Figure 4 shows the overall communication stack. As mentioned in Section 2, SPI and NDAI provide access to BG/Q’s Message Unit and P7-IH’s Hub Chip, respectively. We first introduce our lightweight library, libcomm, that abstracts the hardware and pro-vides reliable and optimized communication, including point-to-point and non-blocking collectives. On top of this communication veneer, we have implemented a minimalistic graph programming abstrac-tion that relies on two basic mechanisms, active messaging and message aggregation.