CS5239 Chap.1 Performance Metrics and Terminology
Performance Metrics and Terminology
Why the vocabulary comes first
Every later tool in the course, from queueing models to the USE method and bottleneck analysis, assumes you can name what a number measures before you reason about it. The opening outcome for the course is explicit on this point: you should be able to explain and interpret latency, utilisation and saturation.
The second lecture extends that list to response time, throughput and bottleneck, and it frames all six around a practical idea.
A metric is chosen to inform a decision, and the decision tells you which word fits: delay felt by a user, work completed, pressure on a resource, or the factor that limits the system.
Time as the user sees it
Response time is the complete span of an operation, from the request being issued to the result arriving, so it bundles together queueing, service and the transfer of the result.
Latency is used in two ways. Its narrow meaning is the waiting portion before service starts; its broad meaning is the whole operation, in which case it is simply another name for response time. Analysts get into trouble when a tool reports the broad figure and a reader interprets it in the narrow sense, because the split between waiting and working is exactly what points to a cause.
Drawing one request on a time axis and labelling each phase is the fastest way to keep the two meanings apart.
Work done and how hard resources work
Throughput is a rate. Communications engineers usually express it as bytes or bits per second, database engineers as operations or transactions per second, and storage engineers as IOPS, the count of reads and writes per second.
Workload is the input that drives all of these figures, such as the queries and commands that clients send to a database. Utilisation describes the resource side. In its time-based form it is busy time divided by the length of the observation period, written U = B/T.
In its capacity-based form it compares delivered throughput with the throughput the component can sustain, and for storage it can simply mean the share of space consumed.
The two forms can disagree widely on one disk, since a stream of small or random requests can keep the device busy almost constantly while it moves only a fraction of its maximum.
From pressure to a limit
Saturation is what happens when demand outruns supply. It starts once capacity-based utilisation reaches 100%, and from that point additional requests have nowhere to go except a queue.
Queued work feeds straight back into the waiting phase of response time, which is why saturation is the usual explanation for latency that climbs without any change in service time. A bottleneck is the resource whose limit caps the performance of the whole system.
Finding it and removing it is one of the central jobs of a performance analyst, and the remaining methodology lectures build structured ways of doing so.
Two starting points for an investigation
The course separates a resource perspective from a workload perspective.
System administrators typically adopt the first, walking through processors, memory, storage devices, network cards and the buses that link them and asking how utilised each is; this suits capacity planning and issue investigation.
Application developers typically adopt the second, beginning with the requests applied, the latency they experience and whether they complete without error, and tracking throughput and latency as the headline numbers. Strong analyses move between both views, and the vocabulary in this chapter is what allows that movement without ambiguity.
What this chapter covers
- 01
Choosing a metric by the decision it informs
- 02
Response time as wait plus service plus transfer
- 03
Narrow and broad meanings of latency
- 04
Throughput, IOPS and workload
- 05
Time-based against capacity-based utilisation
- 06
Saturation, queueing and the bottleneck
- 07
Resource analysis and workload analysis perspectives
Worked example · free
Reading one window of a payment API
- 2Add waiting, service and transfer to obtain response time.
- 2Isolate the waiting phase as narrow-sense latency and give its share.
- 2Divide completed requests by the interval to obtain throughput.
- 2Compute busy time from service time and apply U = B/T, then interpret.
Key terms
- Latency
- A time measure that, in its strict sense, covers only the wait before an operation is serviced, and in looser usage the whole operation.
- Response Time
- The full time for an operation to complete, including waiting, service time and transfer of the result.
- Throughput
- The rate of work performed, such as bytes per second on a link or transactions per second in a database.
- IOPS
- A rate of input and output operations per second; for a disk it counts read and write operations each second.
- Workload
- The input or load applied to a system, such as the queries and commands clients send to a database.
- Utilisation
- The share of time or capacity a resource spends working, either busy time over an interval (U = B/T) or the share of deliverable capacity in use.
- Saturation
- The backlog of requests a resource cannot serve yet, beginning at 100% capacity-based utilisation.
- Bottleneck
- The resource that limits the performance of the whole system and is the target of bottleneck analysis.
Performance Metrics and Terminology FAQ
Is latency the same thing as response time?
Sometimes. In its strict meaning latency is only the time an operation spends waiting before service starts, while response time adds the service itself and the transfer of the result. Some tools and contexts use latency for the whole operation, so always check which sense a reported figure uses before comparing it with another.
Why can a disk show 100% utilisation and still have spare throughput?
Time-based utilisation only asks whether the device was busy, not how much work it moved. A stream of small or random requests can keep a disk occupied for the entire interval while it delivers a modest fraction of the IOPS it could sustain, so the capacity-based reading can be far lower than the time-based one.
When does saturation start?
Saturation begins once capacity-based utilisation reaches 100%, the point at which more work is being requested than the resource can process. Extra requests then wait in a queue, and that queued work shows up as a longer waiting phase inside response time even when service time per request has not changed.
What is the difference between resource analysis and workload analysis?
Resource analysis starts from hardware such as processors, memory, storage devices, network cards and the buses that link them, and centres on utilisation; system administrators use it for investigations and capacity planning. Workload analysis starts from the requests an application receives, tracking latency, completion and errors, and is the natural view for application developers.
How is throughput measured in practice?
It depends on the component. Network links usually report bytes or bits per second, databases report operations or transactions per second, and storage devices report IOPS. Whatever the unit, it is completed work divided by the time over which it was completed, so state the interval alongside the number.
Assessment move
Take any monitoring screenshot or log you can find and label every number with one of the six terms plus the decision it informs. Then redraw one request as wait, service and transfer, and compute both forms of utilisation for one device so the difference between them becomes automatic.
Working through Performance Metrics and Terminology in CS5239? Sia is AskSia’s AI Computer Science tutor — ask any CS5239 Performance Metrics and Terminology question and get a clear, step-by-step explanation grounded in how CS5239 is taught and assessed. Read this chapter free, then take your hardest questions to Sia.
