a checklist or template for your own performance goals—it is a general frame that establishes differences in performance goals when diverse application types are concerned.
Table 1-1. Examples of Performance Goals for Typical Applications
| System Type | Performance Goal | Environment Constraints |
|---|---|---|
| External Web Server | Time from request start to full response generated should not exceed 300ms | Not more than 300 concurrently active requests |
| External Web Server | Virtual memory usage (including cache) should not exceed 1.3GB | Not more than 300 concurrently active requests; not more than 5,000 connected user sessions |
| Application Server | CPU utilization should not exceed 75% | Not more than 1,000 concurrently active API requests |
| Application Server | Hard page fault rate should not exceed 2 hard page faults per second | Not more than 1,000 concurrently active API requests |
| Smart Client Application | Time from double-click on desktop shortcut to main screen showing list of employees should not exceed 1,500ms | -- |
| Smart Client Application | CPU utilization when the application is idle should not exceed 1% | -- |
| Web Page | Time for filtering and sorting the grid of incoming emails should not exceed 750ms, including shuffling animation | Not more than 200 incoming emails displayed on a single screen |
| Web Page | Memory utilization of cached JavaScript objects for the “chat with representative” windows should not exceed 2.5MB | -- |
| Monitoring Service | Time from failure event to alert generated and dispatched should not exceed 25ms | -- |
| Monitoring Service | Disk I/O operation rate when alerts are not actively generated should be 0 | -- |
Note Characteristics of the hardware on which the application runs are a crucial part of environment constraints. For example, the startup time constraint placed on the smart client application in Table 1-1
may require a solid-state hard drive or a rotating hard drive speed of at least 7200RPM, at least 2GB of system memory, and a 1.2GHz or faster processor with SSE3 instruction support. These environment constraints are not worth repeating for every performance goal, but they are worth remembering during performance testing.
When performance goals are well-defined, the performance testing, load testing, and subsequent optimization process is laid out trivially. Verifying conjectures, such as “with 1,000 concurrently executing API requests there are less than 2 hard page faults per second on the application server,” may often require access to load testing tools and a suitable hardware environment. The next chapter discusses measuring the application to determine whether it meets or exceeds its performance goals once such an environment is established.
Composing well-defined performance goals often requires prior familiarity with performance metrics, which we discuss next.
Unlike performance goals, performance metrics are not connected to a specific scenario or environment. A performance metric is a measurable numeric quantity that ref lects the application’s behavior. You can measure a performance metric on any hardware and in any environment, regardless of the number of active users, requests, or sessions. During the development lifecycle, you choose the metrics to measure and derive from them specific performance goals.
Some applications have performance metrics specific to their domain. We do not attempt to identify these metrics here. Instead, we list, in Table 1-2 , performance metrics often important to many applications, as well as the chapter in which optimization of these metrics is discussed. (The CPU utilization and execution time metrics are so important that they are discussed in every chapter of this book.)
Table 1-2. List of Performance Metrics (Partial)
| Performance Metric | Units of Measurement | Specific Chapter(s) in This Book |
|---|---|---|
| CPU Utilization | Percent | All Chapters |
| Physical/Virtual Memory Usage | Bytes, kilobytes, megabytes, gigabytes | Chapter 4 – Garbage Collection Chapter 5 – Collections and Generics |
| Cache Misses | Count, rate/second | Chapter 5 – Collections and Generics Chapter 6 – Concurrency and Parallelism |
| Page Faults | Count, rate/second | -- |
| Database Access Counts/Timing | Count, rate/second, milliseconds | -- |
| Allocations | Number of bytes, number of objects, rate/second | Chapter 3 – Type Internals Chapter 4 – Garbage Collection |
| Execution Time | Milliseconds | All Chapters |
| Network Operations | Count, rate/second | Chapter 7 – Networking, I/O, and Serialization Chapter 11 – Web Applications |
| Disk Operations | Count, rate/second | Chapter 7 – Networking, I/O, and Serialization |
| Response Time | Milliseconds | Chapter 11 – Web Applications |
| Garbage Collections | Count, rate/second, duration (milliseconds), % of total time | Chapter 4 – Garbage Collection |
| Exceptions Thrown | Count, rate/second | Chapter 10 – Performance Patterns |
| Startup Time | Milliseconds | Chapter 10 – Performance Patterns |
| Contentions | Count, rate/second | Chapter 6 – Concurrency and Parallelism |
Some metrics are more relevant to certain application types than others. For example, database access times are not a metric you can measure on a client system. Some common combinations of performance metrics and application types include:
- For client applications, you might focus on startup time, memory usage, and CPU utilization.
- For server applications hosting the system’s algorithms, you usually focus on CPU utilization, cache misses, contentions, allocations, and garbage collections.
- For Web applications, you typically measure memory usage, database access, network and disk operations, and response time.
A final observation about performance metrics is that the level at which they are measured can often be changed without significantly changing the metric’s meaning. For example, allocations and execution time can be measured at the system level, at the single process level, or even for individual methods and lines. Execution time within a specific method can be a more actionable performance metric than overall CPU utilization or execution time at the process level. Unfortunately, increasing the granularity of measurements often incurs a performance overhead, as we illustrate in the next chapter by discussing various profiling tools.
PERFORMANCE IN THE SOFTWARE DEVELOPMENT LIFECYCLE
Where do you fit performance in the software development lifecycle? This innocent question carries the baggage of having to retrofit performance into an existing process. Although it is possible, a healthier approach is to consider every step of the development lifecycle an opportunity to understand the application’s performance better: first, the performance goals and important metrics; next, whether the application meets or exceeds its goals; and finally, whether maintenance, user loads, and requirement changes introduce any regressions.
- During the requirements gathering phase, start thinking about the performance goals you would like to set.
- During the architecture phase, refine the performance metrics important for your application and define concrete performance goals.
- During the development phase, frequently perform exploratory performance testing on prototype code or partially complete features to verify you are well within the system’s performance goals.
- During the testing phase, perform significant load testing and performance testing to validate completely your system’s performance goals.
- During subsequent development and maintenance, perform additional load testing and performance testing with every release (preferably on a daily or weekly basis) to quickly identify any performance regressions introduced into the system.
Summary
This chapter served as an introduction to the world of performance metrics and goals. Making sure you know what to measure and what performance criteria are important to you can be even more important than actually measuring performance, which is the subject of the next chapter. Throughout the remainder of the book, we measure performance using a variety of tools and provide guidelines on how to improve and optimize applications.