By off-loading all the packet handling to the DSNIC board, as
discussed in Section 8.2,
all the packet handling interrupts, which are O(nr of
packets), can be avoided. Avoiding these interrupts has no effect
on the communication latency, since this latency is caused by the
O(nr of messages) operations. Since latency and
throughput are directly coupled in Comms1, there is no effect on the
throughput graph. The gain is to be expected in CPU load. For ATLAS,
which requires packet size 1 Kbyte, this can result in a
performance gain of up to 53 %, see
Table 3. This potential gain is modelled using
, see graph B in Figure 12.