This is my last entry for Fedora-ARM and SBR600 @ Seneca College project "To Thumb or Not To Thumb?".
When I first read about the project, frankly, I had no idea where to begin from. The company ARM and their chips, guruplugs, beagle boards and Fedora-ARM architecture; all of these were new to me and as I was trying to find connections between them, it became more interesting. So my project Description is as follows,
Fedora-ARM does not use thumb. The purpose of this project is to discover whether thumb provides any significant savings in terms of code size, whether programs compiled to thumb execute more quickly or more slowly than non-thumb programs on common ARM processors, whether a thumb compilation takes more or less time than non-thumb, and whether there are any other factors that would influence the decision to support thumb. Ultimately, this project should make a recommendation on the use of the thumb instruction set for the Fedora-ARM secondary architecture.
What I have done mainly to complete this project was
- Download different source rpms that are compiled with gcc.
- Using rpmbuild, recompile the source rpm in Thumb mode.
- Using rpmbuild, recompile the same source rpm without Thumb mode.
- Compare if Thumb speeded up the process or not.
- Compare if Thumb shrunk the file size or not.
- Compare if applications compiled with Thumb runs faster or not.
I worked alone for this project, using cdot-guru-4-1 (Fedora 13, armv5tel), cdot-beagleXM-0-3 and hongkong.proximity.on.ca at Seneca Centre for Development of Open Technology (CDOT). My project page can be found in Seneca Wiki: To Thumb or Not to Thumb.
Major tools, commands and configuration files I used are
- rpmbuild
- /usr/lib/rpm/redhat/rpmrc
- /usr/lib/rpm/redhat/macros
- time
- screen
- yumdownloader
- rpm2cpio
In the beginning, I compiled, benchmarked and posted my result without understanding rpmbuild fully, wasted tones of time. During my SBR600 class, the professor, Chris Tyler told the class that the directories in rpmbuild are for following purposes.
rpmbuild/SOURCES — Contains the original sources, patches, and icon files.
rpmbuild/SPECS — Contains the spec files used to control the build process.
rpmbuild/BUILD — The directory in which the sources are unpacked, and the software is built.
rpmbuild/RPMS — Contains the binary package files created by the build process.
rpmbuild/SRPMS — Contains the source package files created by the build process.
So here are the file size comparisons after compiling source rpm packages in THUMB, No THUMB and THUMB2. The rpm macro settings were identical except -march=armv5tel, -march=armv5tel -mthumb -mthumb-interwork and -march=armv7 -mthumb.
The numbers inside the table represent file/directory sizes in bytes. File sizes are checked with ll and directory sizes are checked with du --bytes.
nled-2.52
Produced RPM: nled-2.52-6.fc12.armv5tel.rpm
Directories and files: 2 directories, 1 file
Binary File: usr/bin/nled
| w/o Thumb | THUMB | THUMB2 | THUMB / NON-THUMB | THUMB / THUMB2 | |
| RPM | 20776 | 20204 | 20328 | 0.9724 | 0.9939 |
| Directory | 50892 | 43084 | 42684 | 0.8465 | 1.0093 |
| Binary | 38604 | 30796 | 30396 | 0.7977 | 1.0131 |
wget-1.11.4
Produced RPM: wget-1.11.4-5.fc12.armv5tel.rpm
Directories and files: 86 directories, 48 files
Binary File: usr/bin/wget
| w/o Thumb | THUMB | THUMB2 | THUMB / NON-THUMB | THUMB / THUMB2 | |
| RPM | 422360 | 420788 | 419852 | 0.9962 | 1.0022 |
| Directory | 1863679 | 1811467 | 1810431 | 0.9719 | 1.0006 |
| Binary | 226464 | 174252 | 173216 | 0.7694 | 1.0060 |
httpd-2.2.13
Produced RPM: httpd-2.2.13-4.fc12.armv5tel.rpm
Directories and files: 37 directories, 334 files
Binary File: usr/sbin/httpd
| w/o Thumb | THUMB | THUMB2 | THUMB / NON-THUMB | THUMB / THUMB2 | |
| RPM | 748180 | 738632 | 736500 | 0.9872 | 1.0029 |
| Directory | 2752985 | 2425373 | 2408025 | 0.8810 | 1.0072 |
| Binary | 287988 | 233660 | 229712 | 0.8113 | 1.0172 |
gimp-2.6.7
Produced RPM: gimp-2.6.7-2.fc12.armv5tel.rpm
Directories and files: 206 directories, 1204 files
Binary File 1: usr/bin/gimp-2.6
Binary File 2: usr/bin/gimp-console-2.6
| w/o Thumb | THUMB | THUMB2 | THUMB / NON-THUMB | THUMB / THUMB2 | |
| RPM | 12423104 | 12431184 | 12350600 | 1.0006 | 1.0065 |
| Directory | 45780103 | 43891439 | 43730427 | 0.9587 | 1.0037 |
| Binary 1 | 4284848 | 3450936 | 3371440 | 0.8054 | 1.0236 |
| Binary 2 | 2131228 | 1688860 | 1656092 | 0.7924 | 1.0198 |
abiword-2.8.1
Produced RPM: libabiword-2.8.1-1.fc12.armv5tel.rpm
Directories and files: 39 directories, 617 files
Binary File: usr/lib/libabiword-2.8
| w/o Thumb | THUMB | THUMB2 | THUMB / NON-THUMB | THUMB / THUMB2 | |
| RPM | 5561088 | 4779800 | 4742828 | 0.8595 | 1.0078 |
| Directory | 21295155 | 17055836 | 16987340 | 0.8009 | 1.0040 |
| Binary | 5806908 | 3450936 | 3371440 | 0.5943 | 1.0236 |
tar-1.22
Produced RPM: tar-1.22-8.fc12.armv5tel.rpm
Directories and files: 79 directories, 49 files
Binary File: bin/tar
| w/o Thumb | THUMB | THUMB2 | THUMB / NON-THUMB | THUMB / THUMB2 | |
| RPM | 744340 | 742784 | 742180 | 0.9979 | 1.0008 |
| Directory | 2737195 | 2681159 | 2675975 | 0.9795 | 1.0019 |
| Binary | 251844 | 195808 | 190624 | 0.7775 | 1.0272 |
cpio-2.10
Produced RPM: cpio-2.10-5.fc12.armv5tel.rpm
Directories and files: 49 directories, 30 files
Binary File: bin/cpio
| w/o Thumb | THUMB | THUMB / NON-THUMB | |
| RPM | 189908 | 189276 | 0.9967 |
| Directory | 842194 | 816194 | 0.9691 |
| Binary | 117064 | 91064 | 0.7779 |
gzip-1.3.12
Produced RPM: gzip-1.3.12-15.fc12.armv5tel.rpm
Directories and files: 9 directories, 33 files
Binary File: bin/gzip
| w/o Thumb | THUMB | THUMB / NON-THUMB | |
| RPM | 116560 | 115736 | 0.9929 |
| Directory | 263922 | 252086 | 0.9551 |
| Binary | 61708 | 49872 | 0.8082 |
bzip2-1.0.5
Produced RPM: bzip2-1.0.5-6.fc12.armv5tel.rpm
Directories and files: 7 directories, 21 files
Binary File: /usr/bin/bzip2
| w/o Thumb | THUMB | THUMB / NON-THUMB | |
| RPM | 49476 | 49232 | 0.9951 |
| Directory | 109233 | 103449 | 0.9470 |
| Binary | 33024 | 28076 | 0.8502 |
In average, THUMB reduced the sizes about
2.24% for *.rpm files,
7.67% for extracted directories, and
22.16% for executables, compared to NON-THUMB.
While it might save some disk space for executable files, it doesn't make a big different for rpm packages since rpm has it's own compressing mechanism.
In average, THUMB produced files that are about
0.24% for *.rpm files,
0.45% for extracted directories, and
1.86% for executables, greater than THUMB2.
I noticed that THUMB produces files that are almost as small as THUMB2's products.
In terms of performance of binaries that are built with THUMB and non-THUMB, there weren't too many programs that I could efficiently test in the command line interface.
I've downloaded source code for openoffice (openoffice.org-3.1.1-19.34.fc12.src.rpm), extracted twice to two separate directories, and compiled each one of them with tar binary that was compiled with THUMB, and without THUMB. I've done every test 4 times for better accuracy.
System: cdot-guru-4-1
File: Extracted openoffice.org-3.1.1-19.34.fc12.src.rpm
Command Used: time ./tar -cvf nothumb.tar ~/temp
| w/o Thumb | THUMB | |
| 1st Attempt | real 0m33.013s user 0m0.090s sys 0m3.800s | real 0m45.746s user 0m0.170s sys 0m3.710s |
| 2nd Attempt | real 0m38.410s user 0m0.150s sys 0m3.860s | real 0m40.039s user 0m0.130s sys 0m3.760s |
| 3rd Attempt | real 0m37.572s user 0m0.110s sys 0m3.820s | real 0m42.034s user 0m0.140s sys 0m3.820s |
| 4th Attempt | real 0m36.934s user 0m0.150s sys 0m3.670s | real 0m43.051s user 0m0.100s sys 0m3.830s |
The tar that was compiled with THUMB was compressing little bit slower in every attempts. Just to confirm that both tars compressed correctly, I've checked sizes of tarballs.
-rw-rw-r-- 1 thlee3 thlee3 218183680 2010-11-26 10:43 nothumb.tar
-rw-rw-r-- 1 thlee3 thlee3 218183680 2010-11-26 10:41 thumb.tar
And they seemed ok.
As a second test, I've tried decompressing one of the tarballs I created in the first test, using both tars.
System: cdot-guru-4-1
File: nothumb.tar
Command Used: time ~/tar.mthumb-interwork/bin/tar -xvf nothumb.tar
| w/o Thumb | THUMB | |
| 1st Attempt | real 0m48.399s user 0m0.170s sys 0m3.560s | real 0m54.197s user 0m0.160s sys 0m3.480s |
| 2nd Attempt | real 0m46.306s user 0m0.170s sys 0m3.280s | real 0m48.234s user 0m0.160s sys 0m3.680s |
| 3rd Attempt | real 0m45.602s user 0m0.130s sys 0m3.800s | real 0m49.501s user 0m0.150s sys 0m3.490s |
| 4th Attempt | real 0m46.546s user 0m0.110s sys 0m3.420s | real 0m52.897s user 0m0.120s sys 0m3.460s |
Again, the tar that was built with THUMB decompressed the tarball little bit slower. I should focus more into performance difference between binaries that are built with THUMB and non-THUMB in this project.
In addition to the benchmark above, I've also compared performances of two bzip2 and gzip that are compiled with/without THUMB.
I extracted & decompressed gimp-2.6.7-2.fc12.armv5tel.rpm file, archived the extracted directory with tar and tried compressing & decompressing using bzip2 and gzips that are compiled with/without THUMB.
System: cdot-guru-4-1
File: gimp.tar (44MB)
Command Used: time ./bzip2 ~/gimp.tar
| w/o Thumb | THUMB | |
| 1st Attempt | real 1m4.616s user 0m59.250s sys 0m0.650s | real 1m1.379s user 0m57.240s sys 0m0.720s |
| 2nd Attempt | real 1m4.985s user 1m0.530s sys 0m0.790s | real 0m59.899s user 0m57.190s sys 0m0.600s |
| 3rd Attempt | real 1m5.214s user 0m58.050s sys 0m0.620s | real 1m7.306s user 1m0.100s sys 0m0.530s |
| 4th Attempt | real 1m4.292s user 0m56.880s sys 0m0.700s | real 1m4.859s user 0m57.690s sys 0m0.770s |
| 5th Attempt | real 0m59.811s user 0m57.520s sys 0m0.470s | real 1m1.930s user 0m57.210s sys 0m0.540s |
| 6th Attempt | real 1m3.452s user 0m57.040s sys 0m0.590s | real 1m1.532s user 0m57.930s sys 0m0.790s |
| 7th Attempt | real 1m0.604s user 0m57.710s sys 0m0.650s | real 1m2.430s user 0m57.420s sys 0m0.730s |
For compression with bzip2, it shows almost no difference in performance.
System: cdot-guru-4-1
File: gimp.tar.bz2
Command Used: time ./bzip2 -d ~/gimp.tar.bz2
| w/o Thumb | THUMB | |
| 1st Attempt | real 0m18.828s user 0m17.430s sys 0m0.820s | real 0m19.662s user 0m17.850s sys 0m0.990s |
| 2nd Attempt | real 0m19.549s user 0m17.610s sys 0m1.160s | real 0m19.647s user 0m17.940s sys 0m0.850s |
| 3rd Attempt | real 0m19.462s user 0m17.600s sys 0m0.950s | real 0m19.271s user 0m17.620s sys 0m0.870s |
| 4th Attempt | real 0m23.303s user 0m17.800s sys 0m0.850s | real 0m22.863s user 0m17.290s sys 0m1.120s |
| 5th Attempt | real 0m24.544s user 0m18.120s sys 0m1.060s | real 0m19.155s user 0m17.560s sys 0m1.030s |
| 6th Attempt | real 0m19.397s user 0m17.540s sys 0m0.800s | real 0m22.998s user 0m17.400s sys 0m0.970s |
| 7th Attempt | real 0m19.462s user 0m17.460s sys 0m1.090s | real 0m19.688s user 0m17.910s sys 0m0.870s |
Similar to compression, decompression with bzip2 shows no noticeable difference in performance.
System: cdot-guru-4-1
File: gimp.tar
Command Used: time ./gzip ~/gimp.tar
| w/o Thumb | THUMB | |
| 1st Attempt | real 0m32.809s user 0m30.620s sys 0m0.780s | real 0m35.226s user 0m33.390s sys 0m0.770s |
| 2nd Attempt | real 0m43.640s user 0m38.300s sys 0m0.790s | real 0m38.300s user 0m34.380s sys 0m0.650s |
| 3rd Attempt | real 0m45.471s user 0m33.280s sys 0m0.670s | real 0m42.994s user 0m36.880s sys 0m0.710s |
| 4th Attempt | real 0m42.212s user 0m32.690s sys 0m0.510s | real 0m41.896s user 0m34.540s sys 0m0.740s |
| 5th Attempt | real 0m43.309s user 0m31.400s sys 0m0.740s | real 0m39.998s user 0m35.240s sys 0m0.840s |
| 6th Attempt | real 0m45.217s user 0m31.850s sys 0m0.550s | real 0m39.942s user 0m34.540s sys 0m0.670s |
The numbers were up and down for compressing with tar. I've actually made about 15 attempts and the results weren't similar at all.
The numbers didn't fluctuate as much for decompressing with tar. And it seems that the tar that is compiled with THUMB performs little bit slower overall.
System: cdot-guru-4-1
File: gimp.tar.gz
Command Used: time ./gzip -d ~/gimp.tar.gz
| w/o Thumb | THUMB | |
| 1st Attempt | real 0m3.341s user 0m2.520s sys 0m0.750s | real 0m3.688s user 0m3.040s sys 0m0.530s |
| 2nd Attempt | real 0m3.702s user 0m2.740s sys 0m0.630s | real 0m4.876s user 0m2.950s sys 0m0.650s |
| 3rd Attempt | real 0m3.372s user 0m2.680s sys 0m0.600s | real 0m4.889s user 0m2.970s sys 0m0.580s |
| 4th Attempt | real 0m3.512s user 0m2.760s sys 0m0.470s | real 0m4.800s user 0m2.890s sys 0m0.650s |
| 5th Attempt | real 0m4.413s user 0m2.800s sys 0m0.470s | real 0m4.016s user 0m2.930s sys 0m0.620s |
| 6th Attempt | real 0m3.316s user 0m2.430s sys 0m0.800s | real 0m4.033s user 0m2.960s sys 0m0.640s |
Lastly, I have compared THUMB and NON-THUMB using cpio. I tried archiving same extracted gimp directory as above benchmarks and measured the time.
System: cdot-guru-4-1
File: ~/gimp/ (206 directories, 1204 files)
Command Used: time find ~/gimp/ -depth -print | ./cpio -ov -O ~/gimp.cpio
| w/o Thumb | THUMB | |
| 1st Attempt | real 0m7.311s user 0m0.330s sys 0m2.300s | real 0m7.432s user 0m0.290s sys 0m2.480s |
| 2nd Attempt | real 0m8.064s user 0m0.300s sys 0m2.460s | real 0m7.574s user 0m0.230s sys 0m2.510s |
| 3rd Attempt | real 0m6.415s user 0m0.280s sys 0m2.430s | real 0m6.954s user 0m0.300s sys 0m2.510s |
| 4th Attempt | real 0m7.167s user 0m0.250s sys 0m2.500s | real 0m7.129s user 0m0.240s sys 0m2.520s |
| 5th Attempt | real 0m8.562s user 0m0.220s sys 0m2.540s | real 0m6.228s user 0m0.290s sys 0m2.500s |
| 6th Attempt | real 0m6.851s user 0m0.210s sys 0m2.580s | real 0m7.883s user 0m0.240s sys 0m2.610s |
| 7th Attempt | real 0m7.382s user 0m0.290s sys 0m2.420s | real 0m7.855s user 0m0.320s sys 0m2.430s |
2 cpios did not show major difference in performance.
At this point, I came up with a conclusion that THUMB does not greatly effect in software's performance, but it varies depending on software.
At this point, I came up with a conclusion that THUMB does not greatly effect in software's performance, but it varies depending on software.
I've also made records of the duration rpmbuild took for compiling. I'd like to say that I did not notice any major difference in compiling times, although it'd be hard measure this accurately because compiling durations fluctuated in every attempts I made.
| w/o Thumb | THUMB | |
| nled | real 0m13.701s user 0m12.190s sys 0m1.020s | real 0m14.821s user 0m11.830s sys 0m1.060s |
| wget | real 3m57.207s user 3m1.920s sys 0m40.610s | real 3m59.251s user 3m0.750s sys 0m39.810s |
| gimp | real 112m40.989s user 92m48.050s sys 10m34.130s | real 110m46.166s user 90m13.740s sys 10m32.150s |
| abiword | real 148m58.630s user 130m32.050s sys 8m51.320s | real 140m30.739s user 123m38.600s sys 8m46.910s |
Please check out my project page in Seneca Cdot Wiki at
This concludes my project To Thumb? or Not to Thumb? and SBR600 in Seneca College. I would like to thank Chris Tyler and Paul Whalen for all the assistance. It has been a great experience to be in SBR600 and be part of CDOT. Thank you.
No comments:
Post a Comment