<html>
<head>
<meta content="text/html; charset=UTF-8" http-equiv="Content-Type">
</head>
<body text="#000000" bgcolor="#FFFFFF">
Am 19.01.2013 19:24, schrieb Filippo Spiga:
<blockquote
cite="mid:BBA2C4C0-4278-433D-871D-1C4A9FD1A53E@gmail.com"
type="cite">
<meta http-equiv="Content-Type" content="text/html; charset=UTF-8">
<div>You mean "OpenMP-only parallelization gives low performance
than 4 MPI without OpenMP"?</div>
</blockquote>
Yes.<br>
<blockquote
cite="mid:BBA2C4C0-4278-433D-871D-1C4A9FD1A53E@gmail.com"
type="cite">
<div> You mentioned ATLAS in a previous email. If you have Intel
compiler you also have Intel MKL library...</div>
</blockquote>
1. I had no Intel compiler when I built that version and knew a far
less than now about high performance computing.<br>
2. There are copyright limitations on that Intel software so I am
also interested in alternative tools to avoid them.<br>
<blockquote
cite="mid:BBA2C4C0-4278-433D-871D-1C4A9FD1A53E@gmail.com"
type="cite">
<div>In case of QE-GPU, having 4 MPI on a multi-core workstation
means share the GPU among them and split by a factor of the
number of the MPI processes the available RAM on the GPU board
(in your case 4). I do not know what kind of GPU do you have but
you would like to avoid this scenario.<br>
</div>
</blockquote>
I use GeForce GTX 460 with 1Gb of memory onboard and I definitely
want to avoid this scenario. I've just rebuilt QE.<br>
<br>
And, probably, I should leave the heaviest tasks for x86-64 system
with more memory...<br>
<br>
Thanks again for clearing the situation.<br>
<br>
Anton S. Lytvynenko,<br>
<br>
L.V.Pisarzhevskii Institute of Physical Chemistry of the National
Academy of Sciences of Ukraine.<br>
<br>
</body>
</html>