type "?" to learn navigation, HOME
nasa travelSecond Trip to Washington, DC for NASA's Biodiversity and Ecological Forecasting Team Meeting
Collecting Urban Heat Island Data with Carly Ziter
Using OpenBLAS to speed up matrix operations in R (linux)
I use the
doParallel packages in R to speed up my work
that can be easily parallelized. However, sometimes work can't be
easily parallelized and things are slower than I'd like. An example
of this might be fitting a single very large and complex model. Andy
Finley, who resently stopped by UW-Madison to give a workshop on
hierarchical modeling, taught us about OpenBLAS as a way to speed up
matrix operations in R. Here are the notes about computing from the
BLAS is Basic Linear Algebra Subprograms. R and other higher level languages call BLAS to do matrix operations. There are other versions of BLAS, such as OpenBLAS, which are faster than the default BLAS that comes with R because they are able to take advantage of multiple cores in a machine. This is the extent of my knowledge on the topic.
Below is how I installed OpenBLAS locally on our linux server and pointed R to use the OpenBLAS instead of its default BLAS. A benchmark test follows.
cd src # move to src directory to download source code wget http://github.com/xianyi/OpenBLAS/archive/v0.2.19.tar.gz # your version may be different tar xzf v0.2.19.tar.gz cd OpenBLAS-0.2.19/ make clean make USE_OPENMP=1 #OPENMP is a threading library recommended by Andy Finley mkdir /home/erker/local make PREFIX=/home/erker/local install # You will have to change your install location
Pointing R to use OpenBLAS
I have R installed in my
~/local directory. libRblas.so is the default
BLAS that comes with R. For me it is located in
Getting R to use OpenBLAS is as simple as changing the name of the
default BLAS and creating a link in its place that points to OpenBLAS:
mv libRblas.so libRblas_default.so ln -s ~/local/lib/libopenblas.so libRblas.so
Deleting the link and reverting the name of the default BLAS, will make R use the default BLAS again. Something like:
rm libRblas.so mv libRblas_default.so libRblas.so
I copied how to do this benchmark test from here. The benchmark test time was cut from about 146 to about 38 seconds on our server. This is a very significant speed up. Thank you OpenBLAS and Andy Finley.
curl http://r.research.att.com/benchmarks/R-benchmark-25.R -O cat R-benchmark-25.R | time R --slave
Loading required package: Matrix Loading required package: SuppDists Warning messages: 1: In remove("a", "b") : object 'a' not found 2: In remove("a", "b") : object 'b' not found R Benchmark 2.5 =============== Number of times each test is run__________________________: 3 I. Matrix calculation --------------------- Creation, transp., deformation of a 2500x2500 matrix (sec): 0.671333333333333 2400x2400 normal distributed random matrix ^1000____ (sec): 0.499666666666667 Sorting of 7,000,000 random values__________________ (sec): 0.701666666666667 2800x2800 cross-product matrix (b = a' * a)_________ (sec): 10.408 Linear regr. over a 3000x3000 matrix (c = a \ b')___ (sec): 4.877 -------------------------------------------- Trimmed geom. mean (2 extremes eliminated): 1.31949354763381 II. Matrix functions -------------------- FFT over 2,400,000 random values____________________ (sec): 0.220333333333334 Eigenvalues of a 640x640 random matrix______________ (sec): 0.717666666666664 Determinant of a 2500x2500 random matrix____________ (sec): 3.127 Cholesky decomposition of a 3000x3000 matrix________ (sec): 4.15 Inverse of a 1600x1600 random matrix________________ (sec): 2.364 -------------------------------------------- Trimmed geom. mean (2 extremes eliminated): 1.74407855808281 III. Programmation ------------------ 3,500,000 Fibonacci numbers calculation (vector calc)(sec): 0.503999999999981 Creation of a 3000x3000 Hilbert matrix (matrix calc) (sec): 0.259999999999991 Grand common divisors of 400,000 pairs (recursion)__ (sec): 0.301000000000007 Creation of a 500x500 Toeplitz matrix (loops)_______ (sec): 0.0393333333333317 Escoufier's method on a 45x45 matrix (mixed)________ (sec): 0.305999999999983 -------------------------------------------- Trimmed geom. mean (2 extremes eliminated): 0.288239673174189 Total time for all 15 tests_________________________ (sec): 29.147 Overall mean (sum of I, II and III trimmed means/3)_ (sec): 0.87211888350174 --- End of test --- 144.64user 0.94system 2:25.59elapsed 99%CPU (0avgtext+0avgdata 454464maxresident)k 0inputs+0outputs (0major+290577minor)pagefaults 0swaps
cat R-benchmark-25.R | time R --slave
Loading required package: Matrix Loading required package: SuppDists Warning messages: 1: In remove("a", "b") : object 'a' not found 2: In remove("a", "b") : object 'b' not found R Benchmark 2.5 =============== Number of times each test is run__________________________: 3 I. Matrix calculation --------------------- Creation, transp., deformation of a 2500x2500 matrix (sec): 0.689666666666667 2400x2400 normal distributed random matrix ^1000____ (sec): 0.499 Sorting of 7,000,000 random values__________________ (sec): 0.701 2800x2800 cross-product matrix (b = a' * a)_________ (sec): 0.163000000000001 Linear regr. over a 3000x3000 matrix (c = a \ b')___ (sec): 0.228 -------------------------------------------- Trimmed geom. mean (2 extremes eliminated): 0.428112796718245 II. Matrix functions -------------------- FFT over 2,400,000 random values____________________ (sec): 0.224333333333332 Eigenvalues of a 640x640 random matrix______________ (sec): 1.35366666666667 Determinant of a 2500x2500 random matrix____________ (sec): 0.140666666666667 Cholesky decomposition of a 3000x3000 matrix________ (sec): 0.280333333333332 Inverse of a 1600x1600 random matrix________________ (sec): 0.247000000000001 -------------------------------------------- Trimmed geom. mean (2 extremes eliminated): 0.249510313157146 III. Programmation ------------------ 3,500,000 Fibonacci numbers calculation (vector calc)(sec): 0.505000000000001 Creation of a 3000x3000 Hilbert matrix (matrix calc) (sec): 0.259333333333333 Grand common divisors of 400,000 pairs (recursion)__ (sec): 0.299333333333332 Creation of a 500x500 Toeplitz matrix (loops)_______ (sec): 0.039333333333334 Escoufier's method on a 45x45 matrix (mixed)________ (sec): 0.256999999999998 -------------------------------------------- Trimmed geom. mean (2 extremes eliminated): 0.271216130718114 Total time for all 15 tests_________________________ (sec): 5.88666666666666 Overall mean (sum of I, II and III trimmed means/3)_ (sec): 0.30712894095638 --- End of test --- 176.85user 12.20system 0:38.00elapsed 497%CPU (0avgtext+0avgdata 561188maxresident)k 0inputs+0outputs (0major+320321minor)pagefaults 0swaps
From comments here, I have heard that OpenBLAS doesn't play well with
doParallel. I will have to test these next. If it is
an issue, I may have to include a shell code chunk in a literate program
to change between BLAS libraries.
Application Essay: Catalyzing Advocacy in Science and Engineering: 2017 Workshop
I just applied to the CASE 2017 Workshop in Washington, DC. The application process led to some interesting thoughts, so I thought I'd share the essay.
"How do we know the earth is 4.5 billion years old?" I loved asking my students this question when I taught high school science. The students (and I) were hard pressed to explain how we know this to be true. Most of us don't have the time to fully understand radiometric dating, let alone collect our own data from meteorites to verify the earth's age. So unless it's a topic we can investigate ourselves, we must simply trust that scientists are following the scientific method and evaluate their results within the context of our own experience.
Trust between scientists and the public is therefore the necessary foundation upon which our society accepts scientific research, incorporates it into policy, and supports more science. The communication of science's benefits to society maintains this trust. Unfortunately, the public and scientists disagree in many critical areas of research, such as genetic modification, climate change, evolution, vaccinations, and the age of the earth (1) (2). I believe scientists must do more to directly address these discrepancies.
As a scientist I have the incredible opportunity to conduct research that I think will improve society, and I'm honored that the public pays me to do it. I'm making a withdrawal from the bank of public trust and feel strongly that I need to pay it back with interest. I see scientific communication as the way to do so. Effective scientific communication goes way beyond publishing quality work in reputable journals and requires that we place our findings into the public consciousness. I have taught at the university and have led a few guest labs at an area high school, but I want to have a greater impact. The CASE 2017 workshop excites me with the opportunity to learn how to make this impact.
My hope is that CASE will orient me to the landscape of science advocacy, policy, and communication. Despite benefiting from federal funds for science, I am mostly ignorant of how our nation allocates resources to research, and I look forward to CASE demystifying this process. I hope to learn effective methods to communicate science with the public and to discuss with elected officials the value of research for crafting smart policy.
Because scientists understand their work best, they are best suited to advocate for it. CASE will provide a unique opportunity to learn how to be an advocate for science and a leader in strengthening the trust between the scientific community and the public whom we serve. If selected, I would like to work with the other selected graduate student and the graduate school's office of professional development to host a mini-workshop to bring the knowledge and skills from CASE to our campus. I'd like to replicate the Capitol Hill visits at a state level and work to get more graduate students engaged with elected officials from across the state.
Installing R, gdal, geos, and proj4 on UW Madison's Center for High Throughput Computing
R is the language I use most often for my work. The spatial packages of R that I use very frequently like rgdal, rgeos, and gdalUtils depend on external software, namely gdal, proj4, and geos.
Here I show how I installed gdal, proj4, and geos on chtc, and pointed the R packages to these so that they install correctly.
The R part of this tutorial comes from chtc's website. Their site should be considered authoritative. I quote them heavily below. My effort here is to help people in the future (including myself) to install gdal etc. on chtc.
Create the interactive submit file. Mine is called
I save it in a directory called "Learn_CHTC"
universe = vanilla # Name the log file: log = interactive.log # Name the files where standard output and error should be saved: output = process.out error = process.err # If you wish to compile code, you'll need the below lines. # Otherwise, LEAVE THEM OUT if you just want to interactively test! +IsBuildJob = true requirements = (OpSysAndVer =?= "SL6") && ( IsBuildSlot == true ) # Indicate all files that need to go into the interactive job session, # including any tar files that you prepared: # transfer_input_files = R-3.2.5.tar.gz, gdal.tar.gz # I comment out the transfer_input_files line because I download tar.gz's from compute node # It's still important to request enough computing resources. The below # values are a good starting point, but consider your file sizes for an # estimate of "disk" and use any other information you might have # for "memory" and/or "cpus". request_cpus = 1 request_memory = 1GB request_disk = 1GB queue
transfer interactive submit file to condor submit node
erker to your username and if you don't use
that too. You'll have to be inside the directory that contains
"interactive_BuildR.sub" for this to work.
rsync -avz interactive_BuildR.sub firstname.lastname@example.org:~/
log into submit node and submit job
ssh submit-3.chtc.wisc.edu condor_submit -i interactive_BuildR.sub
wait for job to start
Installing GDAL, Proj4, Geos
Each install is slightly different, but follows the same pattern. This worked for me on this date, but may not work in the future.
GDAL: Download, configure, make, make install gdal, then tar it up
wget http://download.osgeo.org/gdal/gdal-1.9.2.tar.gz # download gdal tarball tar -xzf gdal-1.9.2.tar.gz # unzip it mkdir gdal # create a directory to install gdal into dir_for_build=$(pwd) # create a variable to indicate this directory (gdal doesn't like relative paths) cd gdal-1.9.2 # go into the unzipped gdal directory ./autogen.sh # run autogen.sh ./configure --prefix=$dir_for_build/gdal # run configure, pointing gdal to be installed in the directory you just created (You'll have to change the path) make make install cd .. tar -czf gdal.tar.gz gdal #zip up your gdal installation to send back and forth between compute and submit nodes
Proj4: Download, configure, make, make install proj4 then tar it up
wget https://github.com/OSGeo/proj.4/archive/master.zip unzip master.zip mkdir proj4 cd proj.4-master ./autogen.sh ./configure --prefix=$dir_for_build/proj4 make make install cd .. tar -czf proj4.tar.gz proj4
wget http://download.osgeo.org/geos/geos-3.6.0.tar.bz2 tar -xjf geos-3.6.0.tar.bz2 # need to use the "j" argumnet because .bz2 not gz mkdir geos cd geos-3.6.0 ./configure --prefix=$dir_for_build/geos # no autogen.sh make make install cd .. tar -czf geos.tar.gz geos
Add libs to
I don't actually know what this path is exactly, but adding
geos/lib to the
LD_LIBRARY_PATH resolved errors I had
related to files not being found when installing in R. For rgdal the error was
Error in dyn.load(file, DLLpath = DLLpath, ...) : unable to load shared object '/home/erker/R-3.2.5/library/rgdal/libs/rgdal.
and lines like this:
... ./proj_conf_test: error while loading shared libraries: libproj.so.12: cannot open shared object file: No such file or directory ... proj_conf_test.c:3: error: conflicting types for 'pj_open_lib' /home/erker/proj4/include/proj_api.h:169: note: previous declaration of 'pj_open_lib' was here ./proj_conf_test: error while loading shared libraries: libproj.so.12: cannot open shared object file: No such file or directory ...
For rgeos the error was
"configure: error: cannot run C compiled programs"
Run this to fix these errors
export LD_LIBRARY_PATH=$LD_LIBRARY_PATH:$(pwd)/gdal/lib:$(pwd)/proj4/lib # this is to install rgdal properly export LD_LIBRARY_PATH=$LD_LIBRARY_PATH:$(pwd)/geos/lib # and rgeos
If you run:
The output should look something like
R: download, untar and move into R source directory, configure, make, make install
As ofR 3.3.0 or higher isn't supported on chtc
wget https://cran.r-project.org/src/base/R-3/R-3.2.5.tar.gz tar -xzf R-3.2.5.tar.gz cd R-3.2.5 ./configure --prefix=$(pwd) make make install cd ..
Install R packages
The installation steps above should have generated an R installation in the lib64 subdirectory of the installation directory. We can start R by typing the path to that installation, like so:
This should open up an R console, which is how we're going to install any extra R libraries. Install each of the library packages your code needs by using R's install.packages command. Use HTTP, not HTTPS for your CRAN mirror. I always download from wustl, my alma mater. For rgdal and rgeos you need to point the package to gdal, proj4 and geos using configure.args
Change your vector of packages according to your needs.
install.packages('rgdal', type = "source", configure.args=c( paste0('--with-gdal-config=',getwd(),'/gdal/bin/gdal-config'), paste0('--with-proj-include=',getwd(),'/proj4/include'), paste0('--with-proj-lib=',getwd(),'/proj4/lib'))) install.packages("rgeos", type = "source", configure.args=c(paste0("--with-geos-config=",getwd(),"/geos/bin/geos-config"))) install.packages(c("gdalUtils", "mlr", "broom", "raster", "plyr", "ggplot2", "dplyr", "tidyr", "stringr", "foreach", "doParallel", "glcm", "randomForest", "kernlab", "irace", "parallelMap", "e1071", "FSelector", "lubridate", "adabag", "gbm"))
Exit R when packages installed
Edit the R executable
The above will open up the main R executable. You will need to change the first line, from something like:
Save and close the file. (In nano, this will be CTRL-O, followed by CTRL-X.)
Move R installation to main directory and Tar so that it will be returned to submit node
mv R-3.2.5/lib64/R ./ tar -czvf R.tar.gz R/
Exit the interactive job
Upon exiting, the tar.gz files created should be sent back to your submit node
Cool Science Image contest
I created this image of Madison's lakes using hyperspectral imagery from NASA's AVIRIS sensor for the Cool Science Image Contest. I threw it together the week before the contest and was very pleased to be selected, but I wish that it had been more related to the science that I do. It is a minimum noise fraction transformation which is a way to transform/condense the data from the ~250 bands into the 3 visible channels (rgb) for maximum information viewing. Originally I intended to create an image over land, but had great difficulty getting the mosaicing of the 3 flightlines to be seamless. You can see the band across the northern part of lake Mendota from fox bluff to warner bay that is due to image processing, not something real in the water. The image is no doubt cool, but I wish I could say more what the colors meant (If you're a limnologist and see some meaning, please let me know). I think that pink may be related to sand, and green to bright reflections on the water. There's probably some algae detection going on too. My goal for next year is to make an image that is heavier on the science and still very cool.
Field work in northern Wisconsin
Field work provides the opportunity to be outside, help out on lab-wide projects, and to learn about new research that isn't exactly in my wheelhouse. September 8-10 I went to the north woods to help collect foliar samples as part of a NEON and Townsend lab project to ultimately predict foliar traits such as morphology, pigments, and other chemical constituents from hyperspectral imagery to create maps of these traits. This was the first year of a five year project. There's much more to the science behind the goal. But the aim of this post is not to explain all that, but rather, to share some images and the joy of being in the north woods.
Flux tower of Ankur Desai's research group, much smaller than NEON's. Maples creating lovely dappled light.
orgmodeMaking this website
I use emacs org-mode as the core application for my research. It makes sense to use the great org publishing features to create a website without having to learn many new skills. I had considered using jekyll, but ultimately realized that I could make a website that is just as beautiful and functional with emacs org-mode.
I've looked at tons of websites made with org-mode. I like Eric Schulte's best for an academic personal page, and I wanted to use the org-info.js for a blog with keyboard shortcuts for navigation and search.
If you're not familiar with org mode, check it out.
If you are already familiar with org mode, spend twenty minutes reading about exporting to html and publishing. The manual is pretty clear. Once you have a published webpage, check out some css stylesheets from other org sites that you like. Mine is a modified version of the stylesheet of eric schulte, who I asked permission from to use.
I spent no more than 3 hours setting up the site. Deciding that this was the approach I wanted to take and generating the content took a couple days.
You can clone the github repo to see how I have it set up.
It is great to be able to work on the content of the website in a very familiar way and export it to the internet with one command. Amazing.
nasa travelTrip to Washington, DC for NASA's Biodiversity and Ecological Forecasting Team Meeting
bikeRemoving Stuck Aluminum Seatpost from a Steel Frame
Use a sodium hydroxide solution with proper protection and ventilation. Be patient. Use rubber stoppers to block holes in frame (bottom bracket and water bottle braze-ons.
My seatpost had been stuck in my steel frame for years. Fortunately it was at the proper height, so it didn't bother me. When my headset broke and needed to be replaced, I figured I'd take care of the seatpost at the same time. I wasted an incredible amount of time trying to remove the seatpost and ruined my paint in the process which required a costly repowdering. This post is to share my experience so that you don't have to go through the same thing.
- What didn't work:
- Pipe wrench with 5 foot bar
- combinations of the above
- Tying it between two trees and trying to pull it apart with 3 men and a 6-1 mechanical advantage system.
- What did work:
- Remove everything from the frame except the seatpost
- Use a hacksaw to remove seat and create hole to pour solution down. Leave as much of the post as possible to reduce splashing, while still creating a large hole to pour solution down. post in frame, side view
- Stop up bottom bracket and braze-ons (any holes that will let the sodium hydroxide leak out of the seat tube) with rubber or cork stoppers. I got many of different sizes for less than a dollar at the hardware store.
- Place frame in well ventilated area on something to catch any spills (I used a plastic sled in my driveway). setup
- Add sodium hydroxide salt to water (not water to salt). I did this in an old milk jug. Sodium hydroxide is sold at your local hardware store as lye or drain cleaner. Check chemical composition to verify it is NaOH. I didn't measure the concentration of the solution that I used, but you don't want it to be so concentrated that it bubbles violently out of seat tube and destroys your paint. Also, the dissolving of NaOH is exothermic and the milk jug will get quite warm, or hot if it's very concentrated.
- Pour solution into seat tube. The solution needs to be up to the top of the tube so that the part of the post inside the tube will dissolve, but filling it up this high risks spashes. Fill up the tube part way to make sure there isn't a ton up bubbling and splashing, then fill up to top of tube (not post). If you didn't saw off too much of the post, this length of post sticking out of tube will help give you a splash buffer. I cut mine too short and the paint was destroyed
- Be patient. My seat post wall was quite thick, at least 2 mm. This will take a long time to dissolve. Wait until the solution is finished reacting with aluminum (you can hear the production of hydrogen gas), which may take a few hours. Then pour out the solution from your frame and dispose of the dark grey liquid (because I wasn't sure if the NaOH was completely used, I added vinegar in an attempt to neutralize the base).
- Repeat steps 5-7 until the post is completely dissolved or you can pull the post out.
- I had apex custom coating in Monona, WI repaint my frame.
They did a great job and the price was lower than everywhere else I looked, but it still wasn't cheap. Don't let the NaOH stay on your frame long!