very slow performance

I just installed pgi accelerator and cuda5.5 and noticed that the tes programs are very slow. Even the smallest problem takes about 12-13 secs.

I notived that pgaccelinfo -v also runs about 12 sec. What can I do?

There are teslas in the machine (m2090) and I use nvidia 319.37 driver.
Accelinfo gives me many lines with the proper information of the cards, but the last two lines are libamdcalcl.so not found
libcoi_host.so not found
Maybe it has some connection to my problem.
What do you think?

What happens if you run “pgcudainit” in the background or from another shell?

On Linux, the OS will power down devices when not in use and takes some time to power them back up. “pgcudainit” will hold them open and not let the OS power them down.

However, 12 seconds seems a bit high. I typically see in the range of 1-2 seconds per device. Then again, it could be different on your system.

  • Mat

Mat it was a great idea
now pgaccelinfo -v runs 10 times faster (about 1.2 secs) with pgcudainit in the backgroud.
In this machines there are 4 tesla cards, maybe it explains the extra high time but now 1.2 sec seems okay for me.
I think it is acceptable to have some delay when I turn to other device.
I cannot find any information pgcudainit. Please give me some URL or explain how to run this in the background in a reliable way. Should I start it from system boot with root privileges?


My other question is still not answered:
What do you think about these two lines in the output of pgaccelinfo?
libamdcalcl.so not found
libcoi_host.so not found


Thanks, David

Hi David,

See Chapter 7 of the PGI User’s Guide.

Though, I’m don’t think I’d recommend having it set when the system boots, since this may increase the power consumption of the system when the GPUs aren’t in use. While 12 seconds warm-up time is a lot when you’re running small jobs, for bigger jobs it’s inconsequential.

For the two shared libs not found, pgaccelinfo is just testing if you have other devices installed by looking for the dependent libraries.

  • Mat