Windows S5R3 SSE power App 4.36 available

Brian Silvers
Brian Silvers
Joined: 26 Aug 05
Posts: 772
Credit: 282,700
RAC: 0

Just started up a task with

Just started up a task with 4.36. Seems to be working fine. Graphics are fine, although I didn't switch during a task like some of the rest of you did...

Projections based on the first few minutes of runtime show a shorter completion than any of my previous results processed with version 4.32, but definitely not 20% shorter. Looks right now like about 10%, but this is only an early projection... By the time I know for sure, the result will have finished...as I'm going to bed soon...

Winterknight
Winterknight
Joined: 4 Jun 05
Posts: 482
Credit: 194,922,894
RAC: 113,808

Second unit finished, a few

Second unit finished, a few seconds longer to crunch.
Downloaded and started as 4.36 graphics work here.

Conan
Conan
Joined: 19 Jun 05
Posts: 162
Credit: 5,937,946
RAC: 0

It happens that I have

It happens that I have switched over whilst I am doing frequency 903.85 and sequence 420 (about 50% done with 4.32). This happens to be the trough of the cycle I am running down.

Sequence unit 421 done on 4.32 was under 27,000 seconds, which I was happy with, if this goes faster still I will be VERY happy.

A 10% improvement on say 26,500 seconds will lower the time to 23,850 seconds and a 20% improvement will lower to 21,200 seconds, so I will wait and see how it goes.

I will very happy with 10% improvement.
It also means my peak times are much lower as well.

Congratulations to the Einstein Project Team for the continued improvement to the programmes we are running for them.

Jord
Joined: 26 Jan 05
Posts: 2,952
Credit: 5,779,100
RAC: 0

First estimates show my new

First estimates show my new task to go run in just under 12 hours. The one started with 4.26 (continued with 4.32) ran in 17 hours. The one started with 4.32 (continued with 4.36) ran in 15 hours. I have good hopes.

Richard Haselgrove
Richard Haselgrove
Joined: 10 Dec 05
Posts: 2,041
Credit: 746,086,017
RAC: 1,147,307

RE: First estimates show my

Message 79629 in response to message 79628

Quote:
First estimates show my new task to go run in just under 12 hours. The one started with 4.26 (continued with 4.32) ran in 17 hours. The one started with 4.32 (continued with 4.36) ran in 15 hours. I have good hopes.


First estimates from Mike's Ready Reckoner:

App 4.32:

		Estimated peak runtime = 28911
		Estimated average runtime = 24106
		Estimated trough runtime = 21363
		Estimated runtime variance = 0.261


App 4.36:

		Estimated peak runtime = 20473
		Estimated average runtime = 19143
		Estimated trough runtime = 18384
		Estimated runtime variance = 0.102


Same host, same frequency.

Jord
Joined: 26 Jan 05
Posts: 2,952
Credit: 5,779,100
RAC: 0

Bernd, I build my own

Bernd,

I build my own BOINC 5.10.43 with all the latest fixes. Copied over my BOINC data directory to a testing directory. Put the BOINC 5.10.43 files in it and started running it. Then checked the graphics.

In 5.10 I see everything in the graphics, so including my own data, the clock and the search information, thereby excluding my video card to be the culprit. As a conclusion, it's BOINC 6.1.8 which doesn't play nice yet.

Bernd Machenschalk
Bernd Machenschalk
Moderator
Administrator
Joined: 15 Oct 04
Posts: 4,031
Credit: 218,232,376
RAC: 49,769

RE: Bernd, I build my own

Message 79631 in response to message 79630

Quote:

Bernd,

I build my own BOINC 5.10.43 with all the latest fixes. Copied over my BOINC data directory to a testing directory. Put the BOINC 5.10.43 files in it and started running it. Then checked the graphics.

In 5.10 I see everything in the graphics, so including my own data, the clock and the search information, thereby excluding my video card to be the culprit. As a conclusion, it's BOINC 6.1.8 which doesn't play nice yet.


I see. For the bug report: there must be a "key" the BOINC core client writes into the "boinc init data" that tells the the graphics app where to find the shared memory segment with the progress information. Looks like this is broken (or at least different) in the 6.1.8 core client.

BM

BM

Jord
Joined: 26 Jan 05
Posts: 2,952
Credit: 5,779,100
RAC: 0

RE: For the bug report:

Message 79632 in response to message 79631

Quote:
For the bug report: there must be a "key" the BOINC core client writes into the "boinc init data" that tells the the graphics app where to find the shared memory segment with the progress information. Looks like this is broken (or at least different) in the 6.1.8 core client.


Not in any of the xml files. I have just about checked them all and can't find anything different between them for Einstein. I'm happy to send you the init_data.xml files and client_state.xml files for you to scrutinize.

But I'm pretty sure it's something in the way that boinc.exe executes the graphics.

Else just tell me what you need and I'll zip it up and email it to you, per BOINC version.

archae86
archae86
Joined: 6 Dec 05
Posts: 3,021
Credit: 4,995,557,190
RAC: 3,013,716

RE: RE: First estimates

Message 79633 in response to message 79629

Quote:
Quote:
First estimates show my new task to go run in just under 12 hours. The one started with 4.26 (continued with 4.32) ran in 17 hours. The one started with 4.32 (continued with 4.36) ran in 15 hours. I have good hopes.

First estimates from Mike's Ready Reckoner:

App 4.32:

		Estimated peak runtime = 28911
		Estimated average runtime = 24106
		Estimated trough runtime = 21363
		Estimated runtime variance = 0.261

App 4.36:
		Estimated peak runtime = 20473
		Estimated average runtime = 19143
		Estimated trough runtime = 18384
		Estimated runtime variance = 0.102

Same host, same frequency.


I've been wondering whether 4.36 gives a big reduction in variance. Your estimate suggests this is spectacular. (.261 going down to .102)

This morning I did some selective pausing to move a few results pretty near peaks into play. With that aid to accuracy, my E6600 shows:

          Estimated peak runtime = 16305
		Estimated average runtime = 14697
		Estimated trough runtime = 13779
		Estimated runtime variance = 0.155

for 4.36, compared to

		Estimated peak runtime = 23119
		Estimated average runtime = 19229
		Estimated trough runtime = 17008
		Estimated runtime variance = 0.264

for 4.32 which is almost a 31% improvement in estimated productivity, and considerable reduction, in variance, though not to so low a level as the estimate you show.

For my Q6600:

		Estimated peak runtime = 16923
		Estimated average runtime = 15232
		Estimated trough runtime = 14266
		Estimated runtime variance = 0.157

for 4.36

		Estimated peak runtime = 23909
		Estimated average runtime = 19926
		Estimated trough runtime = 17652
		Estimated runtime variance = 0.262

for 4.32

I have some reservations about the model fit, especially for the Q6600, where there is more noise in the CPU times than is good for Ready Reckoner's digestion, but I have no reservations about saying that for Conroe-type Windows XP hosts 4.36 is a major improvement over 4.32, by rather more than 20% on the average, and possibly near or even over the upper end of my first estimate of 20 to 30% productivity per hour improvement, with a considerable reduction of variance from recent versions.

Bikeman (Heinz-Bernd Eggenstein)
Bikeman (Heinz-...
Moderator
Joined: 28 Aug 06
Posts: 3,516
Credit: 471,006,529
RAC: 54,275

RE: I have some

Message 79634 in response to message 79633

Quote:

I have some reservations about the model fit, especially for the Q6600, where there is more noise in the CPU times than is good for Ready Reckoner's digestion, but I have no reservations about saying that for Conroe-type Windows XP hosts 4.36 is a major improvement over 4.32, by rather more than 20% on the average, and possibly near or even over the upper end of my first estimate of 20 to 30% productivity per hour improvement, with a considerable reduction of variance from recent versions.

Thanks for the comparison data!!

A considerable reduction of variance is exactly what was hoped for and expected by one of the improvements that went into 4.36. Basically, there are two relatively tiny parts of the science app that consume the vast majority of processing time. Of these two "hot-spots" or "hot-loops", one has constant processing time while the second one is more data dependent and causes this cyclic runtime variations.

The new improvements in 4.36 speed up the second hot-loop more than the first, so the variance should drop. Because of this, the speed-up of 4.36 over 4.32 will not be a uniform factor, but will be smaller for "runtime trough" tasks and greater for "peak runtime" tasks.

CU
Bikeman

Comment viewing options

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