mARGOt: A Dynamic Autotuning Framework for Self-Aware Approximate Computing

Abstract

In the autonomic computing context, the system is perceived as a set of autonomous elements capable of self-management, where end-users define high-level goals and the system shall adapt to achieve the desired behaviour. Runtime adaptation creates several optimization opportunities, especially if we consider approximate computing applications, where it is possible to trade off the accuracy of the result and the performance. Given that modern systems are limited by the power dissipated, autonomic computing is an appealing approach to increase the computation efficiency. In this paper, we introduce mARGOt, a dynamic autotuning framework to enhance the target application with an adaptation layer to provide self-optimization capabilities. The framework is implemented as a C++ library that works at function-level and provides to the application a mechanism to adapt in a reactive and a proactive way. Moreover, the application is capable to change dynamically its requirements and to learn online the underlying application-knowledge. We evaluated the proposed framework in three real-life scenarios, ranging from embedded to HPC applications. In the three use cases, experimental results demonstrate how, thanks to mARGOt, it is possible to increase the computation efficiency by adapting the application at runtime with a limited overhead.

Tunable approximations to control time-to-solution in an HPC molecular docking Mini-App

“…Figure 2 shows the Call Graph report: It groups the individual functions by the caller. 19 set the right fragment to best angle found; 20 end 21 return the overlap score of the ligand; Algorithm 1: Pseudo-code of the MatchProbesShape kernel, which changes the shape of the ligand to maximize the overlap score.…”

“…In this section, we explain how tuning can be automatically done by the application itself according to an execution time budget allocated by the end-user. We used the mARGOt [19] framework to select the configuration of the software knobs that maximizes the accuracy of the result given the time budget.…”

Dynamic Application Autotuning for Self-aware Approximate Computing

“…The application is able to change the application requirements at runtime according to the system evolution. Moreover, if the EFPs are input-dependent, the application can provide features of the actual input to adapt proactively [9].…”

“…Figure 7.4 shows the distributions of the prediction error on the time-to-solution with six pocket from the RCSB Protein Databank (PDB) [32]. We use a chemical library with heterogeneous ligands [9]. For example, the number of their atoms range from 28 to 153.…”

“…The decision process is based on the application knowledge, which describes the relationship between software-knob configurations and EFPs. It is possible to learn the application knowledge either at design-time, by using well-known Design Space Exploration techniques [8], or at runtime by using an external component that coordinates a distributed DSE [9]. The benefits of the latter approach are the following: (1) the application can observe its behaviour with the same execution environment of the production run, (2) it can leverage features of the production input, and in the context of a distributed computation, (3) it can leverage the number of application instances to lower the learning time.…”