Performance Engineering

Introduction

HPC is about high application performance requirements. There are a lot of options to improve the performance of an application code, this also includes algorithmic optimisations. In the following it is assumed that a given algorithm is executed on a given HPC system.

The following factors influence performance:

Implementation of the algorithm (programming language, optimisation techniques)
Compiler and compiler options
Machine and operating system configuration
Runtime setup (pinning and resource allocation)

Basic principles of application performance

Strategies to speedup execution

There are three elementary ways to speed up execution on a computer:

Increase the execution rate (means increase clock)
Introduce parallel (concurrent) execution
Specialisation of execution resources

In computers all three strategies are used, with parallel execution being the most important one.

Illustrating example: Consider the high level task of drilling 20 holes into a wall. At the beginning a single worker drills one hole after the other. The first optimisation would be to hire a worker, who drills every hole quicker. That's increasing the execution rate. Next, one could hire several workers, for example 4, which drill holes concurrently. Ideally the work should then be finished in a quarter of the time. Finally you could buy a special purpose drilling machine which can drill five holes in one step. Now one worker can accomplish five times as much work in the same time. Of course you can combine all of this: Hire multiple fast workers and equip them with the special purpose drilling machine.

Mapping of work on compute resources

In order to understand the following parts, it is important to introduce the relation between work, execution resources and time to solution. The application requires work to accomplish a high level task. What this work actually is, is specific to the application. It may be requests answered, voxels processed or lattice points updated. To accomplish this work, the application uses the resources offered by the computer. A computer's only notion of work are instructions and, triggered by instructions, data transfers from entities that can store data to registers. This fact complicates performance optimisation as execution efficiency always happens on the instruction level. The application work has to be mapped on instruction work and this mapping might add overhead not related to the initial application work. The resulting instruction and data transfer work can then be mapped on compute resources. The mapping finally results in a time to solution.

Runtime contributions and critical path

Everything happening during program execution takes time. The time some activity or task takes is called a runtime contribution. In the simplest case, runtime contributions form the total time to solution by simply being summed up. On modern computers, however, many things happen concurrently where different runtime contributions can overlap with each other. The critical path is the series of runtime contributions, which cannot overlap with each other, and form the total runtime. A runtime contribution that adds to the total execution time is said to be on the critical path. Anything that takes time is only relevant for optimisation if it appears on the critical path.

Parallel execution

Parallelism happens on many levels on computers. For performance engineering one has to separate what happens within one instruction stream (executed on one processor core), and effects introduced by solving a problem using multiple of such instruction streams.

Parallel execution implies additional constraints and overheads. Work must consist of tasks that can be executed concurrently. In simple cases, tasks are embarrassingly parallel, but in most relevant applications there are dependencies between tasks, which require synchronisation or communication of data. Often in order to exploit parallelism additional work has to be introduced to efficiently implement it.

The following factors limit the achievable speedup by parallel execution:

Load balancing: Can work be perfectly distributed on parallel execution resources?
Dependencies: Exist dependencies that add to the critical path?
Communication overhead: Is there any data transfer time that adds to the critical path?
Instruction overhead: Additional work that is required to implement an efficient parallel execution.

From the standpoint of a single instruction stream all those influences but the last one (which adds to the processor work) are experienced as additional waiting times which may or may not add to the critical path during execution.

Basic iterative procedure for performance engineering

The following steps are required for a minimum performance engineering process:

Define a relevant test case which reflects production behaviour
Acquire a runtime profile to determine on which parts of the code the runtime time is spent
For all code parts (hot spots) of the runtime profile perform:
- Static code analysis
- Application benchmarking
- Parallel case: Create and analyse runtime traces
- Perform hardware performance counter profiling
Based on the data acquired by above activities narrow down performance issues
Optional: Formulate an analytic performance model
Improve performance by changing runtime configuration or implementation

Those steps need to be repeated multiple times for every hot spot until a required or good enough performance is reached. After applying an optimisation, it must be ensured that the optimised variants are used and also taking effect in regular production.

To carry out above procedure, multiple special skills beyond standard software engineering are required:

Strategies for performance analysis

After definition of a benchmark case, application benchmarking and performance profiling, the interpretation and analysis of the results is the first difficult task in any performance engineering effort. While there is no silver bullet for performance analysis, multiple strategies provide guidelines for different levels of expertise. It must be noted that in complicated cases the software developer carrying out the process must possess a certain level of experience to succeed. Therefore it is recommended to consult an experienced HPC consultant in the local HPC center if no progress is achieved using the simpler approaches.

Three approaches are described in more detail:

ProPE PE Process: Threshold-based performance analysis process based on the proven EU COE [POP project] approach for a rough initial performance analysis suited also for beginners
Performance Patterns: Performance-pattern-based process for more complicated cases targeted at experienced software developers
Performance Pattern List: List of performance patterns