X-Plane can be configured to model real-world weather conditions. It does this by downloading weather report data from a server, operated by Laminar Research. That server, in turn, periodically fetches three total files from two different sources at NOAA: one for METAR-format weather condition data, and another source for two grib2-format files describing, separately, winds and turbulence. Internally, we tend to refer to this system, overall, as Live Weather, or, in the context of anything server-related, just Weather.
Around March 22nd, it broke down.
What went wrong
As is often the case: several things.
X-Plane began reporting that it couldn’t fetch all the data it needed for real-world weather. This was a little surprising since our weather server is supposed to cache the latest data it received, even if it’s outdated, to smooth over any connectivity or availability issues with NOAA’s servers. Further investigation revealed that only one of the three files it needed was both missing from its local cache, and unavailable from NOAA.
The root cause of that data being absent from cache hasn’t been tracked down, for the reason that that part of the system has been replaced (see below). There may be a bug in the caching logic. The server may have suffered a silent crash-and-restart some time after NOAA stopped serving the file. We’re not sure.
Investigation into the causes of the missing data was sidetracked for a little while because, right around the same time, NOAA happened to suffer a widely-reported, significant disruption to their networks and server operations. Naturally, we initially assumed this was the cause of the problem, and our efforts were mainly aimed at mitigating the harm of a temporary outage, not dealing with a longer-term problem.
In fact, what had happened was a longer-term problem, and the outage was, at most, only part of the cause of the disruption we were seeing.
In which we miss a NOAA systems upgrade
Effective on or about March 22, 2021, beginning with the 1200 Coordinated Universal Time (UTC) run, the National Centers for Environmental Prediction (NCEP) will upgrade the GFS and Global Data Assimilation System (GDAS) from version 15.3 to 16.0.
NOAA Service Change Notice 21-20 (Updated)
What that means is that X-Plane’s Live Weather system was about to break.
The system mentioned in that notice is the one that provides turbulence and wind data for the Live Weather system. These are the grib2-formatted files mentioned above.
Some time into this minor crisis, we noticed that both, not just one, of those files were missing—one had been properly cached, as intended, so it’d initially escaped our notice—and that the directory structure and file naming conventions on the NOAA server from which we source them had changed. Then we found the PDF containing the notice quoted above. Nestled deep in that PDF was the line:
Remove WAFS blended product at 1.25 deg
This refers to the file we used for turbulence data. The wind data we’d been using we could still find, albeit in a new location, but here was confirmation that the turbulence data we’d been using was simply gone.
In which various solutions fail
Our situation at this point was: we had found our wind-data grib2 file, and we had several WAFS (World Area Forecast System) files that looked, judging solely from their names, like they might contain the turbulence data we needed.
None of this was true.
We learned that the NOAA system upgrade had included an upgrade to the compression used on their grib2 files. X-Plane couldn’t understand it, and in internal testing actually crashed when attempting to use these files. That meant that though we had found a file containing the same data we needed for wind, we couldn’t actually use it.
Also, none of the other WAFS-related files contained the same turbulence data we’d been using, nor anything readily translatable to same.
So much for the easy solutions.
Fortran: friend or foe?
NOAA provides a program called cnvgrib to convert grib files to other sorts of grib files—for instance, as one might guess from references above to grib2, there is a grib1 that predated it, and cnvgrb can turn one into the other. It can also, most usefully to our situation, change the compression scheme used by grib2 files, so it should let us turn these new grib2 files into something we can use without having to modify X-Plane itself. Missing turbulence data aside, at least we could convert the wind data grib2 files to be usable.
Fortran is a… let’s say venerable programming language, like, for instance, COBOL. It’s a bit obscure these days, outside certain very long-lived systems and niche applications.
Cnvgrib is written in Fortran, and no recent compiled binary version is available. The Fortran (with some support from C) source code is provided by NOAA, though, along with instructions for compilation.
One Linux VM, a couple hours, and 17GB(?!) of Intel compilers and tools later, and we have a compiled copy of cnvgrib. As hoped, it can turn the new-compression grib2 files into old-compression grib2 files that X-Plane can understand.
There’s still the matter of the turbulence data, though.
Enter the Matrix
I skipped ahead a little. At this point in the troubleshooting process we hadn’t yet noticed that the turbulence data from NOAA wasn’t the same as what we’d been getting before, and once we could convert it to grib2 files that X-Plane could handle, it even looked like it was working.
So we deployed it.
By “it” I mean a very low-tech shell script that periodically downloads the two files we need from NOAA and uses cnvgrib to make them usable by X-Plane.
Though this apparently worked, we quickly realized something was very wrong when support began fielding reports of way-too-powerful wind in X-Plane, when using the real-world weather feature. This was caused by X-Plane applying inappropriate turbulence modifiers to wind effects, because we were feeding it (from its perspective) bad data.
Having no new source of the turbulence data we had been using, and with the new data causing serious issues, what could we do without cutting a new release of X-Plane?
As it happens, we’d snagged exactly one of the old WAFS-Blended files, containing the kind of turbulence data X-Plane could use, from the NOAA servers, before they dropped off forever.
The clear stop-gap solution, then, was to simply serve the working, reasonable turbulence-data file that we’d saved. All the time. Even though it’s increasingly outdated and definitely not “Real-World” anymore.
That’s right: so far as turbulence data goes, if you’re playing X-Plane with real-world weather turned on, you’re trapped in an old facsimile of the real world. You’re in The Matrix.
Fortunately, the effect of turbulence is minor enough, overall, that this doesn’t throw things off too badly (most of the time) and having some outdated real-world data is better than having none. Still, if you notice things behaving just slightly strangely—or if you have a sense of deja vu with real-world weather turbulence while flying—that may be why. The Matrix has you.
Are we going to keep lying to X-Plane about turbulence forever, then?
Nope! We’re working on translating other forms of turbulence data into something usable by existing copies of X-Plane, and have plans to make future versions of the Live Weather system more robust so that, hopefully, it won’t fail again—at least not in the same way.
Every week for the last ten weeks I’ve thought “I should really write a dev blog post” and then…not done that. This isn’t because all is quiet on the Western front – on the contrary, everyone on the team has been really, really busy, and the dev blog is never the loudest thing shouting for attention. But now we have a new RC available, so here we are.
Mi Memory Es Tu Memory
11.53 fixes one bug, and it’s a rare bug, but it’s “exciting” when it happens. It turns out that if a Lua plugin requests a really huge amount of memory, instead of saying “no,” X-Plane gives the Lua program someone else’s memory. This is not good! If the bank gave you someone else’s money, that’d be a bad bookkeeping error. This bug is too, and the consequences of the bug are typically “really insane stuff happens later,” which is hard to sort out. The plugin that crashes may not be the plugin that requested the memory.
X-Plane 11.53 fixes this – large allocations that cannot be fulfilled are denied, which should cause the Lua plugin to halt the affected script without destabilizing the system.
Script authors, if you’re wondering “now can I allocate a lot of memory in my Lua script,” the short answer is “no.” The longer answer is: when your Lua plugin uses a new version of LuaJIT that can use 64-bit addressing, this limitation will go away via a new plugin, without a change to X-Plane. Since the limitation is in LuaJIT, it’s out of our hands.
G2 Controller Support
Since we were doing a bug fix release, we have included support for the HP Reverb G2. For reasons I don’t fully understand, controller support didn’t “just work” in 11.52, so we had to create a new profile.
G2 users should be able to use their controllers with X-Plane 11.53. However you should also read our KB article for any additional issues with controllers, especially with misalignment. This version also includes a CLI option to adjust this if needed.
Tyler Has Left the Building
After almost seven years, Tyler has joined the ranks of Laminar Research alumnae. You may know him from such hits as:
The X-Plane 11 User Interface
X-Plane Mobile’s global scenery
X-Plane Mobile’s mass multiplayer
He will be missed! It took several weeks just to figure out all the things he maintains.
We Need More Jims
A few weeks ago, we posted a developer opening – I am pleased to announce Jim Keir as the newest member of the X-Plane development team. Jim is already fixing our screwed up code contributing bug fixes and learning the insides and outs of X-Plane’s almost 1 million lines of code. Jim brings our count of Jims up to two, which is still less of a namespace collision than our three Dan*s.
Multicore and Plugins
Most of what we are working on is still in the lab and hasn’t escaped yet. A few weeks ago we did have a discussion with developers in our third party developer Slack channel about multi-core and plugins.
The short story is this: in X-Plane 11.50, Sidney added a widget to the plugin admin window that shows how much main thread time they’re consuming, which in turn reveals how much each add-on is impacting FPS.
Plugin authors responded! Lots of plugins moved their CPU processing time to a worker thread. This is mostly great – other cores tend to be underutilized on high-end machines so this gets us more FPS.
Here’s the concern: a lot of plugins are doing this, and they are each moving work to other cores in their own private way. There is no coordination between plugins, and one day we are going to wake up and X-Plane will stutter because plugins were (just for a frame) using all of the cores and leaving too few for X-Plane itself.
We are looking at a mechanism for plugins to use the background processing system that X-Plane has built in. The win would be that X-Plane could play traffic cop between plugins and the sim itself, and prevent background plugin loading from causing frame stutters.
I will write up a Request For Comments (RFC) as a future blog post, so that a wider audience can comment on this.
X-Plane 11.52r1 is now available & release notes are here. Steam will update automatically if you’re in the proper beta channel; Laminar Research customers will need to run the installer to check for the new beta.
We cut this release primarily to fix our number 1 auto crash report issue, which is a crash in the networking code. We do not yet have a fix for the G2 controllers–that should be next.
Thanks to the users who filed bugs (especially Bill), we now understand the issues with the HP Reverb G2, but the fix is not in 11.51r1. The fix needs more testing than will fit into this patch. If you have an HP Reverb G2 and haven’t filed a bug, please do, so we can find you to send you a possible test build. In the meantime, we won’t hold up 11.51 and the new Gateway airports; rather we can work on the G2 in parallel.
When we released X-Plane 11.51 beta 2 last week, we included up-to-date Aftermath support. Aftermath is an NVidia driver feature that catches detailed information when the GPU crashes (e.g. “device loss” crashes).
Full Aftermath debugging slows X-Plane down. The sim is still flyable, but you might go from 50 to 35 fps, for example. It’s noticeable, so we didn’t just turn Aftermath on for all eligible users. It’s too big of a perf hit to just leave it on all the time.
Unfortunately, while we know from auto-crash reporting that device loss errors are happening, we also know that no one is using Aftermath to capture detailed information that we could use to find and fix the potential problems in X-Plane.
So: if you hit device loss errors while flying with Vulkan on Windows with an NVidia card, please follow these instructions and run with Aftermath for a little bit. If you can drop us a few auto-crash-reports with Aftermath enabled, it could get us the key breakthrough we need to fix device losses.
Tyler also fixed some low level networking bugs. This doesn’t change how multiplayer fundamentally works – if you can’t do a LAN flight across your WAN or you need command line magic to get the right NIC, that’s all still true and really not in the scope of what we’re fixing in 11.51. This whole beta run is targeted bug fixes.
X-Plane 11.51 Beta 2 is now available. (Release notes here.) Here’s a few more details on bug fixes we are working on.
Device Loss Errors on Windows
A device loss error occurs when shaders running on the GPU crash. In the old days this might hang or blue screen your computer, but fortunately we live in the age of enlightenment – the GPU catches the error, stops running X-Plane’s shaders and leaves a note for the Vulkan driver to tell X-Plane “hey, you your code died.”
NVidia’s “Aftermath” is a diagnostic tool that can tell us why our shaders crashed on the GPU. When we tried to use it in the past it crashed, but NVidia has since updated their drivers, so we are trying again.
Aftermath can collect lightweight or heavy crash info; the lightweight crash info doesn’t hurt FPS, so it is now always on. Heavy crash info significantly lowers FPS, and must be turned on by the command line –aftermath flag.
So…we are looking for a few brave volunteers. If you:
Have an NVidia card with driver 457.09 on Windows 10 and
Sometimes see device loss errors during your flights and
Can live with some FPS loss for a little bit to fix these errors
Please run with
and auto-report any crashes that come up. If you get a device loss error with Aftermath running, the automatic crash report will contain all of the info we need.
Device Lost Errors on OS X
Device loss errors can happen on OS X in Metal, too – the mechanics are the same as Vulkan. We are aware of one AMD driver bug that caused them which we have worked around in 11.51b1. If you still see device loss errors in 11.51 betas, please file a bug, as we don’t have automatic reporting for these.
HP Reverb G2
We are looking into controller problems with the HP Reverb G2. In X-Plane 11.51b2, the grip trigger should start working again, but the default configuration will still be weird.
The problem appears to be that SteamVR identifies the first-gen and G2 WMR controllers the same way; we are still looking into this. If you have these controllers and haven’t yet sent us a bug report, please try them with 11.51b2 then send us your log with the additional diagnostics.
Clicking Settings Crashes on Windows
A few users have seen crashes when opening the settings menu on Windows, before ever turning on Vulkan. As best we can tell, the crash happens when we open the settings screen because we have to go inspect the Vulkan driver to see if it is usable, and that code goes off the rails. We have a possible fix for this; our theory is that it happens to users who have various third party “layers” (layers are basically plugins for the Vulkan driver) that have gone off the rails.
We are still fixing bugs in the Gateway airports export; it’s not ready for beta 2, but it should be in the next beta after this one.
On Tuesday Apple announced new Macs powered by Apple’s M1 chip, a custom ARM system-on-a-chip based on the Apple A-series System on a Chip (SoC) from the iPhone and iPad.
The rest of this post is probably only of interest to Mac users, but for Windows users, it’s worth noting that the M1 chip is fast. It targets laptop and low power use cases, not gamer-class hardware, and it’s not a discrete GPU. Here’s my 27″ iMac – Intel says the i9 in it is a 95W part:
The take-away here is that Apple doesn’t just have fast chips for their new machines, they might have the fastest ones.
Now, how is this going to work with X-Plane and plugins?
X-Plane 11 is an x86_64 app, as are all plugins ever written for it. So if you run it on an Intel Mac, it just works, and if you run it on one of the new ARM Macs, it will run using Rosetta, which will translate the code as you fly.
In the future, we will have an X-Plane build that is “universal”–that is, it contains ARM and x86_64 code, and we will have a plugin SDK that contains both ARM and x86_64 code. At this point, plugin authors can start recompiling plugins to contain both types of code as well. Users with ARM Macs will have the choice to (1) run ‘natively’ in ARM for higher performance and use only plugins that are universal or (2) continue to run x86_64 code under Rosetta, so that all plugins work.
(This option is available for all apps that are universal on an ARM Mac – you turn “Use Rosetta” on or off in the app properties.)
This situation is exactly the same as the PPC->x86 transition we went through years ago.
Plugin developers: once Big Sur and the new X-code are out and we have an ARM plugin SDK, you can add a new architecture to your project and that should be it, as long as you don’t use any x86 assembly code in your add-on.
It’s been a while since I have posted about what the team is working on, and given all that has happened in the last few weeks, it feels like a million years. Here’s a run-down of…stuff.
X-Plane 11: Beta Time
Today we are putting out X-Plane 11.51 beta 1. This is a bug-fix patch for X-Plane 11.50 with a handful of random fixes that we have accumulated over the last few weeks. Release notes here. You will not be auto-notified of this beta–you have to pick it in the installer if you really want it.
I expect the beta to be relatively short, as we’re just trying to put out fixes for things we’ve found since we’ve shipped 11.50, improve diagnostics, reduce crashes, etc.
11.50 beta 1 does not have new Gateway airports. We’ll include them very soon–probably in beta 2–we had a few last minute snags, so I pulled them out of beta 1 to avoid delay.
Road Map: Graphics and Performance
X-Plane 11.50 represents the first step in our long term performance road map: moving to modern, low overhead, high-performance rendering APIs. These APIs are multi-core friendly; for X-Plane 11.50 this results in better overall FPS and smoother performance, but only an incremental increase in multi-core use.
One stealth performance feature in X-Plane 11.50: plugin object instancing. X-Plane has had an instanced drawing API for several years now, but with 11.50 we saw widespread plugin adoption. This is going to be very important for performance going forward; the instancing APIs are designed for efficiency, particularly in a multicore environment.
We have now switched gears and we are working on new features in the engine itself, e.g. we are working on what we draw and not so much how we draw it. In other words, we are working on graphic enhancements, new features, etc.
The new features are, as they are being coded, already taking advantages of new tech made possible by Vulkan and Metal, e.g. GPU compute kernels, GPU-based culling, etc.
Once we finish rendering features, we can pivot back to performance and push hard on multicore. The next multicore goal is to be able to render multiple views in parallel using multiple cores. Parallel rendering has several benefits:
An X-Plane frame often has sub-views rendered to form the main view (e.g. shadows, water reflections, cube maps, in-cockpit cameras, etc.). Any concurrency we expose makes the sim faster in these scenarios, and they are common.
Right now while multi-monitor is possible with X-Plane, it is very expensive performance-wise. Having a frame that can be farmed out to multiple cores would make multi-monitor less of a performance hit.
Note that multi-core multi-monitor would still be single GPU, and it would be a win because right now CPU time limits multi-monitor setups.
What about multiple GPUs? That’s something we’ll have to look at after we have multicore on the CPU–without it second GPU support doesn’t help.
Big Sur & the Mac
There’s been a lot of Apple news this week that’ll have to wait for a separate post. We recommend waiting on Big Sur for a few days until we’ve had a chance to test it a bit. Hopefully that’s an easy ask, as right now the download servers appear to be overloaded.
Back in the day, the way you put in a feature request for X-Plane was to email Austin – his email address was all over the website. So was his phone number – if you really wanted the request you could just call him, but it might hurt your chances. Austin had a big text file where he copied all of the emails, and then he randomly jammed things into X-Plane. It was the wild west.
Fast forward twenty years – we have a development team, we have customer support, we have an art team, and X-Plane has a lot more users than it used to. So we made an official place to record feature requests: feedback.x-plane.com.
The feedback request board allows for voting, so please look through the existing requests and up-vote it if it’s already there–this lets us easily find very popular requests.
The request board covers X-Plane desktop and mobile. You can also request features for end users (“better clouds”) or for third party developers (“scenery packs can edit the mesh around airports”)–it’s all good.
Looking through the “most wanted” right now, it looks like our internal high priority items (most of which are in progress) match the most wanted list pretty well, which I think is a good sign for our upcoming dev work.