PARSEC與SPLASH-2之工作特性分析

Abstract

在這篇論文中，我們首先試著去分析PARSEC與SPLASH-2這兩個常用的評測基準程式(benchmark suite)之工作特性。每個工作量(workload)都會在各種不同的模擬機器配置下進行實驗來研究他們多樣化的特性。此外，我們也試著啟動一個軟體預取機制(software prefetching scheme)來評估對這些不同工作量所造成的影響。根據這些實驗，我們發現到當執行這些工作量時，在互連網路上常常會有大量對共享資料做讀取動作的傳輸。 針對平行程式中大量的共享資料讀取失誤，我們提出一個稱為Snooping Fetch (SF)的方法。這個方法藉由系統中固有的快取一致性協定(cache coherence protocol)以及監聽協定(snooping protocol)所廣播的資訊來試著減少這些共享資料的讀取失誤。我們可以利用這個廣播的特性，在一筆資料真正要被使用之前先把資料讀進快取記憶體中以隱藏讀取時的延遲。在我們的實驗中，這個方法針對減少L1快取記憶體失誤量的部分可以得到平均8%的可能性，而在L1快取失誤率的減少量約有平均0.5%的可能性，以及平均0.4%的效能加速可能性。

In this thesis, we first characterize two popular benchmark suites, PARSEC and SPLASH-2. Different configurations of the simulated machine are set for each workload in the suites to study their diverse properties. Besides, we also try to evaluate the influence of different workloads when enabling a software prefetching scheme. Based on the experiments, we find that there is plenty of shared read traffic on the interconnection in these workloads. To reduce these shared read misses, a policy, called Snooping Fetch (SF), is proposed for the shared data in parallelized applications, which takes advantage of the inherent cache coherence protocol and the broadcasting information from the snooping protocol. The SF policy utilizes the broadcasting property to fetch shared data before the data is actually needed. In our experiments, the SF policy provides a reduction potential of 8% in average in the amounts of L1 cache misses, a reduction potential of 0.5% in average in the L1 cache miss rate, and a potential of 0.4% speedup in average.