Hmm, yes I see. TBs? Interesting. I'd like to hear a talk about these things. Naively I would expect that (as is the case for my MUCH smaller scale system) tha

Hmm, yes I see. TBs? Interesting. I'd like to hear a talk about these things. Naively I would expect that (as is the case for my MUCH smaller scale system) that I can compensate for network/disk-bound and non-multithreaded stages by merely running two concurrent frames. On a larger scale I would expect to be able to estimate RAM-cheap frames, and always have one of them running per machine, but at SCHED_IDLE priority, so that they only get CPU when the "main" frame is blocked on disk or network, or a non-parallelizable stage. By starving one frame of CPU, it's much more likely that it'll need CPU the short intervals when it's allowed to get it.