Optomized S5 SSE3

[B@H] Ray
[B@H] Ray
Joined: 4 Jun 05
Posts: 621
Credit: 49583
RAC: 0

James I have to agree with

Message 39320 in response to message 39318

James
I have to agree with that 100%.
On the download pages the info about the bad ones I think can stay, but the links to the DAT files should be eliminated.

I have a listserver at my server, I can set the list up if you think if would be used.

Cheers
Ray

Quote:

I neglected to fully explain one of my concerns noted above. When a patch/app is known to be invalid but is not posted as such in the 'official' read-only thread for downloads people keep downloading them.

This is not just a 'waste of time' for the people who download patches that are known to produce invalid thread. It has 'knock on' effects for the rest of the project. That is, by continuing to use invalid patches a third machine must validate the two already reported results. This is a problem; basically what should be an alpha project with minimal effects on the 'stock' project is turning into a project where the stable clients are 'subsidizing' the optimized clients that are no good.

So...people check out the download page more frequently than the invalids. All invalid producing patches must be marked as such (if not removed).

How about a listserv for those of us using the optimized clients? That might be a good way to keep everyone up to date on what patches not to use anymore.

With some of these improvements implemented we might be able to cut down on the need for the stable clients to have to verify our invalid results:) It would also be helpful as the ultimate aim of this project is to produce the maximum number of valid results, which can even be done with the test patches as long as people are aware of a 'bad' patch.



Try the Pizza@Home project, good crunching.

Yin Gang
Yin Gang
Joined: 23 Feb 05
Posts: 52
Credit: 120187750
RAC: 0

The credit of S5 WU is

Message 39321 in response to message 39319

The credit of S5 WU is pre-assigned on server side.

Quote:

After finishing the work I had, I made an app_info.xml from the code Stick quoted, made sure I had a 4.10 named exectuable, and patched it up to S5T0712 which was the newest version I saw with valid results reported here. I got a few short work units when I turned on work requests again, and one has been sent in. It finished in less than 1 hour and was done in about 70% of the time as the first short work unit I did with the standard application -- nice improvement. It hasn't validated yet, but given the positive reports of the S5T0712 patch here, I expect it to be good. And it reported back as "E@H S5R1 4.10 0712 TEST", so the modified version number has been picked up correctly from app_info.xml.

Aside: It looks like Einstein has change the credit granting system so that it's more uniform and less dependent on time, benchmarks, or whatever. Does that mean I can turn off calibration by my client, report the real CPU time with the optimized application, and still be fine? I'd rather report the real crunching time if calibration isn't necessary to claim and get the right credit (which it was before). Thanks.


Welcome To Team China!

Terry
Terry
Joined: 9 Feb 05
Posts: 55
Credit: 1781475
RAC: 0

RE: Aside: It looks like

Message 39322 in response to message 39319

Quote:

Aside: It looks like Einstein has change the credit granting system so that it's more uniform and less dependent on time, benchmarks, or whatever. Does that mean I can turn off calibration by my client, report the real CPU time with the optimized application, and still be fine? I'd rather report the real crunching time if calibration isn't necessary to claim and get the right credit (which it was before). Thanks.

Credit is predetermined server side regardless of your claim, so no calibration is needed.

Terry

edit: Oops, looks like Yin Gang and I were giving the same answer at the same time.

Dronak
Dronak
Joined: 21 Mar 05
Posts: 28
Credit: 10402879
RAC: 0

Simultaneous answers happen

Simultaneous answers happen sometimes. :) Thanks, I just wanted to check and make sure before I changed anything. I'll have to see if calibration is actually needed on my other projects; I might be able to just dump the calibrating client all together.

My other two short work units should report back within a few hours. Maybe I'll have real validation status by tomorrow morning when I wake up. (It's almost bed time for me here.)

[B@H] Ray
[B@H] Ray
Joined: 4 Jun 05
Posts: 621
Credit: 49583
RAC: 0

I just started a email list

I just started a email list for the anouncement and discussion of optomized Einstein Applications. All are invited to join and share info there, finding things in the Archives (now empty as new) should be easier by the subject rather than searching through hundreds of postings in one thread.

You can get info on this and sign up at:
This link

Ray Brown


Try the Pizza@Home project, good crunching.

Nuadormrac
Nuadormrac
Joined: 9 Feb 05
Posts: 76
Credit: 229259947
RAC: 45

RE: RE: S5T0713.dat -

Message 39325 in response to message 39227

Quote:
Quote:

S5T0713.dat

- eliminated double jumps
- reduced amount of FPU macro ops
- removed double loads on general purpose registers

- better SSE register usage
- reduced memory and integer register usage
- optimized branch structure
- faster FPU comparisons

- SSE3 truncation
- some reordered instructions
- automatic SSE/SSE3 usage
- less FPU-memory-FPU operation
- shorter patch file :-)

CPU: ALL

app: einstein_S5R1_4.10_windows_intelx86.exe

And now a new ZIP file containing S5T0713...

more info some posts below...

thx for the zipped .exe and all :)

Metod, S56RKO
Metod, S56RKO
Joined: 11 Feb 05
Posts: 135
Credit: 826903095
RAC: 84910

S5T0711 won't run on AMD

S5T0711 won't run on AMD Opteron 280:

5.4.7

- exit code -1073741819 (0xc0000005)

2006-06-26 10:00:08.9896 [normal]: E@H S5R1 4.02 0711 TEST
2006-06-26 10:00:08.9896 [normal]: Started search at lalDebugLevel = 0
2006-06-26 10:00:09.6771 [normal]: Checkpoint-file 'Fstat.out.ckp' not found.
2006-06-26 10:00:09.6771 [normal]: No usable checkpoint found, starting from beginning.
Detected CPU type 1

Unhandled Exception Detected...

- Unhandled Exception Record -
Reason: Access Violation (0xc0000005) at address 0x00400400

S5T0709 seems to work.

Metod ...

Pepperammi
Pepperammi
Joined: 20 Feb 05
Posts: 131
Credit: 437943
RAC: 0

S5T0712- Valid and credit

S5T0712- Valid and credit granted.
For now it's still 4.02 till i can run my cache out.
http://einsteinathome.org/task/35150680

Standard: ~41,800
S5T0712: 27,518
About 35% faster

LiborA
LiborA
Joined: 8 Dec 05
Posts: 74
Credit: 337135
RAC: 0

RE: S5T0712- Valid and

Message 39328 in response to message 39327

Quote:

S5T0712- Valid and credit granted.
For now it's still 4.02 till i can run my cache out.
http://einsteinathome.org/task/35150680

Standard: ~41,800
S5T0712: 27,518
About 35% faster

You don't need waiting to cache out. You can edit your client_state.xml :
- all strings "einstein_S5R1_4.02_windows_intelx86" replace "einstein_S5R1_4.10_windows_intelx86"
- all strings "version_num>402410"

And of course put into E@H project directory files app_info.xml, einstein_S5R1_4.10_windows_intelx86.exe and einstein_S5R1_4.10_windows_intelx86.pdb

Example of WU which was edited by this technique is here

Pepperammi
Pepperammi
Joined: 20 Feb 05
Posts: 131
Credit: 437943
RAC: 0

RE: There are a few

Message 39329 in response to message 39316

Quote:

There are a few 'issues' with the threads on the new clients.
1.) The application/.dat thread is not being updated while new patches are released in this thread. The app/.dat thread also is not marking apps known to produce invalid results as such. That's an issue if people are downloading 'abandoned' apps - it's a waste of time.

I would say that that not marking patches that produce invalids is one of the biggest 'issues'. This is mostly because the patches are coming out so frequently.

2.) The app_info.xml is a major issue as it is not addressed in the app thread. It is addressed here and is helpful; the problem is that this thread has lots of posts and is probably not the greatest place to put instructions on how to modify files, etc. That is, since we're probably wanting a standardized setup and to mitigate the confusion (and arguments :)) it might be helpful to start updating the app/.dat thread.


I get the impression that Akofs is using that thread more as a public release of the 'Stable' apps only. Good idea really as not everyone wants to be a tester which is what we are doing in this thread. Testing all the latest patches to determine the 'stable' ones. When Akosf believes a specific patch can be considered stable he can release it for everyone in that thread (just a figure of speech- of course anyone can just come to this thread and download whichever they like) and they can use that 'stable' patch knowing it has been fully tested and should be safe and cause them no trouble.
I agree the app_info could be clarified by either Bernd or Akosf and let us all know exactly what them mean and what they would like us to do more clearly. Why i asked

Quote:

Once we get all this sorted it might be a good idea to start this thread afresh. Starting with a new set of instructions and fresh links to the latest stable/test app and a known working app_info.xml file. Basically all that’s needed to continue from here with as little confusion as possible. This thread is getting very long too.

I'd prefer if I could ask that Akosf be the one to decide when its time to do this when he's decided what the next 'stable' release is maybe, and when he and Bernd have sorting things out in the background.

Quote:

but i doubt we at that stage yet and yea your right it would be nice if people could edit their posts to include info on invalid patches but sometimes its too late so maybe they could remove the files wherever so cant be downloaded?

Mush easier-if your worried about getting invalids then stick to the latest 'Stable' releases only. They're updated fairly regularly and you can be sure there safe. Or stay a few patches behind the latest 'beta' patch and you'll get responses from people that have tried them informing if they get valids or invalids. These are all just suggestions. There's always risk when running test patches/apps.

PS Sorry for the very long post.

Comment viewing options

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