Nt1310 Unit 3 Ggression Analysis

section{Evaluation} label{sec-analyze} vspace{-0.08in}

We evaluate Tarax with the six popular server applications described above. We first perform experiments to compare the performance and code sizes of the Tarax-optimized kernels and the vanilla kernel. We then perform dynamic profiling on the kernels to collect detailed statistics on instruction cache misses and branches. Finally, we switch on specific GCC optimizations with and without profile feedback, respectively, to collect performance numbers. We use these experiments to answer the following questions:

egin{itemize}

item What are the performance benefits of kernels optimized by Tarax in comparison to the vanilla kernel? Are the optimized kernels application-specific? (S

Figure~ ef{fig-cross} shows the result matrix, where all numbers shown are normalized to the application performance on its corresponding optimized kernel.

egin{figure} centering includegraphics[width=2.5in]{fig-cross.pdf} caption{Application performance on different optimized kernels. memph{For each application, the numbers are normalized to the performance when it runs on the kernel optimized for itself.}} label{fig-cross} end{figure} If the optimized kernel is best-suited for the target application, the numbers on the diagonal should be the highest; all other numbers should be below 1 since they are running on kernels optimized for other applications. We can see most of the results follow this pattern, with all applications except Memcached achieving the best performance on their corresponding optimized kernel.

For example, when we run Nginx on all six kernels, the performance of it running on other kernels ranges from 91\% to 97\% of the performance on its own kernel. On the other hand, although the performance of Memcached is generally good while running on other kernels, the performance of other applications running on the Memcached kernel could be as low as 93\% of their best performance.

The results show that we have created emph{truly} application-specific Linux kernels for

Running an application on an arbitrary (albeit) optimized kernel could degrade its performance by as much as 9\%.

Note that Apache and Nginx exhibit different behaviors on these kernels even though they are both web server applications, which indicates that it is sometimes difficult to build a uniformly good kernel even for a set of applications with similar functionalities.

%vspace{-0.05in} subsection{Sensitivity Analysis} label{perf-sensivity} %vspace{-0.05in}