For questions pertaining to the measurement or improvement of code and application efficiency.
The performance of applications is often a paramount concern for mission-critical systems. If your question pertains to optimization, whether it be database queries, algorithms, reducing network/transactional overhead, resource contention, or anything that deals with speed or capacity, consider using this tag.
A good question states performance goals that need to be achieved as well as other restrictions. Trying to optimize something without measuring is not a "performance" question or work, but most likely personal entertainment - expect a question without goals/measurements to be treated as such.
Performance for many programs is represented in big O notation, which classifies how an algorithm's resource requirements change in response to a change in the input size.
This tag can also represent system performance, which is one of the key non-functional requirements of an application or system.
The two main measures of performance are
- Throughput (how many in a time frame). Example of units: transactions per second (TPS), megabytes per second (MB/s), gigabits per second (Gb/s), messages/request/pages per second.
- Latency (how long for an action). For example, seek time of 8 ms and search time of 100 ms.
Latency is often qualified with a statistical measure. Note: latencies usually don't follow a normal distribution and have very high upper limits compared with the average latency. As such the standard deviation is not useful.
- Average latency. The average of all the latencies.
- Typical or median latency. The mid-point of the range of possible latencies. This is usually 50% to 90% of the average latency. As this is the lowest figure it is often reported by vendors.
- Percentile latency. The figure which it is less than or equal to N% of the time. That is, 99 percentile if the latency is not more than this, 99 times out of 100.
- Worst or maximum latency. The highest latency measured.
When seeking to improve performance: prototype and measure first, optimize only if and where needed.
See also: optimization profiling assembly compiler low-latency benchmarking
