We now consider how this enhancement affects the access time of non-block transfer requests (those remote requests that have nothing to do with block transfers). The access time of these requests will increase for two reasons. First, LONG_READ instructions have an unfair advantage over normal read or write instructions when it comes to accessing a busy memory. The requests caused by both types of instructions fight to access a remote memory identically; that is, a remote node interface will reject an incoming request whenever it is busy, regardless of the type of incoming request. However, when a block read request is accepted by a node interface, the entire block will be read before any other requests destined to the same node could possibly be accepted. In the base multiprocessor, however, each read that is part of a block transfer must compete individually. Thus, the memory contention seen by non-block transfer requests will be smaller in the base multiprocessor than in the enhanced system.
Second, in a slotted ring, if many consecutive slots are used to transfer a block, the clustering of the full slots will tend to increase the average ring access time. Network issues are discussed in detail in [5].