Recent Changes in projects

[TennesseeTony]

Thanks @emoga !

Disable VPN to sign up.

TAC alert, we need a TeAm created. edit: Thanks @Ken g6

 

[TennesseeTony]

Asteroids@home: New CUDA102 application was released

Today we released a new CUDA102 application v102.14 for both Windows & Linux.
A bug was fixed where application causes 100% utilisation of whole CPU core (thread in hyperthreaded CPUs)

Running a pair of test WUs now, on a GTX 1070 under Windows. Estimated 33min until completion. Looks like Asteroids is STILL a CPU project (Ryzen 3000 series runs a CPU task in under an hour on average). First one finished in 26m19s.

Notes: 100% GPU usage, but only 55-60% of the power target, almost no CPU usage (some spikes occasionally). Progress bar and time remaining freeze then jump ahead significantly.

 

[StefanR5R]

Rosetta@home
After a brief test of the https URL scheme at Ralph@home, Rosetta@home switched its project URL from http://boinc.bakerlab.org/rosetta/ to https://boinc.bakerlab.org/rosetta/. They made this change because companies, who want to contribute computer time, asked for it.

If you have still the old URL on your computers, you can keep using it. Switch over only at a time when you have no Rosetta tasks on the computer.

Furthermore, the application version was updated from v4.15 to v4.20 just half a day ago. I haven't looked up yet what the changes are. I wasn't attentive at Ralph@home for the testing period of v4.20, which lasted just 17 hours.

Edit,
v4.20 has a fix for occasional failures to create initial task files. This problem of the previous version would cause tasks failing with error very early during execution, AFAIK. I haven't found out yet whether there are any other fixes in it.

Edit 2,
the v4.20 update includes according to a forum post:

1. extraction of the Rosetta database into the project directory with all following jobs reading from the same database rather than extracting into the slot directory for every job. This significantly reduces the disk usage per job.
2. checkpointing in the Rosetta comparative modeling protocol. This should significantly reduce wasted cpu time if jobs are preempted often, particularly for jobs that take a long time to produce models.

 

[StefanR5R]

PrimeGrid
An LLR2 applications is being tested since ~a week now. Its important change over LLR is that it incorporates runtime self-tests for software and hardware errors, by Robert Gerbicz. It will not replace LLR completely; LLR will continue to be used in subprojects which test k*2^n-1, as well as on 32bit hosts.

First step of integration includes partial replacement of current LLR by LLR2 in "Gerbicz runtime error check" mode for some projects and platforms:

• Platforms: Windows-64 and Linux-64 only;
• Projects: all except base-2 "-1" tests, because LLR2 uses another testing method and incompatible residues.

Remaining platforms and projects will continue to use old LLR (residues are compatible).

Please note that it will not enable fast validation scheme yet. Only Gerbicz runtime error check will be enabled. It will allow to catch all types of hardware and, what is most important, software errors. Ideally, we should not see mismatching residues anymore. Even completely broken/overheating computer should be either eventually complete the test, even abort it after few attempts, but never return wrong residue.

The importance of Gerbicz check is that it is very good at detecting local hardware and software errors almost for free, and provides the way to recover from them. We should see a dramatic drop in invalid results. That's why it's important to start using it asap.

(found via a post from pschoefer in the SETI.Germany forum)

Translated from German:

In the long run, double checking will be revolutionized by this. More on this in due course.

Update:
The LLR2 application is already live. It's version 8.10 among the LLR-based applications:
CUL, DIV, ESP, MEGA, PPS, PPSE, PSP, SOB.

 

[StefanR5R]

PS,
the potential big impact for less double-checking might come from the so-called "fast double-check" feature, alias "Proth/PRP test proof verification/ certification scheme" proposed by Pavel Atnashev. This feature is already implemented in LLR2, but not yet implemented in PrimeGrid's validator, hence not yet used in PrimeGrid's 8.10 application versions.

 

[emoga]

New project: MLC@Home (link)

studying ways to extract information from neural networks so we can better understand their uses and limitations.

Linux only at the moment. Windows app coming soon™
source.

Edit: Around 700 MB per WU. Ouch.

 

[Fardringle]

View attachment 24724

New project: MLC@Home (link)


Linux only at the moment. Windows app coming soon™
source.

Edit: Around 700 MB per WU. Ouch.

The tasks I'm getting are only 5 - 7 MB each...

Edit: Completed tasks receive 260 credits, regardless of how long they take. So far I've had run times anywhere from 10.5 minutes to 75 minutes on my Ryzen 9 3900X, with most of them in the 65-70 minute range.