Core Concepts

Let us start with an example from manufacturing by observing an assembly line in a factory. The line is assembling bikes from ready-made components. Bikes are built on this line in a sequential manner. Monitoring the line in action, we measure that it takes four hours to complete a bike from the moment the frame enters the line until the assembled bike leaves the line at the other end.

assembly line

Continuing our observations we can also see that one bike is assembled after each minute, 24 hours a day, every day. Simplifying the example and ignoring maintenance windows, we can forecast that in any given hour such an assembly line assembles 60 bikes.

Equipped with these two measurements, we now possess crucial information about the current performance of the assembly line in regards of latency and throughput:

  • Latency of the assembly line: 4 hours
  • Throughput of the assembly line: 60 bikes/hour

Notice that latency is measured in time units suitable for the task at hand – anything from nanoseconds to millennia can be a good choice. Throughput of a system is measured in completed operations per time unit. Operations can be anything relevant to the specific system. In this example the chosen time unit was hours and operations were the assembled bikes.

Having been equipped with the definitions of latency and throughput, let us carry out a performance tuning exercise in the very same factory. The demand for the bikes has been steady for a while and the assembly line has kept producing bikes with the latency of four hours and the throughput of 60 bikes/hour for months. Let us now imagine a situation where the sales team has been successful and the demand for the bikes suddenly doubles. Instead of the usual 60*24 = 1,440 bikes/day the customers demand twice as many. The performance of the factory is no longer satisfactory and something needs to be done.

The factory manager seemingly correctly concludes that the latency of the system is not a concern – instead he should focus on the total number of bikes produced per day. Coming to this conclusion and assuming he is well funded, the hypothetical manager would immediately take the necessary steps to improve throughput by adding capacity.

As a result we would now be observing not one but two identical assembly lines in the same factory. Both of these assembly lines would be assembling the very same bike every minute of every day. By doing this, our imaginary factory has doubled the number of bikes produced per day. Instead of the 1,440 bikes the factory is now capable of shipping 2,880 bikes each day. It is important to note that we have not reduced the time to complete an individual bike by even a millisecond – it still takes four hours to complete a bike from start to finish.

assembly line

In the example above a performance optimization task was carried out, coincidentally impacting both throughput and capacity. As in any good example we started by measuring the system’s current performance, then set a new target and optimized the system only in the aspects required to meet the target.

In this example an important decision was made – the focus was on increasing throughput, not on reducing latency. While increasing the throughput, we also needed to increase the capacity of the system. Instead of a single assembly line we now needed two assembly lines to produce the required quantity. So in this case the added throughput was not free, the solution needed to be scaled out in order to meet the increased throughput requirement.

An important alternative should also be considered for the performance problem at hand. The seemingly non-related latency of the system actually hides a different solution to the problem. If the latency of the assembly line could have been reduced from 1 minute to 30 seconds, the very same increase of throughput would suddenly be possible without any additional capacity.

Whether or not reducing latency was possible or economical in this case is not relevant. What is important is a concept very similar to software engineering – you can almost always choose between two solutions to a performance problem. You can either throw more hardware towards the problem or spend time improving the poorly performing code.


Latency goals for the GC have to be derived from generic latency requirements. Generic latency requirements are typically expressed in a form similar to the following:

  • All user transactions must respond in less than 10 seconds
  • 90% of the invoice payments must be carried out in under 3 seconds
  • Recommended products must be rendered to a purchase screen in less than 100 ms

When facing performance goals similar to the above, we would need to make sure that the duration of GC pauses during the transaction does not contribute too much to violating the requirements. “Too much” is application-specific and needs to take into account other factors contributing to latency including round-trips to external data sources, lock contention issues and other safe points among these.

Let us assume our performance requirements state that 90% of the transactions to the application need to complete under 1,000 ms and no transaction can exceed 10,000 ms. Out of those generic latency requirements let us again assume that GC pauses cannot contribute more than 10%. From this, we can conclude that 90% of GC pauses have to complete under 100 ms, and no GC pause can exceed 1,000 ms. For simplicity’s sake let us ignore in this example multiple pauses that can occur during the same transaction.

Having formalized the requirement, the next step is to measure pause durations. There are many tools for the job, covered in more detail in the chapter on Tooling, but in this section, let us use GC logs, namely for the duration of GC pauses. The information required is present in different log snippets so let us take a look which parts of date/time data are actually relevant, using the following example:

2015-06-04T13:34:16.974-0200: 2.578: [Full GC (Ergonomics) [PSYoungGen: 93677K->70109K(254976K)] [ParOldGen: 499597K->511230K(761856K)] 593275K->581339K(1016832K), [Metaspace: 2936K->2936K(1056768K)], 0.0713174 secs] [Times: user=0.21 sys=0.02, real=0.07 secs

The example above expresses a single GC pause triggered at 13:34:16 on June 4, 2015, just 2,578 ms after the JVM was started.

The event stopped the application threads for 0.0713174 seconds. Even though it took 210 ms of CPU times on multiple cores, the important number for us to measure is the total stop time for application threads, which in this case, where parallel GC was used on a multi-core machine, is equal to a bit more than 70 ms. This specific GC pause is thus well under the required 100 ms threshold and fulfils both requirements.

Extracting information similar to the example above from all GC pauses, we can aggregate the numbers and see whether or not we are violating the set requirements for any of the pause events triggered.


Throughput requirements are different from latency requirements. The only similarity that the throughput requirements share with latency is the fact that again, these requirements need to be derived from generic throughput requirements. Generic requirements for throughput can be similar to the following:

  • The solution must be able to process 1,000,000 invoices/day
  • The solution must support 1,000 authenticated users each invoking one of the functions A, B or C every five to ten seconds
  • Weekly statistics for all customers have to be composed in no more than six hours each Sunday night between 12 PM and 6 AM

So, instead of setting requirements for a single operation, the requirements for throughput specify how many operations the system must process in a given time unit. Similar to the latency requirements, the GC tuning part now requires determining the total time that can be spent on GC during the time measured. How much is tolerable for the particular system is again application-specific, but as a rule of thumb, anything over 10% would look suspicious.

Let us now assume that the requirement at hand foresees that the system processes 1,000 transactions per minute. Let us also assume that the total duration of GC pauses during any minute cannot exceed six seconds (or 10%) of this time.

Having formalized the requirements, the next step would be to harvest the information we need. The source to be used in the example is again GC logs, from which we would get information similar to the following:

2015-06-04T13:34:16.974-0200: 2.578: [Full GC (Ergonomics) [PSYoungGen: 93677K->70109K(254976K)] [ParOldGen: 499597K->511230K(761856K)] 593275K->581339K(1016832K), [Metaspace: 2936K->2936K(1056768K)], 0.0713174 secs] [Times: user=0.21 sys=0.02, real=0.07 secs

This time we are interested in user and system times instead of real time. In this case we should focus on 23 milliseconds (21 + 2 ms in user and system times) during which the particular GC pause kept CPUs busy. Even more important is the fact that the system was running on a multi-core machine, translating to the actual stop-the-world pause of 0.0713174 seconds, which is the number to be used in the following calculations.

Extracting the information similar to the above from the GC logs across the test period, all that is left to be done is to verify the total duration of the stop-the-world pauses during each minute. In case the total duration of the pauses does not exceed 6,000ms or six seconds in any of these one-minute periods, we have fulfilled our requirement.


Capacity requirements put additional constraints on the environment where the throughput and latency goals can be met. These requirements might be expressed either in terms of computing resources or in cold hard cash. The ways in which such requirements can be described can, for example, take the following form:

  • The system must be deployed on Android devices with less than 512 MB of memory
  • The system must be deployed on Amazon EC2 The maximum required instance size must not exceed the configuration c3.xlarge (8 G, 4 cores)
  • The monthly invoice from Amazon EC2 for running the system must not exceed $12,000

Thus, capacity has to be taken into account when fulfilling the latency and throughput requirements. With unlimited computing power, any kind of latency and throughput targets could be met, yet in the real world the budget and other constraints have a tendency to set limits on the resources one can use.