Approach 1

The first approach recognizes that throttling a node makes its synchronization time negligible. When a node is being throttled, it is made to be slower than all other nodes so as to expose data dependencies. Consequently, the think time between I/Os is all computation (e.g., node 1 in Figure 2 does not have to wait on nodes 0 and 2, because node 1 is the slowest node). The primary advantage of this approach is that it can be used even if an application is using ``untraceable'' synchronization mechanisms such as shared memory. The disadvantage is that I/O sampling can affect the computation calculation. This is discussed more in Section 5.