<html>
  <head>
    <meta content="text/html; charset=ISO-8859-1"
      http-equiv="Content-Type">
  </head>
  <body bgcolor="#FFFFFF" text="#000000">
    <div class="moz-cite-prefix">On 12/12/2013 04:38 PM, Paul Mullowney
      wrote:<br>
    </div>
    <blockquote
cite="mid:CAMJ8fwqkNMCFAPu0PCEoOAuoCx4C1xpjrU2R0_9bm+Gm6TssBw@mail.gmail.com"
      type="cite">
      <div dir="ltr">Provided you have a good parallel sparse direct
        solve for a single SM, you could unleash 32 direct solves (or
        perhaps 16) which run concurrently on the K20x. One only needs
        to set an environment variable to use Hypre Q</div>
    </blockquote>
    On Titan all you need to do is<br>
    <meta http-equiv="content-type" content="text/html;
      charset=ISO-8859-1">
    <table style="margin: 0px !important; padding: 0px !important;
      border-collapse: collapse; border-spacing: 0px; box-shadow: none
      !important; border-top-left-radius: 0px !important;
      border-top-right-radius: 0px !important;
      border-bottom-right-radius: 0px !important;
      border-bottom-left-radius: 0px !important; background-image: none
      !important; border: 0px !important; bottom: auto !important;
      float: none !important; height: auto !important; left: auto
      !important; line-height: 1.1em !important; outline: 0px
      !important; overflow: visible !important; position: static
      !important; right: auto !important; text-align: left !important;
      top: auto !important; vertical-align: baseline !important; width:
      706px; box-sizing: content-box !important; font-family: Consolas,
      'Bitstream Vera Sans Mono', 'Courier New', Courier, monospace
      !important; font-weight: normal; font-style: normal; font-size:
      13px; direction: ltr !important; color: rgb(0, 0, 0);
      font-variant: normal; letter-spacing: normal; orphans: auto;
      text-indent: 0px; text-transform: none; white-space: normal;
      widows: auto; word-spacing: 0px; -webkit-text-stroke-width: 0px;
      background-position: initial initial !important;
      background-repeat: initial initial !important;" border="0"
      cellpadding="0" cellspacing="0">
      <tbody style="margin: 0px !important; padding: 0px !important;
        border-top-left-radius: 0px !important; border-top-right-radius:
        0px !important; border-bottom-right-radius: 0px !important;
        border-bottom-left-radius: 0px !important; background-image:
        none !important; border: 0px !important; bottom: auto
        !important; float: none !important; height: auto !important;
        left: auto !important; line-height: 1.1em !important; outline:
        0px !important; overflow: visible !important; position: static
        !important; right: auto !important; text-align: left !important;
        top: auto !important; vertical-align: baseline !important;
        width: auto !important; box-sizing: content-box !important;
        font-family: Consolas, 'Bitstream Vera Sans Mono', 'Courier
        New', Courier, monospace !important; font-weight: normal
        !important; font-style: normal !important; font-size: 1em
        !important; direction: ltr !important; background-position:
        initial initial !important; background-repeat: initial initial
        !important;">
        <tr style="margin: 0px !important; padding: 0px !important;
          border-top-left-radius: 0px !important;
          border-top-right-radius: 0px !important;
          border-bottom-right-radius: 0px !important;
          border-bottom-left-radius: 0px !important; background-image:
          none !important; border: 0px !important; bottom: auto
          !important; float: none !important; height: auto !important;
          left: auto !important; line-height: 1.1em !important; outline:
          0px !important; overflow: visible !important; position: static
          !important; right: auto !important; text-align: left
          !important; top: auto !important; vertical-align: baseline
          !important; width: auto !important; box-sizing: content-box
          !important; font-family: Consolas, 'Bitstream Vera Sans Mono',
          'Courier New', Courier, monospace !important; font-weight:
          normal !important; font-style: normal !important; font-size:
          1em !important; direction: ltr !important;
          background-position: initial initial !important;
          background-repeat: initial initial !important;">
          <td class="code" style="margin: 0px !important; padding: 0px
            !important; border-top-left-radius: 0px !important;
            border-top-right-radius: 0px !important;
            border-bottom-right-radius: 0px !important;
            border-bottom-left-radius: 0px !important; background-image:
            none !important; border: 0px !important; bottom: auto
            !important; float: none !important; height: auto !important;
            left: auto !important; line-height: 1.1em !important;
            outline: 0px !important; overflow: visible !important;
            position: static !important; right: auto !important;
            text-align: left !important; top: auto !important;
            vertical-align: baseline !important; width: 706px;
            box-sizing: content-box !important; font-family: Consolas,
            'Bitstream Vera Sans Mono', 'Courier New', Courier,
            monospace !important; font-weight: normal !important;
            font-style: normal !important; font-size: 1em !important;
            direction: ltr !important; background-position: initial
            initial !important; background-repeat: initial initial
            !important;">
            <div class="container" style="margin-top: 0px !important;
              margin-right: 0px !important; margin-bottom: 10px;
              margin-left: 0px !important; padding: 0px !important;
              border-top-left-radius: 0px !important;
              border-top-right-radius: 0px !important;
              border-bottom-right-radius: 0px !important;
              border-bottom-left-radius: 0px !important;
              background-image: none !important; border: 0px !important;
              bottom: auto !important; float: none !important; height:
              auto !important; left: auto !important; line-height: 1.1em
              !important; outline: 0px !important; overflow: visible
              !important; position: relative !important; right: auto
              !important; text-align: left !important; top: auto
              !important; vertical-align: baseline !important; width:
              auto !important; box-sizing: content-box !important;
              font-family: Consolas, 'Bitstream Vera Sans Mono',
              'Courier New', Courier, monospace !important; font-weight:
              normal !important; font-style: normal !important;
              font-size: 1em !important; direction: ltr !important;
              background-position: initial initial !important;
              background-repeat: initial initial !important;">
              <div class="line number1 index0 alt2" style="margin-top:
                0px !important; margin-right: 0px !important;
                margin-bottom: 10px; margin-left: 0px !important;
                padding: 0px 1em 0px 0em !important;
                border-top-left-radius: 0px !important;
                border-top-right-radius: 0px !important;
                border-bottom-right-radius: 0px !important;
                border-bottom-left-radius: 0px !important;
                background-image: none !important; background-color:
                white !important; border: 0px !important; bottom: auto
                !important; float: none !important; height: auto
                !important; left: auto !important; line-height: 1.1em
                !important; outline: 0px !important; overflow: visible
                !important; position: static !important; right: auto
                !important; text-align: left !important; top: auto
                !important; vertical-align: baseline !important; width:
                auto !important; box-sizing: content-box !important;
                font-family: Consolas, 'Bitstream Vera Sans Mono',
                'Courier New', Courier, monospace !important;
                font-weight: normal !important; font-style: normal
                !important; font-size: 1em !important; direction: ltr
                !important; white-space: pre !important;
                background-position: initial initial !important;
                background-repeat: initial initial !important;"><code
                  class="plain plain" style="margin: 0px !important;
                  padding: 0px !important; font-style: normal
                  !important; font-weight: normal !important;
                  background-image: none !important; color: black
                  !important; font-family: Consolas, 'Bitstream Vera
                  Sans Mono', 'Courier New', Courier, monospace
                  !important; text-shadow: none !important;
                  border-top-left-radius: 0px !important;
                  border-top-right-radius: 0px !important;
                  border-bottom-right-radius: 0px !important;
                  border-bottom-left-radius: 0px !important; border: 0px
                  !important; bottom: auto !important; float: none
                  !important; height: auto !important; left: auto
                  !important; line-height: 1.1em !important; outline:
                  0px !important; overflow: visible !important;
                  position: static !important; right: auto !important;
                  text-align: left !important; top: auto !important;
                  vertical-align: baseline !important; width: auto
                  !important; box-sizing: content-box !important;
                  font-size: 1em !important; direction: ltr !important;
                  display: inline !important; background-position:
                  initial initial !important; background-repeat: initial
                  initial !important;"><br>
                  $ export CRAY_CUDA_PROXY=1<br>
                  <br>
                </code></div>
            </div>
          </td>
        </tr>
      </tbody>
    </table>
    See here:<br>
    <meta http-equiv="content-type" content="text/html;
      charset=ISO-8859-1">
    <a
href="https://www.olcf.ornl.gov/tutorials/cuda-proxy-managing-gpu-context/">https://www.olcf.ornl.gov/tutorials/cuda-proxy-managing-gpu-context/</a><br>
    <br>
    Cheers,<br>
    Dominic<br>
    <blockquote
cite="mid:CAMJ8fwqkNMCFAPu0PCEoOAuoCx4C1xpjrU2R0_9bm+Gm6TssBw@mail.gmail.com"
      type="cite">
      <div dir="ltr">
        <div>
          <br>
        </div>
        <div>I don't know of any good parallel sparse direct solver for
          small systems.
          <div><br>
          </div>
          <div>-Paul</div>
        </div>
      </div>
      <div class="gmail_extra"><br>
        <br>
        <div class="gmail_quote">On Thu, Dec 12, 2013 at 4:29 PM,
          Dominic Meiser <span dir="ltr"><<a moz-do-not-send="true"
              href="mailto:dmeiser@txcorp.com" target="_blank">dmeiser@txcorp.com</a>></span>
          wrote:<br>
          <blockquote class="gmail_quote" style="margin:0 0 0
            .8ex;border-left:1px #ccc solid;padding-left:1ex">
            <div bgcolor="#FFFFFF" text="#000000">
              <div>Hi Karli,
                <div class="im"><br>
                  <br>
                  On 12/12/2013 02:50 PM, Karl Rupp wrote:<br>
                </div>
              </div>
              <div class="im">
                <blockquote type="cite"><br>
                  Hmm, this does not sound like something I would
                  consider a good fit for GPUs. With 16 MPI processes
                  you have additional congestion of the one or two GPUs
                  per node, so you would have the rethink the solution
                  procedure as a whole.<br>
                  <br>
                </blockquote>
              </div>
              Are you sure about that for Titan? Supposedly the K20X's
              can deal with multiple MPI processes hitting a single GPU
              pretty well using Hyper-Q. Paul has seen pretty good speed
              up with small GPU kernels simply by over-subscribing each
              GPU with 4 MPI processes.<br>
              <br>
              See here:<br>
              <a moz-do-not-send="true"
href="http://blogs.nvidia.com/blog/2012/08/23/unleash-legacy-mpi-codes-with-keplers-hyper-q/"
                target="_blank">http://blogs.nvidia.com/blog/2012/08/23/unleash-legacy-mpi-codes-with-keplers-hyper-q/</a><br>
              <br>
              <br>
              Cheers,<br>
              Dominic<span class="HOEnZb"><font color="#888888"><br>
                  <br>
                  <br>
                  <pre cols="72">-- 
Dominic Meiser
Tech-X Corporation
5621 Arapahoe Avenue
Boulder, CO 80303
USA
Telephone: <a moz-do-not-send="true" href="tel:303-996-2036" value="+13039962036" target="_blank">303-996-2036</a>
Fax: <a moz-do-not-send="true" href="tel:303-448-7756" value="+13034487756" target="_blank">303-448-7756</a>
<a moz-do-not-send="true" href="http://www.txcorp.com" target="_blank">www.txcorp.com</a></pre>
                </font></span></div>
          </blockquote>
        </div>
        <br>
      </div>
    </blockquote>
    <br>
    <br>
    <pre class="moz-signature" cols="72">-- 
Dominic Meiser
Tech-X Corporation
5621 Arapahoe Avenue
Boulder, CO 80303
USA
Telephone: 303-996-2036
Fax: 303-448-7756
<a class="moz-txt-link-abbreviated" href="http://www.txcorp.com">www.txcorp.com</a></pre>
  </body>
</html>