Performance Impact of Copies

We evaluate the impact of copies, necessary as part of data transfer when the entire shared address space cannot be registered with VIA, by presenting a comparison between the performance of the no-copy and copy versions in Figure 2. We present a comparison of the execution times breakdown for both versions, normalized with respect to the executions with copies. We had to run the applications with problem sizes smaller than the ones mentioned in Table 1 so that we could use both versions with the same problem size. The bars on the left, labeled ``NO COPY'', present the performance results for the no-copy version, and the bars on the right, labeled ``COPY'', present the performance results for the version with copies. Each bar presents a percentage breakdown of the different components which make up the execution time on a single node. Computation time is the time spent doing application computation. Page fetch time is the time spent in fetching a page from the home node, on a page miss. Lock time is the time spent in getting the lock from the current owner. Barrier time is the time spent waiting for barrier messages from other nodes, at the barrier. Overhead time is the time spent performing protocol actions. Handler time is the time spent inside the handler, servicing remote requests. Since we used only one processor on each node, for our experiments, the handler competes for the CPU with the application thread to service the messages received via the receive completion queue.

The page fetch time is what increases as a result of the additional copies at the home node and the receiving node during page transfers. We can see that Page Time makes up for a significant percentage of the execution time for Barnes, FFT and Radix, and these three applications show an improvement in performance with copy avoidance. Although avoiding copy is good, data transfer with copies doesn't degrade performance drastically. The performance degradation was maximum for FFT (15%) and very little (less than 5%) for the other applications.