Heads-up: BRP4 might run briefly out of work

Link
Link
Joined: 15 Mar 20
Posts: 121
Credit: 9167872
RAC: 47170

Sid wrote:That probably was

Sid wrote:
That probably was me however my error rate is a bit more thet 30%. Makes no sense to proceed.

No, it was Keith Myers' Tegra X1 machine, has over 50% invalid results.

.

Link
Link
Joined: 15 Mar 20
Posts: 121
Credit: 9167872
RAC: 47170

PorkyPies wrote:For my Pi

PorkyPies wrote:

For my Pi Cluster I have around a 5 to 6% rate of invalids.

Of the invalids I looked at they aren't all Intel GPU, a lot are Android not validating.

Yes, the v1.61 usually validates against an Intel GPU application, but not against Androids.

 

PorkyPies wrote:
The v1.51 (android) seems to match the v1.70 (intel) app and the v1.61 is declared invalid.

Not sure about that, as I didn't see v1.51 that much, but yes, it seems to be slightly different from the rest, no idea however if it needs to be separated completely from all other BRP4 applications or can validate reliably against some of them.

.

Link
Link
Joined: 15 Mar 20
Posts: 121
Credit: 9167872
RAC: 47170

Here is a host running v1.51.

Here is a host running v1.51. I don't see any other pattern than the one reported by PorkyPies: "The v1.51 (android) seems to match the v1.70 (intel) app and the v1.61 is declared invalid." So it's results match those of v1.70 a bit better than those of v1.61.

Aside of that special situation, it's like a 50:50 chance that it will validate against any other application, doesn't matter if it's a 1.4x Android or a 1.70 Intel GPU (couldn't find any 1.61). So it seems to generate results, which are somewhere between those of those two applications and that makes it not very compatible with anything else than itself.

.

Oliver Behnke
Oliver Behnke
Moderator
Administrator
Joined: 4 Sep 07
Posts: 982
Credit: 25170813
RAC: 0

Hi Link, Link wrote: This

Hi Link,

Link wrote:

This is pointless, already 50% invalid and will likely end up at around 80-90% as every returend result becomes inconclusive. Too many Intel GPUs and "v1.61 aarch64-unknown-linux-gnu". The only valid WU so far was one with a wingman running a custom CUDA application on NVIDIA Tegra X1 (and that computer also seems to have over 50% invalid results). BRP4A doesn't work either. Moving the phone to Moo! Wrapper until this is fixed.

may I ask why you're focused solely on BRP4 (all variants)?

Everyone, as I said above, we going to resolve this in about a week from now.

Sorry for the hassle and thanks for your patience,
Oliver

Einstein@Home Project

Link
Link
Joined: 15 Mar 20
Posts: 121
Credit: 9167872
RAC: 47170

Oliver Behnke wrote:may I ask

Oliver Behnke wrote:
may I ask why you're focused solely on BRP4 (all variants)?

Unless I missed something on the applications page, that's the only application you have for "Android running on ARM". My GPU is crunching O3AS (or BRP7 if O3AS isn't available), BRP4 is just for an old Android phone which can't run anything else than either your BRP4 or Moo! Wrapper, no other project will run on it.

 

Oliver Behnke wrote:
Everyone, as I said above, we going to resolve this in about a week from now.

No problem, I'll wait.

.

Link
Link
Joined: 15 Mar 20
Posts: 121
Credit: 9167872
RAC: 47170

In case this helps resolving

In case this helps resolving the issue, here the info on how the 10 BRP4 WUs validated, let's start with the valid ones. My results are from "einsteinbinary_BRP4_1.43_arm-android-linux-gnu__NEON":


p2030.1725495006.G42.81-01.36.N.b3s0g0.00000_783

My result valid against an custom CUDA app running on NVIDIA Tegra X1.


p2030.1725713105.G53.17+02.80.N.b0s0g0.00000_3181

My result valid against v1.70 (opencl-intel_gpu-newer) windows_x86_64.

Marked as invalid: v1.61 aarch64-unknown-linux-gnu.


p2030.1725713105.G42.81-01.36.N.b4s0g0.00000_1521

My result valid against another v1.47 (NEON) arm-unknown-linux-gnueabihf.

Marked as invalid: v1.51 (ASIMDPIE_1X) arm-android-linux-gnu, v1.70 (opencl-intel_gpu-newer) windows_x86_64


p2030.1725713105.G42.81-01.36.N.b4s0g0.00000_1812

My result marked as invalid against 2x v1.61 aarch64-unknown-linux-gnu.


p2030.1725495006.G42.81-01.36.N.b4s0g0.00000_3010, p2030.1725713105.G60.20-03.65.C.b5s0g0.00000_1243, p2030.1725946806.G60.20-03.65.C.b5s0g0.00000_2366

My result marked as invalid against v1.61 aarch64-unknown-linux-gnu and v1.70 (opencl-intel_gpu-newer) windows_x86_64.


p2030.1725713105.G60.20-03.65.C.b5s0g0.00000_2480, p2030.1725946806.G46.27-03.52.N.b5s0g0.00000_3235, p2030.1725946806.G46.27-03.52.N.b2s0g0.00000_451

My result marked as invalid against 2x v1.61 aarch64-unknown-linux-gnu.

Marked as invalid as well: v1.70 (opencl-intel_gpu-newer) windows_x86_64. So v1.61 and v1.70 perhaps don't match as often as I initially thought.

So basically it's just v1.33 and 1.4x that validate well against eachother.

.

Oliver Behnke
Oliver Behnke
Moderator
Administrator
Joined: 4 Sep 07
Posts: 982
Credit: 25170813
RAC: 0

Thanks for the details!

Thanks for the details!

Einstein@Home Project

Sid
Sid
Joined: 17 Oct 10
Posts: 164
Credit: 963214253
RAC: 382839

Oliver Behnke

Oliver Behnke wrote:

 

may I ask why you're focused solely on BRP4 (all variants)?

 

Some of us has a custom application that is using a cuda for BRP4G tasks. This application was built from a source code that is published on einstein site.

As far as I know the source code for BRP7 is not available so it is only one variant to use Jetson family SOCs.

However Jetsons is capable for BRP7 too.

Bernd Machenschalk
Bernd Machenschalk
Moderator
Administrator
Joined: 15 Oct 04
Posts: 4308
Credit: 249839379
RAC: 32477

BRP4G will remain out of

BRP4G will remain out of work. Our intention is to direct the computing power of the official Intel CPU BRP4G apps to FGRP5.

We relaxed the validation for BRP4. Your BRP4 apps (including the custom one for the Jetson Board) should now validate much better.

BTW: The code for BRP7 has a CPU part that seems numerically very sensitive to even the compiler version on Intel CPUs From my experience with ARMs I doubt that a Jetson board would produce results that would validate against the official Intel-based BRP7 Apps.

BM

Link
Link
Joined: 15 Mar 20
Posts: 121
Credit: 9167872
RAC: 47170

Looks good so far. Thanks for

Looks good so far. Thanks for fixing it.

.

Comment viewing options

Select your preferred way to display the comments and click "Save settings" to activate your changes.