All the stuff that runs www.canadarasp.com a website that tries to provide two things:
- Easy to understand summaries in the form of windgrams which provide a soaring forecast for paraglider, hang glider, and sail plane pilots in Canada and the northern United States.
- Complete data to allow a deep or shallow dive into the forecast in the form of a map view.
We also serve a large population of kite boarders, though I do not know whether they mainly use the map view for surface winds or the windgrams.
We also have as customers a few coastal engineering firms that are interested in wind forecasts with regards to interaction with wave and water damage. This is the origin of the BC archive map which is used by a small firm in BC (they paid a small amount of money for me to add that feature) where they like to be able to look and see the forecasts on days where there was storm damage at lake waterfronts. The surface wind archiving system was commissioned by a larger engineering firm and it also hangs out on the web server machine (but takes very little processing --- it just archives some data from the forecasts to S3 and processes it on demand when requested through a simple web service).
This runs in the Amazon cloud in ca-central-1a, which is Montreal. This is mainly because data transfers from ECCC appear fastest from there. The system consists of a t3.small EC2 instance which I will refer to as the web server and intermittently a c5d.4xlarge EC2 instance called the compute server.
The compute stack is hideous, a mix of bash scripts, common lisp for data processing, and some wgrib2 for data processing, and NCL for drawing the windgrams. Historically NCL used to be used for almost all of the data handling when we were just visualizing the area around BC off the HRDPS-WEST model, but once we moved to the continental model it was about 10x too slow to be useful, so I wrote the stuff that needed to be fast in common lisp.
The map overlay views are a bespoke pre-WTMS viewer with simple caching and just a fixed zoom. In principle we might want to move to a WTMS based storage scheme instead of the single zoom layer we use now, but given the limited dynamic range in zoom it is likely overkill. If we were serving the data out to anyone except our own javascript viewer then we should consider doing this. I did some experiments with generating TMS tiles using modern tools but didn't find any benefit in computation times or flexibility over the bespoke system currently in use.
It costs roughly $1300 CAD / year to run the system.
We rely on the freely available data from Environment and Climate Change Canada. The HRDPS model runs four times per day at 00Z, 06Z, 12Z, 18Z. We focus on serving west coast pilots who want a morning forecast before they head out and an evening forecast before bedtime. This well served by the 06Z and 18Z cycles which are ready at roughly 12:30Z and 01:15Z or 4:30AM and 5:15PM pacific time (depending on DST, etc). Can probably tweak the timing a bit here to get the data a bit earlier. We only run the 06 and 18Z data cycles. To better serve the east coast we might add one but it would cost several hundred dollars a year more and I don't know the uptake in the east yet. Not so much terrain on the Canadian side of the border so they are probably well served by the general weather and aviation forecasts.
The data comes in the form of GRIB2 files. These are compressed files each of which contains an individual forecast field across the forecast area: like temperature at a given pressure, or wind components, etc. Compressed an HRDPS run for the fields we use is about 10 GB of data. Some of the processing time comes from uncompressing these files at compute time to get the data we want. There are many libraries and command line tools to handle these sort of files.
For some basic data handling and for cutting out the dynamic windgram tiles, we use WGRIB2
For handling the map pngs we use the GDAL library from the Common Lisp code.
To efficiently generate windgrams, we cannot load the entire data set into memory (actually NCL just refuses to and dies --- it is 10GB compressed, > 400GB uncompressed). So we slice up the data into 2x2 degree squares for each field and then we concatenate all the fields into a 2x2 degree windgram tile. While generating the map pngs we work on each field individually, so we can load them directly though we still output 2x2 degree PNGs for speed of loading in the web browser.
The ECCC team is very responsive. Their point contact person for their data products is Sandrine at [email protected] who is always very helpful. The WCS/WMS stuff is the Geomet team who I have not interacted with yet (as my experiments with those interfaces were disappointing).
A t3.small EC2 instance runs the multiple web-server components (Apache2 static serving and proxying to internal services, dynamic wind grams, a time zone service, and an unrelated surface winds query service found in git repo hrdps-coastal-info). There are no automated builds of this system, you'll have to do it by hand. I regularly keep a couple snapshots of this to allow easy rebuilds if something goes completely haywire.
Multiple times per day (see canadarasp/crontab) the web-server
instance downloads data for the HRDPS or GDPS from ECCC. This is the
part of the system that most likely fails due to ECCC failing to have
data or having issues delivering data on time. This is handled by
canadarasp/aws-utils/download-data-and-start-server.sh MODEL where
MODEL is either gdps or hrdps depending.
SUMMARY: This is the start of the data processing pipeline. Download data and start the compute server
MODEL is either hrdps or gdps.
As seen in canadarasp/crontab, this is invoked at 01:15Z and 12:30Z for HRDPS and 08:15Z for GDPS. These times are chosen so that the ECCC models are finished by then. One could trigger this off of their AMQP message service, but I wasted a few hours fiddling with this and I couldn't figure out how to tell when the data was complete.
The general scheme of this is:
- Create / mount a new EC2 volume called download-box-MODEL ** This fails semi regularly, maybe once every few month due to what look like AWS failures. I have not managed to pin it down or make a reliable workaround. If this happens, reboot the web-server (it is un-recoverable without a reboot --- force unmounting makes it fail again later) and delete the old volume and manually retrigger the run that died if you want.
- Download data to the download-box. This script is reasonably robust and will wait for up to 3 hours for data to be available.
- Unmount / detach the download-box
- Start the EC2 compute-server with the name HRDPS PROD V7 (see default at top of this script). This machine is very expensive to run, which is why we do the downloading on the web-server. When ECCC is flaky sometimes the download can take a significant amount of time (usually it's done in less than 10 minutes for the 10GB of data or so for the HRDPS). We also set a tag on the compute server telling it which model we are giving it data for (just so it knows which download-box / ec2 volume to mount). There is another tag on the compute server called shutdown. If it is true the machine shuts down when finished. If false, it doesn't. This is just convenient for debugging the system. Don't forget to set it back to true when done fiddling.
- Wait three hours and kill the compute server just in case something goes wonky. Avoids unnecessary costs when things go wrong.
we download 48 hours for the HRDPS (it outputs data every hour) and 99 hours for the GDPS (it outputs data every 3 hours).
The compute server is a c5d.4xlarge machine that is spun up three times a day to process the HRDPS and GDPS data. It runs for about 1 hour for each HRDPS run and significantly less for the GDPS data. This is the main cost of the system. I use a c5d.4xlarge so the morning runs can be finished by the time people wake up in the morning on the west coast. To better serve the east coast I would have to add an HRDPS run (say the 12Z run) which would cost several hundred dollars a year more. This is a d. machine because I use the onboard 400GB SSD for storing intermediate results in the data processing pipeline. The performance of an EC2 volume with enough IOPS might be good enough but would cost more in the end.
The compute server when started starts up
continental-test/do-rasp-run-continental.sh through /etc/rc.local
The naming here (continental-test) is just historical and I've been too lazy to change it. continental-test has nothing to do with tests or with testing, it just refers to me branching off the code to handle the huge volume of the continental HRDPS data set.
We use a mix of bash and common lisp as a scripting language (it turns out that bash is too slow for some of the steps, taking minutes to perform operations that take seconds when written in common lisp --- remember this machine is expensive! I only switched to common lisp for the scripting components after I exhausted trying to speed up bash to be useful). Generally all these scripts can handle different model data sets, the differences being encoded in model-parameters.sh / model-parameters.lisp and the equivalent in the windgram ncl scripts.
**Summary: copy data to local disk and setup processing environment **
This script is run on boot by /etc/rc.local on the compute-server.
This script initializes the internal SSD and creates a local swap file (needed for stability later, though we don't actually swap in steady state).
This script attaches and mounts the download box, copies the data to the local SSD, detaches and deletes the download box.
Then it calls continental-test/do-hrdps-plots-continental.sh
**Summary: generate windgrams, map tiles, and dynamic windgram data tiles and upload them all to the web-server **
YEAR MONTH DAY HOUR are the forecast initialization time. These are
just handled by guess-time.sh MODEL. This works fine, though the
design is relatively fragile. We could frob this from the filenames
on the download box. But this works OK for now.
There are six main steps here:
do-generate-new-variables.shwhich computes some new parameters (HCRIT, WSTAR) across the whole domain and are used in the map soaring forecasts. These are re-computed in the windgram scripts because I was too lazy to fix them.clip-wind-to-terrain.shwhich clips the wind fields to terrain (this seems to be broken --- this has to do with which file has the terrain information). This makes a nicer looking map, but isn't truly necessary.do-tile-generation.shwhich generates the grib2 files used by the windgram generation scripts (both static and dynamic generation).do-windgrams-continental.shwhich generates the static windgrams for all the sites in locations.txt and uploads them back to the web server (along with a locations / regions list to be read by the javascript front end). It also uploads the grib2 tile files generated in step 3 back to the web server for the dynamic windgram generation.generate-single-hrdps-plot-continental.lispwhich generates the PNG and pseudo-PNG files for the google maps interface.upload-map-pngs.shuploads the PNG files back to the web server.
Generally for each of these we grab the model-parameters which are encoded in model-parameters.sh and model-parameters.lisp to tell us something about the files we are processing.
The NCL code that draws this is based on some ancient code by TJ Olney, a paraglider pilot from Washington state. His code is a lot like his car and his house were, just piles of stuff organized in a scheme only he could understand. Despite that, he created a visualization that is amazingly intuitive for most pilots. He died paragliding right after retiring from his day job. That code also used to use some fortran code from Dr. Jack Glendening, the creator of "RASP" which was a set of scripts that ran an ancient version of WRF based on NAM input data. I excised the fortran code as Dr. Jack made it expire at some point in time. It was unnecessary. The windgram code was heavily hacked by Peter Spear and myself to handle the different data from ECCC.
The front end is a mixture of javascript, CSS, and html. The map viewer was originally based on code from Paul Scorer, a sail plane pilot from the UK. Most of the bits are gone from that, but some of it still exists. It works, but it isn't particularly pretty.
Generally we are only displaying static PNGs and the like for the map viewer, but we encode wind direction in a pseudo-png and draw the arrows in the client. This saved several hundred dollars a year in compute costs server side.
I would like to change the map viewer so the user can specify the color scales interactively. To do so I would change the static PNGs to all be pseudo-PNGs where I encode the raw data. You can think of this as me doing bespoke compression on the original data. Right now when I write the PNGs out they are significantly smaller (10x) that the original data set. That's because the original data set uses double floats to encode the data (64-bits per point) while we are using something like 4 bits per point (we actually don't used paletted PNGs which would save even more space, but the compression takes care of the repeated colors nicely). Both are compressed with similar algorithms. The size reduction is noticeable --- it's 4.6GB in PNG format. Going back to say 8 bits per point, eg quantizing wind speeds to 1 km/hr would still be manageable. We could also switch to just slicing the entire data set up and leaving it in GRIB2 format and serving that raw and using any of the geo javascript libraries to decode it. That's probably the simplest. Either way, GRIB2 or pseudo-PNG we could then move the windgram generation to the client side if serving those static files is fast enough.
Anyway, we could just use gdal-translate to do the remapping instead of what we use now. Just translate to 8 bit grayscale for everything and put the scale information into some small JSON packets for ingesting in the viewer.
I don't really want to compete with things like windy, etc, as they do their jobs well, but we could make this look a lot nicer especially if we provide a WCS like interface to the data.
NCL is a deprecated scripting language (seems they have a python based replacement for it?). We could move to that, or we could do the above two tasks and rewrite the drawing code in javascript.
I'd like to support the NAM and GFS models. The GFS model would cost little to add aside from having to go through and see how the data is packed, etc. The NAM model, though, would again cost several hundred dollars a year to process and I'm not sure I'd get enough increase in users to cover the cost. Mainly the development time is what's stopping me though. I just don't have the motivation right now.
ECCC provides a WCS and WMS interface to the HRDPS data. My experiments with trying to use the WCS interface found it was about 100x too slow to be useful for generating dynamic windgrams. I did not investigate the WMS interface to replace the map pngs, but I expect it to be as slow as all the data needs to be generated dynamically.
This stuff is mainly for desktop GIS apps I guess. The WCS test took roughly 0.5 seconds to get a single field at a single location (and it doesn't support single points, despite the standard saying that should work, just regions, but whatever).
WTMS is a static image based tiling scheme for maps very similar to what we use, but that handles dynamic zooming and scaled down preview tiles. This doesn't exactly match what we want here because the underlying data is not particularly high resolution, so the dynamic range is well covered by a fixed zoom level but where the user sees the individual pixels. We have basically a bespoke version of this implemented. Regardless, let's see about the standard tool chains and whether we can leverage them to remove some of the canadrasp backend complexity.
I found gdal2tiles.py which appears to do what we want.
Summary: this tool seems best if the underlying data set is much higher resolution... it is way too slow and space wasteful as written to be useful for our backend
Here we can use gdal2tiles.py to generate a TMS compatible set of fixed PNGs. Here is an example
gdal_translate -of GTiff -ot Byte -scale -10 50 0 255 /mnt/input/hrdps/CMC_hrdps_continental_TMP_TGL_2_ps2.5km_2020083100_P001-00.grib2 /mnt/input/hrdps/CMC_hrdps_continental_TMP_TGL_2_ps2.5km_2020083100_P001-00.tiff
this scales from Celsius to 8-bit gray scale (-10 to 50C to 0 to 255). This is relatively fast (0.8 seconds, most of that time is uncompressing the original GRIB file).
time gdal2tiles.py -r near -z1-10 /mnt/input/hrdps/CMC_hrdps_continental_TMP_TGL_2_ps2.5km_2020083100_P001-00.tiff blarg
then generates the PNG files. This takes... 5 minutes instead of 1-2 seconds it takes currently. This is because it is trying to interpolate beyond the original resolution. It's kind of dumb that way. We start with a roughly 3000x2000 image and instead of setting the maximum resolution to that (ie 2.5km/2.5km) and just pixelating the zoom like we do, it spends a lot of time interpolating and the output is too large too because of the over-sampling.
Also would need to rewrite the map viewer to use fixed zoom, etc. All in all this doesn't seem to meet my needs.
The commercial implementation maptiler might work but it costs money.
This might be better but suffers from the same generic problem as above.
Here is how the time is spent:
- Initial start-up (copy files, make swap file) (3 minutes?)
- Generating new variables (3 minutes)
- Clipping wind to terrain (currently does not happen)
- Fixing TGL grib files (1 minute)
- Generate the sub tiles for the windgrams (14 minutes; now it is 24 minutes with new wgrib2... boo)
- Concatenating the sub tiles into single grib files (2.25 minutes)
- Generating and uploading windgrams (3.5 minutes)
- Generate the map pngs (14 minutes)
- Uploading map png files (5 minutes)
Total is about 45 minutes.
About 0.7 seconds * 4800 files is spent opening and uncompressing the same GRIB2 file each time with almost perfect parallelization. So if we combined the tasks 5 and 8 we could save roughly five minutes. Or if we went to keeping all the GRIB2 tiles we could save the windgram drawing (move to client side).
That isn't enough to make or break the whole thing but I feel like I can probably remove most of 5... more like saving 10 minutes.
If we could save 15 minutes then with no cost we could add a single run of the NAM probably.
It looks like generating all the GRIB2 windgram tiles is probably a good idea. Then writing a small access interface which opens, decompresses, and then streams just the single point to the user as a JSON message would probably work. It would take maybe 1 second to decompress each file. Then we can do client side windgram generation.
(They cannot access the full tiles as that is 48 MB per windgram which is too much to push down phones regularly... see below). A single point is tiny, though (roughly 50 kB assuming a JSON/text blow-up of 10x).
Currently the data needed for each windgram is kept on the web-server in GRIB2 format compressed with the complex-2 algorithm (best trade between time and space). Each 2x2 degree grid is roughly 48 MB of data... 1 MB per hour. There are 1288 2x2 grids which gives a total space requirement of 62 GB. I also am keeping two runs around which is wasteful. At 0.10 USD/GB-month and assuming we are only increasing the size by say 50 GB, would cost (* 50 0.1 1.3 12) $78/year which is OK. SO. If we increase the web prod storage by another 50 GB or so we can generate all the windgram tiles. Then we can move windgram drawing to the client side. This saves 10 minutes on the server side per day, or roughly 5 hours a month which costs (* 5 0.768 1.3) $5/month which roughly cancels this out.
Of course if we add the NAM that is double this, but that's OK.