• Can determine the dataset sizes for batch-style parallelism — effectively a latency
• Can determine the processing speed for stream-based parallelism
• Low bandwidth is acceptable with highly-independent parallelism
• High bandwidth is beneficial for highly-dependent parallelism

A node can be connected to one or more SATAnet routers and to one or more ports per router. Using multiple ports allows for multiplication of the bandwidth and also lower average latency.

The following table lists bandwidths for packet transfers of varying sizes on a SATA-I interface.

• Can determine the granularity of the parallelism
• Low-latency is important in synchronous or poorly-parallel applications
• High-latency is tolerable in asynchronous and highly-parallel applications
• Set-up latencies can be offset by merging packets
• In core-rich nodes one core can be allocated to communications to minimise blocking

Packet-switched latencies

The following table lists timings for a packet to transfer through various parts of a Linux-based system.

Transmission on one node is asynchronous to reception on another node, and so the latencies overlap. In practice this means that the overall latency is lower than the simple summation of each part of the transfer. Since the latencies are roughly symmetrical for both transmission and reception, the typical overall latency is T_tad + T_tdh + T_net = 62µs.

Circuit-switched latencies

The following table lists the approximate timings for a 4-byte word to transfer through various parts of a Linux-based system.

Total end-to-end latency is simply T_tah + T_net + T_rha = 0.8µs. The bandwidth achieved for back-to-back transfers of 4-byte words is around 8~12Mbytes/sec.