This should fix the slowdown on Windows if you have a high CPU load.


You can now configure the polling mode sort of like what the old one had.

The --gpu-target-frequency is similar to the -f flag the old one had. The default is 30 hz. This can be used in place of the --responsiveness-factor one I added in the last minor release (or in addition to, although I don't know why you would want to. I would recommend using this one instead and I'll probably remove the other at some point).

The --gpu-polling-mode is similar to the -b flag the old one had. A negative integer will use busy waiting and have a high CPU load (like what always happened in 0.58). 0 will use the same method that was used in 0.59. A positive integer > 0 sets the polling frequency in milliseconds. The default now is to poll every 1 ms which seems to have solved the slowdowns with a high cpu load without increasing the cpu usage much.

Update: Now at 0.62 since I'm good at screwing up

Hello Matt,

thanks for the new version. First tasks ran fine on Win7 64-Bit 2x 5850 (Catalyst 11.3) with high CPU-load from CPU-projects. Load from 0.60 seems to be low, only higher at the end of run. Runtime of the tasks are about the same compared to 0.23 only 2 or 3 seconds higher. No windows lags.

Looking for the next tasks, but it looks like you done a good job.

---

The --gpu-target-frequency <number> is similar to the -f flag the old one had. The default is 30 hz.

How should this be set. I tried

<cmdline>--gpu-target-frequency 40</cmdline>

in the app_info.xml. I verified that --gpu-target-frequency 40 is passed to the app using Process Explorer.
However in the finished results I find "Target frequency 30.000000 ms, polling mode 1, using responsiveness factor of 1.000000".

On the other side I can confirm that GPU load is now high (98% up) with full load on all CPU cores. It's way faster than 0.59 :)

---

Starfire

Hello.
Any of the new tasks 0.60 run a less longer like this:
Specialy the 10_3s_fix_20 runs less than 4min. for now.
I have 2 5850 GPU's and Aqua runs on 6 cores.
Win7 64bit,12GB RAM,Boinc 6.10.60
Greetings.

I now had a similar task like Milchband. It was de_separation_10_3s_fix10_1_463621_1302635252_1. While it was running the UI was really sluggish and the Target frequency in the result is strange. Currently another de_separation_10 is running and the UI is a lot smoother.

<core_client_version>6.12.22</core_client_version>
<![CDATA[
<stderr_txt>
<search_application> milkywayathome_client separation 0.60 Windows x86 double CAL++ </search_application>
Found 1 CAL devices
Chose device 0

Device target:         CAL_TARGET_770
Revision:              2
CAL Version:           1.4.1332
Engine clock:          750 Mhz
Memory clock:          900 Mhz
GPU RAM:               1024
Wavefront size:        64
Double precision:      CAL_TRUE
Compute shader:        CAL_TRUE
Number SIMD:           10
Number shader engines: 1
Pitch alignment:       256
Surface alignment:     4096
Max size 2D:           { 8192, 8192 }

Estimated iteration time 284.281667 ms
Target frequency 5735634568500407000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000.000000 ms, polling mode 1, using responsiveness factor of 1.000000
Dividing into 1 chunks
Integration range: { nu_steps = 640, mu_steps = 1600, r_steps = 1400 }
Using { 1, 1 } chunk(s) of size { 1400, 1600 }
Integration time = 199.189885 s, average per iteration = 311.234195 ms
Integral 0 time = 200.606518 s
Estimated iteration time 71.070417 ms
Target frequency 5735634568500407000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000.000000 ms, polling mode 1, using responsiveness factor of 1.000000
Dividing into 1 chunks
Integration range: { nu_steps = 640, mu_steps = 400, r_steps = 1400 }
Using { 1, 1 } chunk(s) of size { 1400, 400 }
Integration time = 53.654794 s, average per iteration = 83.835615 ms
Integral 1 time = 54.111614 s
Likelihood time = 5.941989 s
<background_integral> 0.000331750494935 </background_integral>
<stream_integral>  109.619376970424070  1951.519873483356200  266.117270577347310 </stream_integral>
<background_likelihood> -3.051034779156822 </background_likelihood>
<stream_only_likelihood>  -4.832203828758965  -4.439834432194514  -4.548461163928033 </stream_only_likelihood>
<search_likelihood> -2.824976493982051 </search_likelihood>
22:26:05 (5984): called boinc_finish

</stderr_txt>

Edit:
Target frequency looks normal for the second task (even through the setting to 40 seems to be ignored).
Target frequency 30.000000 ms, polling mode 1, using responsiveness factor of 1.000000

---

Starfire

Hi Matt,

2671709 1444283 198526 12 Apr 2011 | 20:28:53 UTC 12 Apr 2011 | 20:32:20 UTC Completed and validated 135.33 8.60 267.20 MilkyWay@Home v0.60 (ati14)

This is what I like to see, only 8.60 secs of CPU use.
Thank you for your effort on this issue.

---

Greetings from,
TJ

Well done Matt! It's looking good!

A big thanks to Matt Arsenault!

Great work!

Looking very good here but even after updating the app_info file, I can't get it to work. I'm doing something wrong. Any ideas?

---

Target frequency looks normal for the second task (even through the setting to 40 seems to be ignored).
Target frequency 30.000000 ms, polling mode 1, using responsiveness factor of 1.000000

This was really dumb. Should be fixed now with 0.62

I'll try to add checkpointing for this and do the end likelihood on the GPU (which will stop the CPU spike at the end and save a few seconds) when I get the chance (which will probably be this weekend / sometime next week)

Thanks for the perseverance guys!

---

Should be fixed now with 0.62

Thumbs up. The 40 Hz are there now :)

---

Starfire

Looking very good here but even after updating the app_info file, I can't get it to work. I'm doing something wrong. Any ideas?

If your talking about the frequency parameter - it wasn't working in 0.60 but in 0.62 it is fixed.

PS: Your second computer errors out all tasks complaining about missing milkyway_0.60_windows_intelx86__ati14.exe. When using app_info.xml file you need to put it in the project folder manually.

---

Starfire

Yes, it refuses to download the new app. I think it has finally downloaded it now.

Edit: I was working on .60, didn't see the .62 one until you mentioned it. All is working well now.

---

MW is running fine Matt, but what I see is that the overall performance of the pc is slowed down. The letters are slower on the screen than I type them. Also responces of Internet pages are slower then with the 0.23 app and even with the 0.57 and 0.59 apps.

Can you look at this too over the weekend?
Or is it something only I noticed?

---

Greetings from,
TJ

Mine seems pretty responsive, maybe a touch slower than the old app. Overall, I am pleased so far.
PS I don't have the latest/greatest hardware out.

---

Thumbs up for the 0.62 tasks!! I can now run them on three machines at 95% GPU activity + 2 x Seti CPU tasks as well. Virtually back to where we were with the 0.23 apps.

AND I can access all three machines remotely without them locking up any more. Well done guys :-)

---

Don't drink water, that's the stuff that rusts pipes

MW is running fine Matt, but what I see is that the overall performance of the pc is slowed down. The letters are slower on the screen than I type them. Also responces of Internet pages are slower then with the 0.23 app and even with the 0.57 and 0.59 apps.

Can you look at this too over the weekend?
Or is it something only I noticed?

You can use the options I mentioned to tune this. It tries to pick something to prevent too much lag, but it's based on some rough estimates. If you use the --gpu-target-frequency option with something over 30, things will be responsive but a bit slower.

Mine seems pretty responsive, maybe a touch slower than the old app. Overall, I am pleased so far.
PS I don't have the latest/greatest hardware out.

It is slightly slower. Most of the slowdown is since the likelihood at the end isn't currently done on the GPU which I'm planning on fixing.