The T/16 was a remarkably successful machine for its intended purpose—at one time over 80% of all ATMs in the U.S. were controlled by Tandem systems—but of course there were disadvantages as well. Some, like the higher cost in comparison with conventional systems, are inevitable. Others were not so obvious to the designers.
Tandem was justifiably proud of the near-linear scaling of performance when hardware was added. Horst and Chou (1985), which refers to a later system, the TXP, shows how a FOX cluster can scale linearly from 2 to 32 processors.
Bartlett (1982) shows the downside: the performance of the message system limited the speed even of small systems. A single message with no attached data takes over 2 ms to transmit, and messages with 2,000 bytes of data in each direction take between 4.6 ms (same CPU) and 7.0 ms (different CPUs). This is the overhead for a single I/O operation, and even in its day it was slow. The delay between sequential I/O requests to a disk file was long enough that they would not occur until the data had passed the disk head, meaning that only one request could be satisfied per disk revolution. A program that sequentially reads 2 kB from disk and processes it (for example, the equivalent of grep) would get a throughput of only 120 kB/s. Smaller I/O sizes, such as 512 bytes, could limit the throughput to floppy disk speeds.
As the name “Tandem/16” suggests, the designers had a 16-bit mindset. That is fairly ...