We’re reaching the end of 2015 and as in previous years, I’d like to summarize for donors what Folding@home (FAH) has done in 2015 and where we’re going. In general, FAH has always tried to push the limits in terms of new software infrastructure (new clients, backend software), new algorithms, improving the accuracy of our approaches, and perhaps most importantly the disease areas we’re targeting and making advancement in. Here’s a summary of what we’ve done this year and where we’re going.
As in 2015, key areas in 2016 will also be infrastructure, methods, accuracy, and targets. Below, in each section, after the update for 2015, I’ll give my vision for where FAH is going. As some of these items involve unannounced initiatives with partners, some of these will unfortunately be more in teaser form (sorry – as always, this part will be much more brief than the summary of the previous year’s results). Much like I had to tease our mobile launch in January 2015 in my December 2014 summary, the good news is that hopefully donors won’t have to wait too long for some of these to be more formally announced in 2016.
For those who want to see all of the details, please check out our papers page, which includes abstracts, layman summaries, and links to the specific research results.
1) FAH Infrastructure
Philosophy and vision. As in previous years, we have continued our push in developing new backend and frontend software. This has allowed us to stay on the cutting edge, but does introduce more instability in FAH for donors, so this is often a difficult tradeoff. But, my thinking here has been that we’re here to push the limits and hopefully we can communicate that FAH is more like a piece of scientific equipment than a BMW, always changing and improving, but at the cost of frequent change and upheaval.
New Assignment Server (AS). We rolled out a new AS into production. I’m personally a little wistful since the previous AS was the last piece of code I was the original and primary author. However, in the 15+ years we’ve run Folding@home, the operation has grown significantly from me and a few students doing everything to a team of programmers, a large group of scientists, many affiliated labs, as well as sysadmins, designers, etc.
The new AS has some key features which are important in this “grown up” version of FAH: the old AS had some very onerous requirements for updating its configuration (which worked fine when FAH had 3 scientists) but now with 20+, we needed a whole new approach. Also, the new AS allows for more efficient allocation of FAH donor machines to problems as well as setting the stage for more advanced analytics.
Mobile client. With our launch of our mobile client (where mobile means phones and tablets), we’ve pushed into another type of infrastructure. My vision for mobile is that the trend in computing is clear: people are often going without desktops or laptops in favor of mobile, so FAH needs to evolve with those trends as well. While the current mobile phones do not have the compute power of modern desktop/laptop CPUs, the gap is narrowing quickly, especially with the power of mobile GPUs advancing very quickly as well. So, our mobile client lays the foundation for much of FAH’s potential future.
It’s also worth noting in passing that in the early days of our GPU client (which I’m beaming with pride to say predates other GPU computing in our field by a wide margin, as we were running GPUs on Folding@home even before CUDA existed), the GPU client also didn’t have amazing performance. But it was clear to me that it would take off and that bet has born out very well today. It’s of course way too early to know if mobile will follow a similar path, but I feel that it’s an important bet to make.
Native Client (NaCl). The NaCl client is another sort of experiment akin to mobile. It takes easy installation to the ultimate level by making running FAH just going to our web site, with a Chrome browser. The next steps here for growth is working on viral marketing campaigns to get the word out.
Work Server (WS) updates. Finally, we have continued to make improvements to the WS. These improvements are typically only visible to researchers in FAH, but this is very much the backbone of how FAH works and we continue to make the WS easier to use (simplifications to how researchers configure it), more powerful in features (in serving science calculations), as well as more powerful in its ability to serve up WUs (speed improvements from architectural overhauls to allow a given server to server more donors).
Stats system update. Finally, we have made some updates to the stats system to make it much simpler on the backend. This has led to fewer stats issues and a faster response in handling stats issues.
2016 Vision. As detailed above, to keep FAH on the cutting edge and make the most out scientifically of donor’s contributions.
New clients. We will look to expand the mobile clients more broadly. (Sorry for not saying more here, but like many of our rollouts, we do combined announcements with our partners.)
New Gromacs core. The Gromacs core continues to be a workhorse core for CPU clients and we are working on releasing an updated core with more broad hardware support, including support for AVX, which will give some donors an automatic performance boost.
WS to do more advanced science. In conjunction with new methods below, we will be rolling out new WS features, such as high frequency adaptive sampling (see below).
More advanced analytics. We’ve been working on this area for a while and it’s often pushed behind other priorities, but this is becoming a greater and greater need for us to monitor FAH in a much improved way.
New stats system. Given all of the other infrastructure work, this is possibly the least likely to happen in 2016, but we recognize this as a key need. This is the last part of FAH which is very old and has not been recently updated.
2) New methods
Philosophy and vision. Since FAH has been doing this for so long, it’s now often taken for granted that one can use a huge, loosely coupled set of processors to do huge calculations. But this is far from easy. In fact, when I was first proposing FAH, many people said it was actually impossible. And in a sense it was –– it would have been impossible to use many processors to do traditional parallelization of our calculations, as that requires a super high bandwidth, low latency network. And in the early days of FAH, many people were even on modems (yes! It’s easy to forget what life was like in 2000).
A key part of our success with FAH has been pushing the limits for how we can use a FAH like resource to do greater and greater things, i.e. more complex systems, more predictive models, etc. And this year has been a particularly productive year in advancements in this area. Moreover, the techniques that we develop here very naturally have been applied to cloud computing, further broadening the impact of FAH’s scientific impact.
Markov State Models (MSMs). In the last 10 years, our lab (and a few others) have pioneered MSM methods and they in recent years have very much come of age, becoming in many cases a gold standard for distributed computation. In particular, there were several key advances by Robert McGibbon and other researchers in my lab which have allowed MSM construction to be completely automated as well as to be more accurate. Our MSM code is fully open source, available at http://msmbuilder.org .
OpenMM advancements. Finally, the foundation of our newest clients (GPU and mobile) is OpenMM, our high speed molecular dynamics code. OpenMM is derived from the original GPU code from Folding@home and has now gone to power many other researcher’s work as well. This year’s advances to OpenMM have allowed us to run on modern GPUs, improve speed, and improve accuracy (change notes are available on the OpenMM github page). OpenMM is fully open source, available at http://openmm.org .
2016 Vision. As detailed above, to push our ability to do new science, especially given the extreme algorithmic challenges of a distributed architecture.
MSM advances. We’re continuing to push advances in MSMs, especially in connection with modern machine learning advances. This will also give us a big boost in our ability to understand the deeper meaning in FAH results.
Adaptive sampling. Adaptive sampling has played a key role in many FAH calculations, but it’s been done with a great deal of manual work by the WS project managers. In conjunction with new infrastructure, this will be now automatic.
OpenMM. OpenMM’s development will continue to push into greater speed and functionality, with a particular push into much improved speed for the advanced accuracy AMOEBA force field and support for AMOEBA in Folding@home. Both will come with OpenMM 7.0 which will roll into a new core (likely Core22).
3) More accurate models
Philosophy and vision. A key challenge in simulation in general is the tension between accuracy and computability. Here, I’m especially talking about the model we use to calculate the forces between atoms in our simulations. One could choose a very accurate model, but then not be able to compute very much (in chemistry, the most accurate models, eg FCI, barely can handle a few atoms and in FAH we want to simulate hundreds of thousands of interacting atoms). My vision here has been to have it both ways – to develop high performance models that have considerably improved accuracy, namely accuracy greater than what considerably more expensive approaches would afford. This year has seen the first fruits of this initiative, which was started several years ago, and like many FAH initiatives, donors only see the results at the very end (sorry, that’s the nature of research – there’s nothing we can show until this is ready for publication). But the first results of this initiative are very exciting, with great promise for years to come.
Water first. Our first improvements in models comes with our force field for water, led by Dr. Lee-Ping Wang, who is now a professor in the Chemistry Department at UC Davis. Water is very fundamental to all of our simulations, so it’s the natural place to start. Our water model is considerably more accurate than previous ones due to a radically new approach, where we use heavy computation and statistical methods to make fundamental advances, not unlike how MSMs have helped as well, albeit in a different area.
2016 Vision. Development of force fields more broadly. Going beyond water, we expect to revamp force fields across the board, with broad impact in our calculations for FAH. FAH will also play a key role in quantitatively understanding the improvements these force fields bring.
4) Disease targets
Philosophy and vision. In choosing targets, i.e. disease areas where we use Folding@home to make advances, we have been picking areas where there would be a significant impact combined with where FAH can help the most. In general, FAH projects run a very long time – it takes a lot of compute power to address many of the questions we tackle. Therefore, I’ll cover the newer areas below.
Kinases (key to cancer) now becoming common place. Kinases are key to many cancer therapies as in cancer, cells grow unchecked and inhibiting kinases can stop this. In the past, kinases were a reach for FAH (and I like pushing the team to reach for hard targets) but now they are very much commonplace and even in many cases relatively “easy” targets. This means we can aggressively go after many of them, especially in collaborations with key experimentalists. We have had a key kinase paper published, with more on the way.
Channels. Ion channels are also exciting targets, with very broad pharmacological and disease impact. However, these are very large and challenging systems, a reach much like kinases used to be, so this is an area to watch particularly in 2016.
2016 Vision. As discussed above, we will continue push to bigger and more sophisticated systems, hopefully with channels becoming much easier as kinases did. Also, I have been working on ways we can partner with other groups, including groups from pharma, for “open source” science in this space, making results fully available to improve the whole community’s ability to impact disease. Sorry that this is more brief here – like major software announcements, we hold back details on future targets until we’re ready to roll them out.
Summary. I’d like to thank all of our donors for their contributions to Folding@home. It’s exciting to me to be discussing these amazing results and 15 years ago would not have imagined that we’ve been able to do what we do now and I’m excited in terms of how our vision and plans for the future will be able to push us even farther.
PS For those who want to see all of the details, please check out our papers page, which includes abstracts, layman summaries, and links to the specific research results.