performance.html

<!--#include virtual="header.inc" -->

<div class="navbar navbar-fixed-top">
  <div class="navbar-inner">
    <div class="container">
      <a class="btn btn-navbar" data-toggle="collapse" data-target=".nav-collapse">
        <span class="icon-bar"></span>
        <span class="icon-bar"></span>  
        <span class="icon-bar"></span>
      </a>
      <a class="brand" href="index.html">Grappa</a>
      <div class="nav-collapse">
        <ul class="nav">
          <li><a href="index.html">Home</a></li>
          <li><a href="about.html">About</a></li>
          <li><a href="contact.html">Contact</a></li>
          <!-- <li><a href="#download">Download</a></li> -->
        </ul>
      </div><!--/.nav-collapse -->
    </div>
  </div>
</div>


<div class="container">

  <div class="row">
    <div class="span6">

      <h1>Early results show promise</h1>
      
      <p>Grappa is still under development, but early results suggest
      we can perform well.</p>

      <img src="images/grappa-performance.png" />

      <p>We ran a modified version of the Unbalanced Tree Search (UTS)
	benchmark [Olivier et al. '06] to evaluate random access
	bandwidth and load balancing. We used the T1L tree, with a
	geometric distribution and 100 million nodes </p>

      <p>Grappa & the Cray XMT are both traversing a tree dispersed
	randomly through global shared memory using a parallel loop
	decomposition, with no random number generator calls.</p>

      <p>The MPI execution uses out-of-box UTS and traverses the tree
	implicitly, so there are random number generator calls and
	work-stealing, but no remote references.</p>
      
      <p>The Grappa and MPI runs used our 12-node 144-core 2.66 GHz
	Intel Westmere Xeon system with Mellanox QDR Infiniband. The
	XMT run used PNNL's 128-processsor Cray XMT-I.</p>

    </div>
  </div>
  
  <hr>
  
  <footer>
    <p>&copy; University of Washington CSE 2012</p>
  </footer>
  
</div> <!-- /container -->


<!--#include virtual="footer.inc" -->