Based on the Condor package developed by the University of Wisconsin, Madison, we have implemented the proposed task management mechanism as a software layer that lies between the OS and the application processes:

• We implement the first version entirely outside the OS kernel and at the user level. This eliminates the need to access/change the internals of OS, allows us to concentrate on varying the degree of design complexity, and facilitates later expansion or generalization.
  
  
• We configure the proposed task management mechanism into three daemons, Collector, Schedd, and Startd, all of which reside constantly on each participating node. Two additional processes, Shadow and Starter, run on the home node and on the server node, respectively, whenever a task is remotely executed.

  
  
  

• We configure the module allocation scheme and the transfer, information, and location policies of the load sharing scheme into Schedd, Startd, and Collector, and leave the OS-related process transfer, execution, suspension, resumption, and checkpoint operations of the load sharing scheme in Shadow and Starter.
  
  
• Shadow on the home node and Starter on the server node cooperate to achieve locality transparency. By transparency, we mean a process's behavior should not be affected by its transfer. To this end, we configure Shadow to act as an agent for the remotely executed task in performing system calls.

To facilitate monitoring of the task management system, we have also implemented a Java monitor. The monitor runs on a machine (which needs not be one of the participating workstations), and communicates with a statistics collector daemon that runs at each participating machine. The statistics collector daemon is implemented in such a way that it behaves like a "remote shell" daemon. That is, it takes the command from a socket, executes it locally and then returns the result to the command initiator (i.e., the Java monitor). In this manner, the Java monitor has full control of the commands to invoke, the arguments that goes with the commands, and when to invoke the commands. More details on the Java monitor can be found here.

---

Return to Project Home Page
Date last modified -- August 15, 1998
Direct comments concerning this WWW site to: jhou@ece.osu.edu.