Indiana University

Media Contact

PTI Strategic Initiatives

ptisi [at] indiana [dot] edu

ParalleX at Supercomputing 2011 in Seattle.

ParalleX will be on display at Supercomputing 2011 in Seattle.  The demo will highlight some of the capabilities ParalleX applications gain for free:  active power management and automatic load balancing.  The demo demonstrates an adaptive mesh refinement simulation evolving a hyperbolic system of equations:

November 10, 2011

ParalleX will be on display at Supercomputing 2011 in Seattle.  The demo will highlight some of the capabilities ParalleX applications gain for free:  active power management and automatic load balancing.  The demo demonstrates an adaptive mesh refinement simulation evolving a hyperbolic system of equations:



 Power usage is externally measured using the "Watt's Up" power meter.  Power usage is reported to the ipad interface:




 The ipad interface displays the power usage externally measured by the "Watt's Up" meter.  It also directly communicates with the HPX runtime system to control the thread managers of each core.

 A series of CPU kill-switches will toggle a particular thread manager associated with a specified core.  Turning off a core will reduce the speed of the simulation but also reduce power consumption.  The CPU clock speed can also be changed interactively through the ipad application.

 A maximum power budget can be directly specified using the "Auto Power Management" tool in the upper right hand corner of the ipad application:

 

 The "Auto Power Management" button opens a new dialogue where the user can specify the overall power budget for a given simulation.  This tool will turn on or off thread managers in order to meet the specified power consumption limitation.  When the auto-throttle client is active, the CPU kill-switch is greyed out; however, it will reflect which cores are active based on the power budget supplied.  The max power budget limit is indicated in the graph by the red line.

 ParalleX provides fine grained power management tools and performance counters that are independent of the application.  The will be especially useful in situations where power budgets are constrained and vary in time.