[Solved]: What guarantees do “soft” real-time operating systems actually provide

What does soft realtime mean? Cynics will say “basically nothing”. (…) Many telecomms systems have less strict requirements [than hard realtime], for instance they might require a statistical guarantee along the lines of “a database lookup takes less than 20ms in 97% of cases”. Soft realtime systems, such as Erlang, let you make that sort of guarantee.

And this does provide a useful definition. Soft real-time indicates a design which is optimized towards each individual task taking no more than a certain amount of time, rather than towards optimizing the total time spent to perform all tasks. For example, a soft real-time system would aim to complete 99.9% of the requests in 10ms and process 100 requests per second, where a non-real-time might aim to proceed 200 requests per second but allow the occasional request to block for 50ms or more. A hard real-time would guarantee one request every 15ms no matter what. Soft real-time is used for applications where a missed deadline means a degradation of service, but is not performance-critical. Multimedia and telephony are some typical use cases. Each audio or video frame must be rendered it time, or else it has to be skipped; but skipping a frame is not the end of the world. The designers of Erlang had similar objectives on reliability in other matters: they observed that it was better for a telephone exchange to very occasionally drop a call, but to be absolutely sure that the exchange as a whole would keep working come what may, than to ever risk catastrophic failure in trying to maintain connections at all cost. In contrast, something like controlling a motor requires that the software never misses a deadline. This has costs: the overall performance is typically slower, and only relatively simple behaviors can be achieved. On the other side of the spectrum, a number crunching application typically cares only about overall performance — what matters is how fast the 1000×1000 matrices are multiplied, not how fast each column is calculated.