DeSpErate: Speeding-up Design Space Exploration by using Predictive Simulation Scheduling 论文笔记

abstract

设计空间探索（DSE）阶段是用来调整配置系统的参数。通常是一个多目标优化（MOO）问题。

It is usually done at pre-design phase and consists of the evaluation of large design spaceswhere each configuration requires long simulation.

Several heuristic techniques have been proposed in the past and the recent trend isreducing the exploration time by using analytic

prediction models to approximate the system metrics, effectively pruning sub-optimal configurations from the exploration scope.

以前研究的不足之处：

However, there is still a missing path towards theeffective usage of the underlying computing resourcesused by the DSE process.

本文的主要研究工作：

In this work, we will show that an alternative and almost orthogonal approach — focused on exploiting the available parallelism in terms of computing resources — can be
used to better schedule the simulations and to obtain a high speedup with respect to state of the art approaches, without compromising the accuracy of exploration results.

1. INTRODUCTION

Due to the computational time required to carry out each simulation, the DSE process might become unreasonably long. To reduce the simulation time, a well-known solution is to use analytic models
to predict the simulation results [4]. Once the analytic models are trained, they can be used to prune the design space by focusing the exploration on the most
4000
promising design space regions [4].

研究的目的：

When relying on analytic models to prune the design space we shrink the number of simulations to be run in parallel

and we might reduce the benefits of a parallel simulation environment.

提出的新方法：预测运行模拟的执行时间而不是预测模拟的输出。

To tackle this problem, we suggest for the first time to change the perspective and topredict the execution time required to run a simulation rather than to predict the simulation output itself.

2. BACKGROUND

目前研究的不足：

So far, in the field of optimization of computing systems there has not yet been any comparison between the advantages of using a parallel computing environment and an analytic performance prediction model, nor their joint exploitation.

3.MOTIVATIONS



Our basic idea consists of scheduling more simulations (rather than to prune them). Additional simulations can be efficiently scheduled when some computing resources would be otherwise idle. To this end, we suggest to use approximate
models to predict the time required to execute the simulations (rather than predicting the quality of the architectural configurations).

运行reserve simulations,并不会花更多的时间，同时提升了模拟的吞吐量。因此提供了额外的信息，潜在的帮助找到更好的架构配置

4. THE PROPOSED METHODOLOGY



增强Markovianity-based Optimization Algorithm ( MOA )

MOA belongs to the class of optimization algorithms known as Estimation of Distribution Algorithms (EDAs) since it iteratively estimates the probability distribution of the optimal solutions.

At every iteration, we update a simulation time prediction model t̂. The first iteration proceeds as for the MOA algorithm by sampling n configurations uniformly distributed in the design space. The simulation time t(x) of each
configuration x is collected and used to fit the analytic prediction model t̂. This analytic function is an approximation

of the simulation time t̂(x) ∼ t(x).

In the following iterations, the probability distribution fitting the best m configurations is sampled twice to generate

two sets, i.e. P and R. The set P represents the set of n configurations to be simulated as for the MOA algorithm. The

set R represents a reserve list of candidate configurations to be simulated if computational power is available.

VI. C ONCLUSION

In this paper, we proposed a simulation scheduling technique to exploit a parallel simulation environment during the

optimization process. The proposed technique is based on the definition of an analytic model to predict the time required to execute the different simulations.

The underlying idea of the approach is that, additional simulations taken from a reserve list can

be scheduled when some computing resources are predicted to become idle.