source: docs/Cmpt886_Project_Report/05-conclusion.tex @ 4857

Last change on this file since 4857 was 1122, checked in by lindanl, 8 years ago

Add project report for cmpt886 on parallelizing Parabix2

File size: 1.6 KB
Line 
1\section{Conclusion and Future Work}
2This paper studied and analyzed the structure of Parabix
3and presents a pipeline strategy for XML parsing.
4Performance and energy consumption are evaluated on a quad core machine
5and compared with the sequential Parabix as well as two other XML parsers.
6The parallelized Parabix provides a 2X speedup over the sequential version
7by using about the same amount of energy.
8It also shows a substantially better performance and less energy consumption
9compared with Xerces and Expat.
10
11The presented parallelizing strategy can be applied to other sequential applications
12that share the same data dependency structure as described for Parabix,
13especially for applications based on parallel bitstream technologies,
14where the bitstream processing can be easily divided into different stages.
15
16
17There are many possible optimizations for further research.
18For example, the frequency of each core can be dynamically changed to reduce stall time and save energy.
19A new thread can be created and assigned to help thread one on the first stage since
20the first stage is the bottleneck of the overall performance as shown in Figure \ref{work_balance}.
21However, only the first three passes (first stage) of Parabix are data independent
22and hence can be processed by multiple threads at the same time.
23Therefore, this method might not work for a different work division.
24
25Other future research includes porting the parallelized Parabix to different
26architectures (e.g. more cores, NUMA, asymmetric machine)
27and dynamically assigning the workload by sampling the performance at run time
28instead of predefining work division based on analysis of the sequential Parabix.
Note: See TracBrowser for help on using the repository browser.