4 ms·

I’m surprised that no work has been put into Linux such that one instance of the OS can work across a whole cluster of machines. This seems like the main sellin

by mruts 7y ago

I’m surprised that no work has been put into Linux such that one instance of the OS can work across a whole cluster of machines. This seems like the main selling point of mainframes and you would think cloud providers would be interested in developing a replacement. Or maybe they have, but just on top of the OS? Maybe forcing everyone to use their property crap is a better idea? But from a technical perspective, how hard would it be to turn Linux into something that could seemlessly use resources from many different machines? We already have NUMA machines in which some cores are “different” from others and Linux seems okay at managing that. Would it to be such a large step for it to deal with networked CPUs? And the networked filesystem part seems to be already solved. IIRC Plan9 supported all this for both compute and disk a long-time ago.

zozbot234 7y ago
Plan9 is not a true single-system-image OS (w/ process migration and the whole approach you're describing here). There were many attempts to achieve this with Linux - OpenMOSIX, OpenSSI among others - but they've all been abandoned and bitrotted.
sjg007 7y ago
I think the big issue is latency.. it just takes too long.
detaro 7y ago
It has been tried many times, but it just didn't work that well: network was a lot slower than the interconnects in a NUMA machine are, and even those are noticeable in application performance. If you need to modify your applications to use it properly, you're quickly at the point where you instead are better off to make the application distributed. It seems to come up again a bit in research, given recent hardware developments, but I doubt it'll see widespread use.
njharman 7y ago
> Would it to be such a large step for it to deal with networked CPUs? Yes. Mainframe 90% the hardware and 10% the OS. To get mainframe capability you need mainframe hardware. The use case/need for mainframe capability is very niche.

by mruts 7y ago

zozbot234 7y ago

Plan9 is not a true single-system-image OS (w/ process migration and the whole approach you're describing here). There were many attempts to achieve this with Linux - OpenMOSIX, OpenSSI among others - but they've all been abandoned and bitrotted.

sjg007 7y ago

I think the big issue is latency.. it just takes too long.

detaro 7y ago

It has been tried many times, but it just didn't work that well: network was a lot slower than the interconnects in a NUMA machine are, and even those are noticeable in application performance. If you need to modify your applications to use it properly, you're quickly at the point where you instead are better off to make the application distributed. It seems to come up again a bit in research, given recent hardware developments, but I doubt it'll see widespread use.

njharman 7y ago

> Would it to be such a large step for it to deal with networked CPUs? Yes. Mainframe 90% the hardware and 10% the OS. To get mainframe capability you need mainframe hardware. The use case/need for mainframe capability is very niche.