Hi, I have taken an testing tool from Simons "pre-release" section and modified it to test the real apps on Chirp limited wus (makes them run shorter times by eliminating repetitive operations, as I understand it from Joe Segur). I hope some of you AMD users find this helpful in you're selection of which apps to try and which not to waste your time on.

OK, from this post until notified, here are the conditions used for the upcoming charts. I have editted the batch file to run the quick test for 10 seti applications (8 on 32b machines, 10 on 64 bit OSes), and 13 "quick"(chirp limited) different wus. The file runs one wu through each application, and records the stop/start time of each event (test). I take the "run time" files and copy and paste them to a chart, then create a chart for "elapsed time" of those runs and chart those. I emboldened the shortest times on that chart per WU ran.

Computers run:
AMD64 2800 Clawhammer., 32b OS
Mobile AMD64 3700, 32b OS
AMD64 3700 Sandiego, 32b OS
AMD64 X2 4800 Toledo, 32b OS
AMD64 X2 5200 Windsor, 64b OS
AMD64 X2 6000 Windsor, 64b OS

Applications tested:
From Simon:
KWSN_2.4_MMX_MB
KWSN_2.4_SSE_MB
KWSN_2.4_SSE2-AMD_MB (32bit)
KWSN_2.4B_SSE2-generic_Ben-Joe(32bit, Loadstore adjust modified 2.2B)
KWSN_2.4_SSE2_IPP_Ben-Joe (64 bit)
From Crunch3r
KWSN_2.4V_MMX_MB
KWSN_2.4V_SSE_MB
KWSN_2.4V_SSE2_MB
KWSN_2.4V_SSE2_MB_64 (The real name of this app matches the 32bit one so I added the -64 to show it as 64 bit)
and for a standard I used this app from Seti
setiathome_5.27_windows_intelx86

I checked to make sure that 32b was running 32b and 64b was running as 64b on both 64b windows systems

The WUs' are:(taken from Simons test facility)
testWU-1-0.60488405578199, Linefeed wu with AR 0.60488405578199
testWU-2-0.4437317022742, Linefeed wu with AR 0.4437317022742
testWU-3-0.42587682223996, Linefeed wu with AR 0.42587682223996
testWU-4-1.2796485198966, Linefeed wu with AR 1.2796485198966
testWU-5-0.43995685848661, Linefeed wu with AR 0.43995685848661
testWU-6-0.033857570580327, Linefeed wu with AR 0.033857570580327
testWU-7-0.775, Linefeed wu with AR 0.775
testWU-8-FM0.017.wu, MultiBeam wu with AR 0.017
testWU-9-FM0.369.wu, MultiBeam wu with AR 0.369
testWU-10-FM0.446.wu, MultiBeam wu with AR 0.446
testWU-11-FM0.828.wu, MultiBeam wu with AR 0.828
testWU-12-FM1.181.wu, MultiBeam wu with AR 1.181
testWU-13-FM7.766.wu, MultiBeam wu with AR 7.766

Since these will be finishing at different times I'll start with the AMD64 X2 6000, and add data from other computers as they become available. There'll be two charts for each computer. One chart of the "stop/start" times, and one of the "elapsed times" w/emboldened fastest times (also emboldened where ties for fastest). Both charts are sorted by Angle Range for easier interpretation. The elapsed time will be the most useful. Also there'll be a total elapsed time (how long testing took) displayed.

I'd ask that other posts not be made until the conclusion of the test and a summary post has been shown. It'll keep the data neater and closer together. Thanks, Tony

This thread will end up being very graphic intensive, and will widen the screens for those with 800/600 screen res or lower. Sorry.





Please hold all comments and questions until after I get the rest of the data up tomorrow morning PLEASE

Thanks for not posting before I got a chance to make these additional 4 charts. You may resume posting. I was thinking about a summary, but not sure what I'd say just yet. LOL

Thanks for not posting before I got a chance to make these additional 4 charts. You may resume posting. I was thinking about a summary, but not sure what I'd say just yet. LOL

Nice work Astro. Look forward to seeing other cpu charts.

Regards

PJ

---

to make this easier to see, I cut out the stop/start and put all the elapsed time into one chart. Listed from slowest to fastest puter.

I pruned out the apps which didn't show any "fastest" times for any wu. Narrowed the columns and eliminated redundant rows. (this should be more easily viewed)

Thanks for the effort.

Could you make a .CSV or .XLS file(s) available online somewhere ?

Maybe use a few more colors too as the shade of green you have used is possibly not the best for easy reading.

Thanks, Keith

Thanks for the effort.

Could you make a .CSV or .XLS file(s) available online somewhere ?

Maybe use a few more colors too as the shade of green you have used is possibly not the best for easy reading.

Thanks, Keith

I can email you it, or anyone. PM me with your addy, and it's on the way. Thanks for the color suggestion. The test tools and such are just too large to email, and I don't have a way to host anything.

Now to ponder how it can be displayed differently, or if it's needed at all.

And if possible, a "top" line on each CPU / WU showing the time to complete with the stock s@h app.

This would be great since there have been posts about there being little or no improvement using the mmx/sse apps; but big jumps in the sse2, etc. apps. Also, it would show everyone how much of a crunching improvement over the stock app you can get using the optimized.

This is just a request if someone has the time.

Great job Astro! - I wish I had the time to do this sort of thing...

---

I gotta fever and the only prescription is more cowbell.

This any better? Best times are red, legend is some other funky color.

Tony,
Thanks for the info, and for the time and effort you have put in.

I don't know wether the results show anything or not. But I don't think I we could ask you to do more.
Obviously the different chirp limits for the test WU's have played their part in my confusion. But if I owned an AMD cpu I think, on these tests the KWSN_2.4_SSE2_AMD_MB app has it with its better performance at VLAR.
But we probably haven't run MB long enough to say wether we are going to get enough VLAR's to make any difference.

Thanks again,

Andy

P.S. I haven't got the nerve to ask you to do more tests with your P60.

I agree on the KWSN_2.4_SSE2_AMD-MB choice for the following reasons:

1) The KWSN_2.4B_SSE2-generic_Ben-Joe isn't available to the public.

2) KWSN_2.4V_SSE2_MB performs nearly the same in all wus to KWSN_2.4B_SSE2-generic_Ben-Joe's version.

3) the difference between those two and KWSN_2.4_SSE2_AMD-MB isn't significant across the middle and high ends of the angle range, but KWSN_2.4_SSE2_AMD-MB is noticeably faster on the low end of the angle range.

I guess it comes down to the AR's being issued ATM. Anyone wanna right a program to change apps based upon the AR of a WU??? LOL

One thing is for sure, NEVER use the 64b windows apps on an AMD system.

I agree on the KWSN_2.4_SSE2_AMD-MB choice for the following reasons:

1) The KWSN_2.4B_SSE2-generic_Ben-Joe isn't available to the public.

2) KWSN_2.4V_SSE2_MB performs nearly the same in all wus to KWSN_2.4B_SSE2-generic_Ben-Joe's version.

3) the difference between those two and KWSN_2.4_SSE2_AMD-MB isn't significant across the middle and high ends of the angle range, but KWSN_2.4_SSE2_AMD-MB is noticeably faster on the low end of the angle range.

I guess it comes down to the AR's being issued ATM. Anyone wanna right a program to change apps based upon the AR of a WU??? LOL

One thing is for sure, NEVER use the 64b windows apps on an AMD system.

Agreed, UNLESS the system is running Windows Vista where KWSN_2.4_SSE2_AMD-MB will fail, so KWSN_2.4V_SSE2_MB should be used.

The puzzle now is why on AMD system which don't have SSE2 the optimized SSE versions have been reported as no quicker than stock. Both KWSN_2.4_SSE_MB and KWSN_2.4V_SSE_MB were significantly quicker than stock on these AMD 64 systems. If anyone with an older AMD system running Windows wants to run some quick tests like these, I will gladly supply a test package tailored for the purpose.

                                                               Joe

The puzzle now is why on AMD system which don't have SSE2 the optimized SSE versions have been reported as no quicker than stock. Both KWSN_2.4_SSE_MB and KWSN_2.4V_SSE_MB were significantly quicker than stock on these AMD 64 systems. If anyone with an older AMD system running Windows wants to run some quick tests like these, I will gladly supply a test package tailored for the purpose.

If nobody else is interested in running that test I could run it.

There is just one thing. I have only this machine so it is loaded with all the programs I use daily and of course those programs take their share of cpu cycles. Now I'm afraid that those other programs may affect the test run if they for whatever reason want to do something. For example AV decides to update itself. So, does anyone know any software for Windows that can tell the cpu time an app used? Something like 'time' in Linux.

-Juha

---

The puzzle now is why on AMD system which don't have SSE2 the optimized SSE versions have been reported as no quicker than stock. Both KWSN_2.4_SSE_MB and KWSN_2.4V_SSE_MB were significantly quicker than stock on these AMD 64 systems. If anyone with an older AMD system running Windows wants to run some quick tests like these, I will gladly supply a test package tailored for the purpose.

                                                               Joe

Is my Athlon XP3200+ of any use for this test?

Many thanks for the time & effort Astro.

Saving thread offline for future reference.

---

Astro,
That's some great testing and results you've completed. It seems clear the KWSN_2.4_SSE2-AMD_MB (32bit) is probably the best (on average). The next step would be to run NHST on the values (against the standard app). I hypothesize there's not a significant difference between any KWSN SSE2 versions except compatibility perhaps.

Missing from this batch of tests was the AMD Opteron CPU of any class.

... So, does anyone know any software for Windows that can tell the cpu time an app used? Something like 'time' in Linux.

-Juha

There is Windows task manager.

Right click on the task bar, select Task Manager
Then select 'Processes' tab
Then click on 'View' on the menu bar
In 'View' click on 'Select Columns...'
Then put a check mark in 'CPU Time'
Click on 'OK' button.

Now your should see the CPU run time for all displayed tasks.

---

SL

... So, does anyone know any software for Windows that can tell the cpu time an app used? Something like 'time' in Linux.

-Juha

There is Windows task manager.

An alternative is Process Explorer, which used to be independent, but got acquired by Microsoft. It is GUI-oriented, which you may not like, but it does offer some pretty interesting configurability, included recent history of percent CPU use by process. I like it specifically for watching over BOINC operations. Despite the fact that Microsoft offers it now, installation is a snap, and, so far as I know, it is robust.

---

Astro,
That's some great testing and results you've completed. It seems clear the KWSN_2.4_SSE2-AMD_MB (32bit) is probably the best (on average). The next step would be to run NHST on the values (against the standard app). I hypothesize there's not a significant difference between any KWSN SSE2 versions except compatibility perhaps.

Missing from this batch of tests was the AMD Opteron CPU of any class.

Thanks, NHST? I thought about taking the full length MB wus from simon's site, but after examination, they're all "mid range", so I doubt it show as much about the low and hi ends. I ran it on the AMD's I have, but am willing to accept any XP and/or Opteron offered (would also like a Barcelona, but I don't suppose those come gift wrapped just yet)

The puzzle now is why on AMD system which don't have SSE2 the optimized SSE versions have been reported as no quicker than stock. Both KWSN_2.4_SSE_MB and KWSN_2.4V_SSE_MB were significantly quicker than stock on these AMD 64 systems. If anyone with an older AMD system running Windows wants to run some quick tests like these, I will gladly supply a test package tailored for the purpose.

If nobody else is interested in running that test I could run it.

There is just one thing. I have only this machine so it is loaded with all the programs I use daily and of course those programs take their share of cpu cycles. Now I'm afraid that those other programs may affect the test run if they for whatever reason want to do something. For example AV decides to update itself. So, does anyone know any software for Windows that can tell the cpu time an app used? Something like 'time' in Linux.

-Juha

I put together a package for testing SSE systems, SSEBENCH.exe. That's a self extracting 7zip archive which expands to a SSEBENCH folder and several subfolders. The test script is knabench_vnog.cmd. The reference app is stock 5.27, test apps are the 2.4, 2.4B (patched 2.2B), and 2.4V optimized SSE apps. It's set to run 3 test WUs, those Tony called WU8, WU10, and WU12.
If you're running BOINC as a service, starting the test script will shut BOINC down with a "Net stop", and it will be started again at the end. If you're running it otherwise, do whatever is appropriate to turn BOINC off during the test.

Juha, I think the test would take roughly 3 hours on your system, if you could pick a time period when not much else is likely to interfere or disable some things which might, the results should be OK. There is a method of getting CPU time from the init_data.xml file after a run, but what gets entered in that file is tied to the checkpoint period so varies too much for these quick tests. We basically just run with very little else and use elapsed time to compare.

Robert, your Athlon XP3200+ will probably run the test in about 1.5 hours if you care to try.

When the test completes, there will be a file in the Testdatas subdirectory with a name starting with the day, date, and time and ending bench.txt which contains a lot of detail. Near the end is a section called Quick timetable which has the summary of relative timing for each WU, perhaps you might want to copy and paste that part in a post here.

If you want a longer test, the test WUs Tony called WU9, WU11, and WU13 are in the TestWUs\\Reserve subdirectory, copying them up to TestWUs would make the script run them also and approximately double the test time.

                                                               Joe