Folding on Android

[n_w95482]

I've been running BOINC on my phone and tablet for a few months - Asteroids@Home and Einstein@Home. I'm not sure what throughput desktop hardware normally gets with them, but so far I've accumulated 15720 points for Asteroids and 12687 for Einstein (my phone being the majority for both). Current work unit times on my phone are a bit over 34 hours for 64.2% of Asteroids, and just under 2 hours for 16.5% on Einstein. It's a Galaxy S III with a dual-core 1.5 GHz Snapdragon S4 Plus SoC (older Krait).

The client itself is pretty nifty. By default, it's set to run when only on AC power when nothing else is sucking up a substantial amount of CPU time. If I'm at my computer for a while (where my charger is) or in bed, I plug both devices in and let them do their thing. Once I leave for work and unplug them, they pause automatically. There's also settings for max RAM and storage usage, core and CPU usage limits, max battery temperature, minimum battery level (if set to run on battery), data usage limits and WiFi-only restrictions, and what type of power sources qualify for computation - power adapter, regular USB connection, wireless charger, or battery.

I'm curious how well F@H would run with NEON optimizations and/or GPU support, but I imagine it's a llloooong road to get there even now.

[bruce]

As a general rule, developing a new client costs Stanford more than the small scientific benefit that might be gained. The road would truly be llloooong. A client for a low-productivity platform that only has AC power part of the day would most likely introduce unacceptable delays into the science (remember most FAH WUs must be processed serially). On the other hand, there are a lot of tablets out there that might be used. There's a 3rd party monitoring program being developed so people can check on their home computer. It's an interesting idea, but not likely to be seriously considered any time soon.

Many DC projects use mostly fixed-point hardware (MIPS) whereas FAH needs mostly floating-point hardware (FLOPS) so not all CPUs would be productive. GPU or NEON optimizations can change that, although it makes the scientific code more difficult to write. Do you have any idea how universal NEON optimizations would be? I looked up the specs for my Samsung tablet and they won't tell me what CPU is in it. What percentage of the GPUs have OpenCL or CUDA support?

[PantherX]

When considering portable devices (tablets, phablets, mobiles, etc), heat dissipation can be a serious issue since these devices aren't designed for constant high usage. When you add highly optimized code, the CPU/GPU would be working harder than before and thus, generating more heat than an unoptimized code. This could quickly become an issue for the user.

Hopefully, there can be manageable solutions in the future.

[Leonardo]

Modern smartphones have very powerful cpus.

Compared to a coffee maker or calculator, yes, powerful. Compared to any computer, no, bottom of the barrel. Smartphones are good at what they do, but relevant to real computers, they are simply third rate.

[n_w95482]

Do you have any idea how universal NEON optimizations would be? I looked up the specs for my Samsung tablet and they won't tell me what CPU is in it. What percentage of the GPUs have OpenCL or CUDA support?

At this point, NEON support is pretty much universal with mid/high-end SoCs. The last high-profile one I can think of with no NEON support was the Tegra 2.

OpenCL support is much less common. Not only that, but Google suppresses OpenCL driver support in Android on Nexus devices in favor of Renderscript. Third parties can retain driver support, but it remains to be seen how universal that will be.

CUDA support is nonexistent right now, Tegra K1 will change that but it's not out yet, nor does it have very many design wins so far.

Given the pace of ARM SoC development, I'd say GPGPU support will be in good shape in a year or two.

When considering portable devices (tablets, phablets, mobiles, etc), heat dissipation can be a serious issue since these devices aren't designed for constant high usage. When you add highly optimized code, the CPU/GPU would be working harder than before and thus, generating more heat than an unoptimized code. This could quickly become an issue for the user.

Hopefully, there can be manageable solutions in the future.

Yeah, it'll definitely be a problem with quad-core SoCs, especially in phones. With my Galaxy S III, it's currently running BOINC at 52-56 C at 918 MHz (auto throttle from 1.5 GHz). Sometimes it'll stay at 1.5 GHz and get up around 60-65 C. The phone gets noticeably warm when running - about 43 C on the lower part of the screen and 40 C on the opposite side, with the phone facing up and running at 1.5 GHz/64 C. I've had zero issues so far with running it at full load.

As for my tablet (HP TouchPad with webOS+Android 4.2.2 dualboot), it doesn't get warm at all. Unfortunately, it caps the CPU at 702 MHz when the screen is off. I haven't found a workaround yet for forcing it to run at 1.5 GHz all the time.

The SoC I'd be most worried about running full blast is the Exynos 5 Octa. The fast "half" of the CPU has four Cortex-A15 cores. From what I can gather, A15s use a substantial amount of power at full load - Qualcomm's SoCs tend to perform better in this regard.

[shiryunaga]

Ok can F@H have similiar client like this

www.gizmag.com/power-sleep-mobile-proce ... app/30838/

[[WHGT]Cyberman]

Take a look at BOINC: https://play.google.com/store/apps/deta ... oinc&hl=en

There are a few worthwile projects that take whatever computing power they can get. (Personally, I recommend WCG.)

[bruce]

Exactly. Even if it could run, I don't think it'd be worth it. In order to maximize productivity you'd have to plug it in all the time in order to avoid draining the battery....

Plugged in 24x7, a smart phone or tablet might produce a percent or two as much work as the simplest of computer CPUs. There may be a lot of android/iOS devices available for folding but even if they were all harnessed for FAH, it's a lot less science than is currently being processed with home computers.

See Leonardo's comment, above.

[X1900AIW]

Plugged in 24x7, a smart phone or tablet might produce a percent or two as much work as the simplest of computer CPUs. There may be a lot of android/iOS devices available for folding but even if they were all harnessed for FAH, it's a lot less science than is currently being processed with home computers.

Such a 24/7 android machine is running now for a week for SIMAP (hostid 718980), you can compare it with different CPU from other home computers, i.e. with i7-4770K in wuid 45191801. Runtimes 8900 seconds versus 2200 seconds. Or against Opteron 6272 (wuid 44247848), 8500 seconds versus 4200 seconds.

The ARM CPU is a Rockchip RK3188, designed for tablets, capable in my box version up to 1,4 Ghz but running constant @1 Ghz, full usage all cores running each SIMAP the power usage of the complete box (Bluetooth + WLAN off, LAN on) is 4 watts (in words four).

Yes it produces less science but with split energy costs. A lot less science ? Referred to your´s criteria*.

These boxes have cons, they can not be upgraded to more RAM or memory, in a multiple purchase they are too expensive ATM related to a small budget x86 quadcore, but they have Ethernet Lan, two perfect usable USB interfaces, need no fans for cooling, are very small in required place (screenshot reverse side: 1xHDMI can be removed, 1x LAN for network, 1x power supply). I provoked exactly 4 error workunits on that android machine when switching off the running box by mistake, the power button topside can be hitten easily. I love it and would love to use it for folding. And I know folding needs quick return results, IMO that is the main reason efficient android cruncher are far away to become an alternative plattform.



* Calculate the TCO for 1,2 .. 5 years in contrast to increasing as country specific level of energy costs.

[7im]

Cruncher? Crunch what? A bunch of database queries? Before you show me a bug sitting on a leaf in a puddle as proof that you could build a cruise ship to carry a 1000 people across an ocean, you best show me proof of an ability to build at least build a boat.

What kind of crunching is done by SIMAP on your device? Does it actually process protein data and calculate new protein structures like fah? Or is SIMAP just a fancy search engine? What runs SIMAP faster, more flops or more threads? If more threads is the answer...that is not a sea worthy boat.

[PantherX]

I believe that SIMAP stands for The Similarity Matrix for Proteins (http://boincsimap.org/boincsimap/). However, I am unsure of how SIMAP and F@H compare in the workload/computational area.

[7im]

I believe that SIMAP stands for The Similarity Matrix for Proteins (http://boincsimap.org/boincsimap/). However, I am unsure of how SIMAP and F@H compare in the workload/computational area.

Thanks. Helpful for those who still don't have that googler thing figured out. Does that link lead to any info about computational workload?

[davidcoton]

Well, the page starts with the sort of semi-truth that gives popular science a bad name:

Many smartphones are a supercomputer!

Oh? By the standards of 1980, yes. By today's standards, no. Though possibly a defensible statement if "many" is a very large number (many phones to make one project with compute power equivalent to a supercomputer -- but without the IO capacity to match).

It appears that simap is basically a database matching task for identifying "protein sequences" -- noting similarities in static structures. This is very different from FAH which uses floating point calculations to simulate protein folding -- a dynamic process which is highly compute-intensive. I would guess simap generates a lot of net traffic but relatively little stress on the CPU (and no FPU or GPU activity), though I might be wrong.

It looks like simap can use a smartphone effectively in a DC project. FAH is different in essence, and almost certainly can find better ways to keep programmers busy than writing a FAH client for Android.

David

[Ivkosky]

Hi there


Some of you have probably already noticed that both global players - Samsung as well as HTC - have almost simultaneously launched new applications dedicated to distributed computing for Android smartphones.

Samsung Power Sleep
https://play.google.com/store/apps/deta ... powersleep

HTC Power to Give
https://play.google.com/store/apps/deta ... om.htc.ptg

I would like to ask people cleverer than me whether it is of any worth trying these apps. HTC's app is very much like Boinc whereas Samsung seems to be closer to FAH in its purpose. I have to plug my phone to charger every night anyway, so I thought I would couple it with something like this...

[7im]

Until fah has an app for that hardware, why not? Just realize, as mention above, these are light weight apps. For example, SIMAP doesn't make any new proteins, just compares the features of existing proteins. It's a glorified distributed file clerk. A necessary clerk that helps science, but does no new research. Good task to fit the abilities of the hardware.