Static Compiler Analyses for Application-specific Optimization of Task-Parallel Runtime Systems

Abstract

Achieving high performance in task-parallel runtime systems, especially with high degrees of parallelism and fine-grained tasks, requires tuning a large variety of behavioral parameters according to program characteristics. In the current state of the art, this tuning is generally performed in one of two ways: either by a group of experts who derive a single setup which achieves good-but not optimal-performance across a wide variety of use cases, or by monitoring a system's behavior at runtime and responding to it. The former approach invariably fails to achieve optimal performance for programs with highly distinct execution patterns, while the latter induces overhead and cannot affect parameters which need to be set at compile time. In order to mitigate these drawbacks, we propose a set of novel static compiler analyses specifically designed to determine program features which affect the optimal settings for a task-parallel execution environment. These features include the parallel structure of task spawning, the granularity of individual tasks, the memory size of the closure required for task parameters, and an estimate of the stack size required per task. Based on the result of these analyses, various runtime system parameters are then tuned at compile time. We have implemented this approach in the Insieme compiler and runtime system, and evaluated its effectiveness on a set of 12 task parallel benchmarks running with 1 to 64 hardware threads. Across this entire space of use cases, our implementation achieves a geometric mean performance improvement of 39%. To illustrate the impact of our optimizations, we also provide a comparison to current state-of-the art task-parallel runtime systems, including OpenMP, Cilk, HPX, and Intel TBB.

A Novel Prediction Model for Compiler Optimization with Hybrid Meta-Heuristic Optimization Algorithm

“…Peter et al [3] offered a series of new static compiler studies aimed at identifying programme properties that influenced the best settings for a task-parallel runtime environment. The parallel configuration of job spawn, the precision of specific activities, the memory capacity of closures needed for task variables and an estimation of the stack dimension necessary each task were all examples of such aspects.…”

Towards a Novel Framework for Automatic Big Data Detection

“…As it can be observed via Table 4, several works predict execution time [13]- [15], [17], [31], [33], [63], while others estimate memory consumption [16], [18], [29]. It can be observed that in [13]- [18], [63], the prediction is done for big data workloads, whereas in [29], [31], [33], the considered applications are routine ones. In [14], a performance prediction framework Ernest is proposed, which can predict the execution time on a hardware configuration, given a job and its input.…”

“…Moreover, our instrumentation technique is simpler as compared to [31], [33], as only the reachable functions and basic blocks list is generated. For predicting memory consumption, [29] estimates the stack frame size of a given task using static compiler analysis similar to our approach. Table 4, no existing work is found for estimating the heterogeneity ratio.…”

“…Additionally, the prediction models proposed in [13]- [18], [63] work for a specific platform like Spark and MapReduce. Whereas, similar to [29], [31], [33], the proposed prediction model is generic because it is implemented at the compiler stage. The proposed model can work with any computational engine, because it extracts core application features via compiler Intermediate Representation (IR), instead of using abstract application properties present in specialized implementations.…”

“…Additionally, the proposed model has been tested with different types of datasets like text, graph, and image. Whereas, mostly existing solutions [13], [14], [18], [29], [31], [33] are designed for text datasets only, [17] for image dataset, few [15], [16] for text and graph both, while only [63] considers graph, image, and text. It can be observed via Table 4, the metrics predictions in existing works have been done in a different environment, using different techniques and workloads, hence a direct comparison of these with proposed work is not feasible.…”