The impact of operating system on the performance of user-level data transfer is quite limited. In VMMC, the OS involvement is needed only when the address translation is not available in the UTLB translation table. To evaluate this effect, we first measure the performance in the common case in which all the address translation is in the UTLB translation table and therefore has no operating system involvement. We then measure the UTLB overheads when the translations are not available in the translation table. Finally we measure the UTLB overheads in a few real applications and parallel programs.
Figure 4: Remote send latency for small messages
Figure 5: One-way send bandwidth