Asynchronous data transfer to MIC buying anything? https://conferences.xsede.org/c/message_boards/find_message?p_l_id=&messageId=462934 2013-02-19T18:33:50Z 2013-02-19T18:31:31Z

I&#039;m experiencing what appears to be a bug with asynchronous memory transfer offloads to the MIC. I&#039;ve written a short program to expose the issue, which uses a 6GB array of doubles, and measures the time spent by the CPU executing a synchronous memory transfer, and then an asynchronous memory transfer. The array is pre-allocated on the MIC before the measurement portion of the code. <br /><div class="quote"><div class="quote-content"><br />time_sync = omp_get_wtime();<br />#pragma offload_transfer in(data : length(size) alloc_if(0) free_if(0))<br />time_sync = omp_get_wtime() - time_sync;<br /><br />time_async = omp_get_wtime();<br />#pragma offload_transfer in(data : length(size) alloc_if(0) free_if(0)) signal(data)<br />time_async = omp_get_wtime() - time_async;<br /><br />time_wait = omp_get_wtime();<br />#pragma offload_wait wait(data)<br />time_wait = omm_get_wtime() - time_wait;</div></div><br /><br />I am seeing that time_async is approximately equal to time_sync (about 0.9 seconds), and that time_wait is only a fraction of time_async (about 0.01 seconds). By contrast, for asynchronously offloaded computation, I see no pause on the CPU.<br /><br />Is this a bug? Shouldn&#039;t asynchronous data transfer be, well, asynchronous?

2013-02-19T18:31:31Z