Addressing concerns in performance prediction : the impact of data dependencies and denormal arithmetic in scientific codes

To meet the increasing computational requirements of the scientific community, the use of parallel programming has become commonplace, and in recent years distributed applications running on clusters of computers have become the norm. Both parallel and distributed applications face the problem of predictive uncertainty and variations in runtime. Modern scientific applications have varying I/O, cache, and memory profiles that have significant and difficult to predict effects on their runtimes. Data-dependent sensitivities such as the costs of denormal floating point calculations introduce more variations in runtime, further hindering predictability. Applications with unpredictable performance or which have highly variable runtimes can cause several problems. If the runtime of an application is unknown or varies widely, workflow schedulers cannot e�ciently allocate them to compute nodes, leading to the under-utilisation of expensive resources. Similarly, a lack of accurate knowledge of the performance of an application on new hardware can lead to misguided procurement decisions. In heavily parallel applications, minor variations in runtime on individual nodes can have disproportionate effects on the overall application runtime. Even on a smaller scale, a lack of certainty about an application’s runtime can preclude its use in real-time or time-critical applications such as clinical diagnosis. This thesis investigates two sources of data-dependent performance variability. The first source is algorithmic and is seen in a state-of-the-art C++ biomedical imaging application. It identifies the cause of the variability in the application and develops a means of characterising the variability. This ‘probe task’ based model is adapted for use with a workflow scheduler, and the scheduling improvements it brings are examined. The second source of variability is more subtle as it is micro-architectural in nature. Depending on the input data, two runs of an application executing exactly the same sequence of instructions and with exactly the same memory access patterns can have large differences in runtime due to deficiencies in common hardware implementations of denormal arithmetic1. An exception-based profiler is written to detect occurrences of denormal arithmetic and it is shown how this is insufficient to isolate the sources of denormal arithmetic in an application. A novel tool based on theValgrind binary instrumentation framework is developed which can trace the origins of denormal values and the frequency of their occurrence in an application’s data structures. This second tool is used to isolate and remove the cause of denormal arithmetic both from a simple numerical code, and then from a face recognition application.EThOS - Electronic Theses Online ServiceGBUnited Kingdo