overview.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="span5">

      <h2>What is Grappa?</h2>

      <p>Mass-market computer systems are designed to exploit spatial
	locality via cache and local memory to achieve high
	efficiency. Unfortunately, when processing graphs spatial locality is
	often difficult, if not impossible, to express.</p>

      <p>As system size grows,
	edges in a graph distributed across its nodes' memories become
	increasingly likely to join vertices that are far apart. The rate of
	traversal slows. Consequently, even though parallelism and hardware
	resources increase, performance degrades.</p>

      <p>Grappa is a latency-tolerant runtime for mass-market clusters
	that mitigates this degradation, allowing graph processing to scale up
	even in the presence of diminishing locality and increasing
	latency. Grappa works by: 
	<ul>
	  <li>exploiting fine-grained task parallelism to
	    tolerate the increasing latency, and</li>
	  <li>aggregating remote references
	    from disparate tasks to make better use of diminishing bandwidth at
	    scale.</li>
	</ul>
	The application developer need only express parallelism, not
	decide when and how to exploit it.

      <h2>Key idea: Trade latency for throughput</h2>
      <p>Grappa’s core component is a lightweight cooperative threading system. We tolerate latency by context-switching to other work.
	Given this ability, we can increase throughput by increasing the latency of key operations:
	<ul>
	  <li>Context-switch on long-latency memory operations</li>
	  <li>Delay and aggregate messages for better utilization of network bandwidth</li>
	  <li>Migrate synchronization operations to avoid unnecessary serialization</li>
	</ul>
      </p>
      
      <p><a class="btn" href="approach.html">See how Grappa works &raquo;</a></p>
    </div>
  </div>
  
  <hr>
  
  <footer>
    <p>&copy; University of Washington CSE 2012</p>
  </footer>
  
</div> <!-- /container -->


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