Parallel Computer Architecture

7.6 Shared Physical Address Space 511

The Alpha 21064 provides only nonblocking stores. Execution is allowed to pro-

ceed after a store instruction without waiting for the store to complete. Writes are

buffered by a write buffer, which is four deep, and in each entry up to 32 bytes of

write data can be merged. Several store instructions can be outstanding. A "memory

barrier" instruction is provided to ensure that writes have completed before further

execution commences. The Alpha nonblocking store allows remote stores to be

overlapped so that high bandwidth can be achieved in writes to remote memory

locations. Remote writes of up to a full cach ...