source: docs/HPCA2012/06-scalability.tex @ 1370

Last change on this file since 1370 was 1370, checked in by lindanl, 8 years ago

modifications

File size: 3.1 KB
Line 
1\section{Scalability}
2\label{section:scalability}
3\subsection{Performance}
4Figure \ref{Scalability} (a) demonstrates the average XML
5well-formedness checking performance of Parabix2 for each of the
6workloads and as executed on each of the processor cores --- \CO\,
7\CITHREE\ and \SB{}.  Processing time is shown in terms of bit stream
8based operations executed in `bit-space' and postprocessing operations
9executed in `byte-space'.  In the Parabix2 parser, bit-space parallel
10bit stream parser operations consist primarily of SIMD instructions;
11byte-space operations consist of byte comparisons across arrays of
12values. Executing Parabix2 on \CITHREE{} over \CO\ results in an
13average performance improvement of 17\% in bit stream processing
14whereas migrating Parabix2 from \CITHREE{} to \SB{} results in a 22\%
15average performance gain. Bit space measurements are stable and
16consistent across each of the source inputs and cores. Postprocessing
17operations demonstrate data dependent variance. Performance gains from
1818\% to 31\% performance are observered in migrating Parabix2 from
19\CO\ to \CITHREE{}; 0\% to 17\% performance from \CITHREE\ to
20\SB{}. For the purpose of comparison, Figure \ref{Scalability} (b)
21shows the performance of the Expat parser on each of the processor
22cores.  A performance improvement of less than 5\% is observed when
23executing Expat on \CITHREE\ over \CO\ and less than 10\% on \SB\ over
24\CITHREE{}.
25
26Overall, Parabix2 scales better than Expat. Simply executing identical
27Parabix2 object code on \SB\ results in an overall performance
28improvement up to 26\%. Additional performance aspects of Parabix2 on
29\SB\ with AVX instructions are discussed in the following sections.
30
31\begin{figure}
32\centering
33\subfigure[Parabix2]{
34\includegraphics[width=0.40\textwidth]{plots/P2_scalability.pdf}
35}
36\subfigure[Expat]{
37\includegraphics[width=0.40\textwidth]{plots/Expat_scalability.pdf}
38}
39\caption{Average Performance Parabix vs. Expat (y-axis: ns per kB)}
40\label{Scalability}
41\end{figure}
42
43
44\subsection{Power and Energy}
45
46Figure \ref{power_Parabix2} shows the average power consumption of
47Parabix2 over each workload and as executed on each of the processor
48cores --- \CO{}, \CITHREE\ and \SB{}.  Average power consumption on
49\CO{} is 32 watts. Execution on \CITHREE\ results in 30\% power saving
50over \CO{}\SB\ saves 25\% of the power compared with \CITHREE\ and
51consumes only 15 watts.
52
53In XML parsing we observe energy consumption is dependent on processing time. That is, a reduction in processing time results in a directly proportional reduction in energy consumption.
54With newer processor cores comes improvements in application performance. As a result, Parabix2 executed on \SB\ consumes 72\% to 75\% less energy than Parabix2 on \CO{}.
55
56
57
58
59\begin{figure}
60\centering
61\subfigure[Avg. Power of Parabix on various hardware (Watts)]{
62\includegraphics[width=85mm]{plots/power_Parabix2.pdf}
63\label{power_Parabix2}
64}
65\hfill
66\centering
67\subfigure[Avg. Energy Consumption on various hardware (nJ per kB)]{
68\includegraphics[width=85mm]{plots/energy_Parabix2.pdf}
69\label{energy_Parabix2}
70}
71\end{figure}
Note: See TracBrowser for help on using the repository browser.