Does it really need to be parallel, though?

Published on 7-Mar-2024


Please. I ask you. No, beg you. Before start spending weeks... months... years... designing a new computational theory to parallelize something. Does it really need to be parallel?

Nowadays, our computers are insanely fast. I am sure that you could pick up a laptop with a chip from 5 years ago and it would do all the things average people do perfectly. In fact, I did and you can. If you stick to one web-browser based application (this is admittedly quite difficult nowadays), you can run all non-professional workflows very well. But these chips even have multiple cores that often sit unused, why don't we start using them?

I am all in favor of using multiple cores... for applications where that makes sense. In the past few weeks I have been in a reading frenzy again. This time about something that are basically very specific hardware tests. The academic field for this is not very developed and most papers were from the 80s and 90s. Those papers are really nice to read... some people really knew how to write good papers back then. I wanted to know what the current state-of-the-art is with these testing tools and I looked at some newer papers. Almost all of them focused on making the tools parallel.

For some time, I was thinking about this problem as well. How do test and simulate digital circuits in parallel? You could do a discrete time simulation which does not scale that well for huge systems. Or what all these papers are doing is event-driven simulation. Parallel discrete time simulation is reasonably simple as it is mostly just matrix multiplications. But parallel event-driven simulation, well that is difficult. Event-driven simulation requires a massive amount of dependencies, bookkeeping and in general coordination. This is even more extreme within these tools than in other event-driven simulations. This task is not-embarrassingly parallel. So, what wins could we expect? At maximum, according to some people, 5x to 10x. Wow!

I was convinced. This is the way forward. We need to make these tools parallel — now. Imagine having to spend 10x less time on something, that is a massive win. Well, I am sad to announce that you and I won't be 10x faster. This is because of a simple fact. We were already using all our cores. Not several cores per test, but one core per test. However, we have 1000 tests and we have to run Monte-Carlo on them all. On top of our tool, we essentially had an embarrassingly parallel system all along.

Fighting Amdahl's Law to parallelize a system that is not the lowest-common-denominator was a mistake. I would call this problem and most of the testing field an Embarrassingly Not-Parallel system. Let's learn and move on from that.