Einstein@Home on Android Devices

A hint to keep it cool, when Boinc is crunching leave the device on a large cool metal surface if possible, heat is conducted away amazingly fast, and it'll remain cool to the touch.

Claggy

Cellphone isn't usually has edge with right angle, so some soft but nice contact material can transfer the hear much faster.

Does it officially?

AFAIK (e.g. from here) the drivers installed on the device are unofficial and thus don't necessarily support all OpenCL features.

Has this changed in any way?

BM

That will be nice.
It will also be nice if you guys start colaborating with parallella guys too. Their processor looks very promising http://www.parallella.org/board/

The Tegra 4 in a Nexus 7 does not support OpenCL (or CUDA for that matter, one needs the new Tegra 5, to be released in 2014 for that).

At the moment, there are phones and tablets available with Android GPUs that support OpenCL: http://e7dev.blogspot.nl/2013/07/android-opencl-gpu.html (Already sent to Rom).

For much the same reason that we do not support iOS: license issues and rather closed systems.

There had been some Distributed Computing activities on the PS3: until recently (a couple of months ago IIRC), Folding@Home was part of the Software pre- installed on the PS3, but this was not using BOINC. It was based on an individual agreement between Sony and the F@H project.

Both XBOX and PS3 are rather closed software ecosystems today.

PS3 used to support booting a Linux system, and several people used (and continue to use) BOINC under Linux on PS3s, but that multi-boot-option was removed by a Sony software update. The number of PS3s that still can boot Linux and run BOINC on it is not big enough to justify the effort of providing an official science app for it. But the BRP4 and GW-search code is Open Source...

There are hopes that the new PS4 will a) have a marketplace that is more open to independent software developers and b) be more Unix like so porting BOINC would be straight forward. We will see.

Cheers
HB

I just started to see info about BOINC on Android in the blogs and on Google+ yesterday (7-24-2013). I used to have Folding @ Home running on my PS3 until an update killed it. When I saw that my shiny new phone could be part of these exciting projects, I had to jump in. I'm hoping that more of the projects available to PC will become available for my Android phone as well. Right now, it's only a few projects but I'm still happy to help.

Wierd issue on new task on my Android phone. It's a Galaxy Note running android 4.0.4 OS.

Task - p2030.20121022.G194.41-02.68.S.b30g0.00000_939_1

I noticed it always seemed to be running, but the % was always at 0.0%. I watched it for awhile and noticed that it would run for 3:00 minutes and then reset to 0:00 minutes. It has probably been doing this for days. I aborted that task and will probably hold off on anymore until I have more time to watch them.

FYI,
Chuck

We have a known issue on some phones where this can happen if the setting "Use xxx % of CPU time" is set to anything less than 100%.

Cheers
HB

From a hardware point of view, the XBOX One , and the PS4, will be good crunchers. Traditionally console ecosystem are quite closed, but this may change a bit with the next generation. Microsoft will announce come details about how independent software developers (those not associated with the big game publishing companies) will be able to write software for the XBOX at Gamescon in August.

Cheers
HB

You don't have enough RAM available:

http://einstein.phys.uwm.edu/host_sched_logs/8245/8245613

Try a different project, like Asteroids@home, RAM use isn't so critical there (at least until the requirements are lowered). (Older Android mobiles have a limited amount of RAM, and of storage)

Claggy

Nexus 7: 3/4 cores assigned to Boinc, mix of Einstein, Pogs and MW
Asus EEPad: 3/4 cores, same mix
HTC: 1/2 cores, same mix
Hardkernel: 4/4 cores, overclocked to 2 GHz, watercooled, Einstein, Pogs

120mm fan? LOL, no idea, look here:
This is my main android device in original (Odroid U2):
hardkernel.com
And this is how it works here:
https://dl.dropboxusercontent.com/u/50246791/Odroid%20watercooled%201.jpg
https://dl.dropboxusercontent.com/u/50246791/Odroid%20watercooled%202.jpg
https://dl.dropboxusercontent.com/u/50246791/Odroid%20watercooled%203.jpg
https://dl.dropboxusercontent.com/u/50246791/odroid%20watercooled%204.jpg
:-)))

To be honest: in terms of performance you are right. But compare the power consumption: This device is overclocked to 4x2GHz and draws 1.8Amp @ 5V. Can any old style PC or laptop compare in any way?
This is the performance @ MW: http://milkyway.cs.rpi.edu/milkyway/show_host_detail.php?hostid=525266
This @ pogs: http://pogs.theskynet.org/pogs/show_host_detail.php?hostid=7962
This is a 50:50 mix of Pogs and MW.
I'll post actual Einstein performance in 3 or 4 days when I switch back to Einstein.

Another point: the technique is very young and in steady developement, device prices drop and performance raises, and someone needs to start with it. And remember the first days of SETI 12 years ago; I counted the dayly credits in tens, not kilo credits. Would you say it was senseless to do it? Let's give it a chance!

And to be correct: Higher temps decrease the lifetime of any electronic device.
And to be a little bit more correct: use chargers with lower charging current to keep the batt temp lower.

Always look at the bright side of live,

Cheers
Alexander

Not necessarily; the only one I bought for crunching only is the Odroid, which comes from south Korea. The other devices have dedicated work to do beside BOINC and would be on duty anyway.

And another thing: how many of us cruchers bought hi-end gpu's just only for crunching? Is your argument applyable to android devices only?

Re Parallella : the architecture is highly optimized for signal processing, but fortunately also for low power at ~ 600mW per core at typical loads ( including board support here, though I hear that may scale down to around 25 - 50mW per core with denser chipsets ). The kits now available are only at evaluation level. At 64 cores ( the current densest chip ) it might just start to be worth using for, say, FFT. That is still a long way from an E@H work unit. Really Parallella ought be viewed as a coprocessor component within another system, and not just for now but even if you do scale up to 4096 cores. Presently it requires an ARM dual core CPU, 1GB RAM and a minimal FPGA allocation to host it at all. On the up side that runs Ubuntu, but I can't honestly say how truly anyone intends to put a JVM ( or Dalvik look-alike ) on it! Plus about the only thing that isn't open about Parallella is the IP license ....

Cheers, Mike.

It's entirely possible that the successor(s) of the current Parallella implementations will do E@H. But that evidently will depend upon the commercial success and progress of the Adapteva guys ....

FWIW my rough/wild guess is that once you get to 256 cores ( 16 x 16 ) on a chip and sufficient per-core memory ( 1MB is addressable by design but only 32KB physically present now ) say 256KB/core then one can put serious slabs of data onto the array. After some setup, then ideally you want a single largest write into the chip, do some parallel magic, followed by a single largest read out ( + rinse/lather/repeat ).

BTW - there is a sub-group of the Parallella backers that is looking at porting Android to it.

Cheers, Mike.

I do have such a device too, my old, outdated POV Mobii.
http://milkyway.cs.rpi.edu/milkyway/show_host_detail.php?hostid=523949

Does not work with official Boinc but works with native Boinc, is attached to Milkyway and works fine. Has not enough memory for other projects. May help to keep your device usable.

Let's see: you're drawing 9 W at the wall for 163 credits/day per core, resulting in an optimal daily throughput of 652 credits.

My HD4000 at 1350 MHz @ ~1.00 V does about 8.8k RAC. When I suspend Einstein@iGPU and BOINC starts nothing else power draw at the wall is reduced by ~20 W, give or take a bit. This is using a voltage offset of -0.11 V, i.e. it's more power-optimized than regular HD4000's, but 1350 MHz is not exactly the most power efficient setup either (in mobile chips the GPU can go below 0.5 V!).

Using these numbers your ARM Quad would be 6 times less power efficient. That's unfair, though, as I'm comparing system power draw at the wall versus just the iGPU power at the wall. And an idle PC can add significantly to the total power budget.. however, modern laptops can idle below 10 W, whereas desktops reach 30 - 40 W. Such system power consumption would become almost negligible if more powerful GPUs would be used.

I'm sure comparisons with current AMD or nVidia GPUs wouldn't put the ARM in any better light, since these are pretty power efficient too. There's not even cost in favor of the ARM, as you'd need 13.5 of them to match even the relatively low-performance HD4000.

And here's the reason why I think we shouldn't expect miracles from these CPUs, nto even long-term: they're CPUs, not GPUs! You could push them for higher performance by including larger SIMD units feds by small and power efficient cores.. but ultimately you'd just end up creating something similar as Intels Knight's Corner / Larrabee. Or you dumb the cores down a bit further and you'd end up creating GPUs.

Personally I don't think it makes all that much sense to invest time & effort into developing for BOINC@ARM now, especially since most of these CPUs are put into devices which are normally not plugged into the wall all the time (and you really really shouldn't crunch on battery). However, if you do have the time and ressources anyway.. well, we've got to start at some point, as Alex said.

MrS

Hi MrS,

I agree, high-end GPU's are much more efficient than cpu's or arm's, no doubt. But the comparison was not made by me, the post says This is very poor performance even compared to a 7 year old low end laptop.
And the post contained a link to my arm devices. So I took the best one of these 4 devices and checked how relevant this comparison is.
As you posted, my device draws ~ 9 watts at full load and a modern laptop can idle below 10 watts; older style laptops draw more, be shure. And how much do they draw during crunching?

Next thing is comparing the performance.
Einstein is not the best project for that; statistics end at 10000 and the hosts there have a RAC that we will not see in the near future from any arm device (nor from a 7 years old low end laptop).
So I have choosen Milkyway for testing. The wu's validate against nVidia and AMD gpu's and, of course, against cpu's, Linux, Windows, and whatever is crunching there. So this is very similar to Einstein.
But the project is smaller and the chance to see these devices in the top computers list is higher.

On monday evening I set up my host to crunch MW only, 4 cores, 7/24. The RAC has reached 856.83 now, expecting to reach a value of ~2400 in 11 days (about one wu finished per hour giving 106 credits each).
After 3 days of crunching the device is listed as 8227 out of 35504 hosts with recent credit.

This is well within the top 25% of hosts there, after 3 days of crunching for the project. With an estimated RAC of 2400 the device will be in the top 15% of the hosts. Would be fair to take a 7 yr old low end laptop and
post the performance at MW here so that everyone can compare for himself.

Another argument was: But instead of this we put efforts on a new RISC-processors having lower stability, uknown yet bugs, and not almost the same productivity and stability
Hmm, let's see: valid 44, invalid 0, error 0.
http://milkyway.cs.rpi.edu/milkyway/results.php?hostid=525266

I do not do this to attack anyone, but people have a picture in their mind showing arm as low end and not worth to invest time and effort into these devices. I do this just simply to correct this picture.

BTW,
1. Hardkernel has announced a new board with the Exynos5 Octa processor, containing 4 A15 cores + 4 A7 cores and OpenCL conform.
2. I wonder why noone has ever tried to use sat-receivers to run boinc; many of these devices run some form of linux, there are many of them out there and plugged in 7/24. X-Box yes, sat-Receiver no?
3. More and more devices are on the market having Android operating system, for example media players and possibly soon even TV's. Let's see if someeone hacks them and uses these devices (without batteries!) to crunch. As posted somewhere else: the amount of the devices is what makes them useful.

I'm happy that you agreee well, we've got to start at some point

Alex

Hi Alex,

I didn't mention the comparison to that "7 year old low end laptop" because I think it doesn't make all that much sense to crunch on these either. Their energy efficiency as a platform may be OK (since they're laptops), but serious crunching will kill them in a matter of months to years. They're old, have probably dusted cooling vents and were not made for crunching. Even if the CPU is cooled well the mainboard gets a lot hotter than usual.. and is one of the first to fail in a laptop. I think there are surely better uses for these devices. And anything prior to Core 1 architecture is far too energy inefficient today.

Regarding Milkyway: do you run nbody or separation tasks? If it's separation: well.. even at 2400 RAC the comparison pales when looking at GPUs. Badly.. but this is true for any CPU doing separation tasks. I poushed my "old" HD6950 quite hard.. unlocked shaders and OC beyond HD6970 clocks, but keeping stock HD6950 voltage. I achieved ~260k RAC at an estimated power consumption of 150 W for the GPU, 300 W for the entire PC crunching 8 CPU tasks. 280 W after switching from Sandy to Ivy.

Granted that was a very efficient setup for MW and is probably above average, even now that we've got 28 nm GPUs. But if even only the top 100 crunchers used one GPU of this caliber, you'd need 10000 of your ARMs to match that performance!

It's a different story if you're running nbody tasks. I never cared about these, because I'd rather spend my CPU cycles for medical projects or Einstein, but here the ARM performance may be more impressive.

And about that comment about lacking stability.. I have no idea where this is comming from. Such a problem is certainly conceivable.. but I'd rather go with the hard numbers here.

Anyway, the future of ARM CPUs will be interesting. "Much you still have to learn my young Padawan".. but the progress is very dynamic!

MrS

Linux is not the argument for the device to start writing apps for it. The main reason is that these devices made up on very different chips thus having a very different architecture (both for ram and for CPU itself). I guess most of them have a C++ compiler, but... even these days we can and often do find bugs even in Intel Standard compiler that as said conforms to some calculation standards. Nothing to say for these new devices. The support for such devices can be made only if the project programmer will study for a new to program the devices from the scratch (even from the asm) to avoid such bugs (that can be in Linux kernel, compiler itself, processor core or in OpenCL-supporting libraries). So the only devices that can be succesfully supported are those that are reported officialy to fully conform to math-standards used in certain project. But most of devices (TVs, receivers and enginering devices like yours) instead have their own architecture and reduced Linux ports that don't conform to almost any standard thus having uknown and untested functionality.
And again, no new device can save your enough money with their energy economy to return the cost of such a device. But a rather old machine can help you to reach the same productivity and save you while that a $100-$200.
P.S. I have a lot of machines produced even in 2000. And they are well even this days crunching all this time for DC (firstly for S@H, then for E@H). Just look after them and they will crunch good for you.

And that drive even with the cost of replacement is much cheaper than a whole new PC. This is the benefit.

Hi!

I'm not sure that is really useful at this moment, the server status page is already quite crowded anyway. For me personally, the CPU cycles from Android devices are just as welcome as the CPU cycles from any other platform, no matter what their share in the overall statistics are (in terms of credits, I would estimate ca 140k credits per day go to Android hosts atm.). The more important effect to me is that we reached out to many new volunteers who were not involved in volunteer distributed computing before. Tablets and similar devices will play an ever increasing role in households in the near future, and rather than wait for things to happen and then follow the development, we like to be a bit ahead of such trends and experiment with new stuff while it's hot.

Cheers
HB

You can get this info at stats sites, e.g. BOINCStats Host OS Breakdown and look for "Android".

BM