V1.73 Beta

Risky64
Risky64
Joined: 1 Feb 25
Posts: 25
Credit: 48308
RAC: 2503
Topic 232086

I have hoped for a performance boost, in comparison with V1.51, but it needs twice as long to compute arecibo units with my tablet(Amrv8).

 

risky

ahorek's team
ahorek's team
Joined: 16 Dec 05
Posts: 39
Credit: 249686477
RAC: 3077

indeed. I tested it on my old

indeed. I tested it on my old phone which can run both apps and the old app seems to be 20% faster.

v1.46 (ASIMDPIE) vs 1.73 (armv8a) (beta test)

I've quickly reviewed the code but didn't find any obvious issues. All optimizations appear to be set up correctly.

@Oliver.Behnke any idea?

Risky64
Risky64
Joined: 1 Feb 25
Posts: 25
Credit: 48308
RAC: 2503

I am sorry, maybe this is a

I am sorry, maybe this is a non issue. I am new to crunching, and thought the beta would be faster, and i have used all 8 processors at the same time. It could be throttling, due to heat.

At this time, i only crunch 3 units together, and they only need 10000 seconds each.

risky

AndreyOR
AndreyOR
Joined: 28 Jul 19
Posts: 84
Credit: 772407213
RAC: 1062755

ahorek's team wrote:indeed.

ahorek's team wrote:

indeed. I tested it on my old phone which can run both apps and the old app seems to be 20% faster.

v1.46 (ASIMDPIE) vs 1.73 (armv8a) (beta test)

I've quickly reviewed the code but didn't find any obvious issues. All optimizations appear to be set up correctly.

@Oliver.Behnke any idea?

I've noticed same thing with other apps too, although to much smaller degree than 20%. Windows BRP7 v.12 (cuda5.5) is 5% faster than v.19 (cuda10.2). Linux O3AS v1.16 is 2% faster than v1.17 (both cuda).

I should compare Windows BRP4 on intel GPU v1.34 vs v1.70.

ahorek's team
ahorek's team
Joined: 16 Dec 05
Posts: 39
Credit: 249686477
RAC: 3077

I made a mistake in

I made a mistake in estimating the runtime. Actually, the new app is faster, but not by a significant margin.

new 17,850.97 https://einsteinathome.org/cs/task/1729374861
old 19,473.63 https://einsteinathome.org/cs/task/1729714538

it's on 2/8 cores ARM Cortex-A53. The main difference is the new app is 64bit, also the compiler and libraries like fftw were updated.

There may be some fluctuations depending on the number of cores and the hardware you have, but there doesn't seem to be anything wrong with the new app. It's always good to double check.

ahorek's team
ahorek's team
Joined: 16 Dec 05
Posts: 39
Credit: 249686477
RAC: 3077

AndreyOR wrote: ahorek's

AndreyOR wrote:

ahorek's team wrote:

indeed. I tested it on my old phone which can run both apps and the old app seems to be 20% faster.

v1.46 (ASIMDPIE) vs 1.73 (armv8a) (beta test)

I've quickly reviewed the code but didn't find any obvious issues. All optimizations appear to be set up correctly.

@Oliver.Behnke any idea?

I've noticed same thing with other apps too, although to much smaller degree than 20%. Windows BRP7 v.12 (cuda5.5) is 5% faster than v.19 (cuda10.2). Linux O3AS v1.16 is 2% faster than v1.17 (both cuda).

I should compare Windows BRP4 on intel GPU v1.34 vs v1.70.

Variations like this can be caused by the compiler e.g., Clang vs. GCC,, Linux vs. Windows or CUDA versions. Some optimizations may benefit one GPU architecture while being slower on others. Additionally, not all work units are exactly the same.

Sry, but a difference of around 2% is within the margin of error...

Risky64
Risky64
Joined: 1 Feb 25
Posts: 25
Credit: 48308
RAC: 2503

ahorek's team

ahorek's team wrote:

Variations like this can be caused by the compiler e.g., Clang vs. GCC,, Linux vs. Windows or CUDA versions. Some optimizations may benefit one GPU architecture while being slower on others. Additionally, not all work units are exactly the same.

Sry, but a difference of around 2% is within the margin of error...

I totally agree. Btw. you do not need to write a checkpoint every minute. about every 200 seconds is plenty.

AndreyOR
AndreyOR
Joined: 28 Jul 19
Posts: 84
Credit: 772407213
RAC: 1062755

Risky64 wrote: ahorek's team

Risky64 wrote:

ahorek's team wrote:

Variations like this can be caused by the compiler e.g., Clang vs. GCC,, Linux vs. Windows or CUDA versions. Some optimizations may benefit one GPU architecture while being slower on others. Additionally, not all work units are exactly the same.

Sry, but a difference of around 2% is within the margin of error...

I totally agree. Btw. you do not need to write a checkpoint every minute. about every 200 seconds is plenty.

BOINC doesn't dictate check-pointing, just sends a request to the task.  I believe most BOINC projects tasks checkpoint at points or intervals that are coded in, regardless of BOINC request. 

Risky64
Risky64
Joined: 1 Feb 25
Posts: 25
Credit: 48308
RAC: 2503

Now it is clear to me. V1.73

Now it is clear to me. V1.73 is about 20% faster than V1.51. More efficient. Could we have this version regularly? Stop the beta phase?

https://einsteinathome.org/de/host/13214403/tasks/0/0?page=3

San-Fernando-Valley
San-Fernando-Valley
Joined: 16 Mar 16
Posts: 565
Credit: 10944342825
RAC: 15144368

Risky64 wrote: Now it is

Risky64 wrote:

Now it is clear to me. V1.73 is about 20% faster than V1.51. More efficient. Could we have this version regularly? Stop the beta phase?

https://einsteinathome.org/de/host/13214403/tasks/0/0?page=3

Don't worry, I'm sure they will do that as soon as reasonable !

So, no rush please .

happy crunching ....

ahorek's team
ahorek's team
Joined: 16 Dec 05
Posts: 39
Credit: 249686477
RAC: 3077

The older apps will likely

The older apps will likely remain available for compatibility. Some older phones may not support 64-bit apps, while certain newer phones no longer support 32-bit apps.

The Intel Arc app has been in Beta for two years, and to my knowledge, there are no open issues with it. Once it's confirmed to be stable enough, it would be ideal to promote it to a stable release, though only server admins can do so.

For us crunchers, the only difference is that you need to enable the app by accepting beta work, and there's a lower limit on the number of tasks that can be processed per host per day.

Comment viewing options

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