More thoughts about performance

Performance is a set of quantifiable properties of a performer.

In this context, performance isn't the execution of an action itself; it's how well something or someone performs. Therefore, we use the adjective performant.

While the how well part can, in general, be described in natural languages, in our limited scope, the focus is on something that is quantifiable as a real number. Real numbers include integers and 0/1 binaries as special cases. Natural language descriptions are still very important. For example, a news article that heavily criticizes Flutter's performance by just using words without any numbers (a quantifiable value) could still be meaningful, and it could have great impacts. The limited scope is chosen only because of our limited resources.

The required quantity to describe performance is often referred to as a metric.

To navigate through countless performance issues and metrics, you can categorize based on performers.

For example, most of the content on this website is about the Flutter app performance, where the performer is a Flutter app. Infra performance is also important to Flutter, where the performers are build bots and CI task runners: they heavily affect how fast Flutter can incorporate code changes, to improve the app's performance.

Here, the scope was intentionally broadened to include performance issues other than just app performance issues because they can share many tools regardless of who the performers are. For example, Flutter app performance and infra performance might share the same dashboard and similar alert mechanisms.

Broadening the scope also allows performers to be included that traditionally are easy to ignore. Document performance is such an example. The performer could be an API doc of the SDK, and a metric could be: the percentage of readers who find the API doc useful.

Answering this question is not only crucial for validating the work in performance, but also for guiding the performance work in order to be more useful. The answer to "why is performance important?" often is also the answer to "how is performance useful?"

Simply speaking, performance is important and useful because, in the scope, performance must have quantifiable properties or metrics. This implies:

1. A performance report is easy to consume.
2. Performance has little ambiguity.
3. Performance is comparable and convertible.
4. Performance is fair.

Not that non-performance, or non-measurable issues or descriptions are not important. They're meant to highlight the scenarios where performance can be more useful.

Performance metrics are numbers. Reading a number is much easier than reading a passage. For example, it probably takes an engineer 1 second to consume the performance rating as a number from 1 to 5. It probably takes the same engineer at least 1 minute to read the full, 500-word feedback summary.

If there are many numbers, it's easy to summarize or visualize them for quick consumption. For example, you can quickly consume millions of numbers by looking at its histogram, average, quantiles, and so on. If a metric has a history of thousands of data points, then you can easily plot a timeline to read its trend.

On the other hand, having n number of 500-word texts almost guarantees an n-time cost to consume those texts. It would be a daunting task to analyze thousands of historical descriptions, each having 500 words.

Another advantage of having performance as a set of numbers is its unambiguity. When you want an animation to have a performance of 20 ms per frame or 50 fps, there's little room for different interpretations about the numbers. On the other hand, to describe the same animation in words, someone might call it good, while someone else might complain that it's bad. Similarly, the same word or phrase could be interpreted differently by different people. You might interpret an OK frame rate to be 60 fps, while someone else might interpret it to be 30 fps.

Numbers can still be noisy. For example, the measured time per frame might be a true computation time of this frame, plus a random amount of time (noise) that CPU/GPU spends on some unrelated work. Hence, the metric fluctuates. Nevertheless, there's no ambiguity of what the number means. And, there are also rigorous theory and testing tools to handle such noise. For example, you could take multiple measurements to estimate the distribution of a random variable, or you could take the average of many measurements to eliminate the noise by the law of large numbers.

Performance numbers not only have unambiguous meanings, but they also have unambiguous comparisons. For example, there's no doubt that 5 is greater than 4. On the other hand, it might be subjective to figure out whether excellent is better or worse than superb. Similarly, could you figure out whether epic is better than legendary? Actually, the phrase strongly exceeds expectations could be better than superb in someone's interpretation. It only becomes unambiguous and comparable after a definition that maps strongly exceeds expectations to 4 and superb to 5.

Numbers are also easily convertible using formulas and functions. For example, 60 fps can be converted to 16.67 ms per frame. A frame's rendering time x (ms) can be converted to a binary indicator isSmooth = [x <= 16] = (x <= 16 ? 1 :0). Such conversion can be compounded or chained, so you can get a large variety of quantities using a single measurement without any added noise or ambiguity. The converted quantity can then be used for further comparisons and consumption. Such conversions are almost impossible if you're dealing with natural languages.

If issues rely on verbose words to be discovered, then an unfair advantage is given to people who are more verbose (more willing to chat or write) or those who are closer to the development team, who have a larger bandwidth and lower cost for chatting or face-to-face meetings.

By having the same metrics to detect problems no matter how far away or how silent the users are, we can treat all issues fairly. That, in turn, allows us to focus on the right issues that have greater impact.

The following summarizes the 4 points discussed here, from a slightly different perspective:

1. Make performance metrics easy to consume. Do not overwhelm the readers with a lot of numbers (or words). If there are many numbers, then try to summarize them into a smaller set of numbers (for example, summarize many numbers into a single average number). Only notify readers when the numbers change significantly (for example, automatic alerts on spikes or regressions).

2. Make performance metrics as unambiguous as possible. Define the unit that the number is using. Precisely describe how the number is measured. Make the number easily reproducible. When there's a lot of noise, try to show the full distribution, or eliminate the noise as much as possible by aggregating many noisy measurements.

3. Make it easy to compare performance. For example, provide a timeline to compare the current version with the old version. Provide ways and tools to convert one metric to another. For example, if we can convert both memory increase and fps drops into the number of users dropped or revenue lost in dollars, then we can compare them and make an informed trade-off.

4. Make performance metrics monitor a population that is as wide as possible, so no one is left behind.

Unless stated otherwise, the documentation on this site reflects the latest stable version of Flutter. Page last updated on 2024-04-04. View source or report an issue.