Mainframes staying busy at 98% is because they have batch workloads with no guarantees of completion time. It’s effectively using the “idle loop” of the OS to handle programs that might complete today / tomorrow / whenever. It also does a great job of limiting the physical pages of memory that these background jobs can use.

About 25 years ago I was working for a very large travel provider and they were so proud that their largest TPF system could run at 100% all day and still provide great response times. My team did some detailed analysis of the logs and we found that about 0.1% of transactions actually failed, but they were using 28% of the CPU. Actual useful workload was 72%

There’s nothing in a mainframe that can beat the physics of queuing theory, they just use these numbers for marketing.

Also, as I understand it, Parallel Sysplex is not shared memory, the coupling facility is used to coordinate locking and exchange of data pages.

A.

The Fujitsu A64FX and Fugaku supercomputer demonstrate that it’s possible to create a high-performance computing architecture that is better than a heterogeneous GPU-based system both in raw performance and the ease in which scientists can apply that performance to real problems.

On the other hand, the Power9 on its own had noticeably mediocre float-point performance even at release. Trying to remedy that with tightly coupled GPU accelerators created a hardware platform too expensive for entry-level usage with an inflexible proprietary software stack–good riddance of that idea.

The only way to sell big systems is if a successful customer can start small and grow. The only way to sell small systems is if they are easy to program and can be scaled to big systems. While the huge systems work well, both Power and Z are missing performant entry-level systems that entice customers who need to start small and grow. In my opinion, Power10 needs to be powerful enough on its own to create an attractive entry-level system.

The hype of composable compute infrastructure sounds great, especially in that interview with William Starke on this site last year. The reality, however, could be much different: an impossible to program, afford or control Frankenstein style of machine. Conversely, the way in which scalable vector extensions, tofu and high-bandwidth memory were added to the ARM architecture machine by Fujitsu resulted in architecture in which the features needed for high-performance computing blended together well.

Thanks for the thought provoking article!