3 ms·
18 exabytes isn't that much if you want to uniquely address every memory byte in a (future) datacenter. Papers on RDMA-based unified memory systems are current
by pclmulqdq 5y ago
18 exabytes isn't that much if you want to uniquely address every memory byte in a (future) datacenter. Papers on RDMA-based unified memory systems are currently producing some very compelling performance results in workloads like pub-sub and mostly-read databases. It's conceivable that in the next 15 years, there will be special-purpose datacenter-scale computers networked with a unified memory space that need address spaces greater than 64 bits, especially with the high-capacity optane DIMMs that Intel produces.
- whatshisface 5y agoYou will need something much more exotic than a data center for the individual processors to need to address more than 18 exabytes. An architecture where every node can directly read the RAM of every other node is called a supercomputer. In a data center, dividing global address into node IDs and true addresses would handle everything you'd want to do. If the top N bits of the global addresses are only ever read by routing software they don't need to be represented in the processor hardware.
- remram 5y agoDoes any system marketed as a "supercomputer" with exabytes of memory actually allow unified addressing of the RAM? My understanding is that they are used through interfaces such as MPI which doesn't work that way.
- whatshisface 5y agoI think you are on to something insofar as noticing that supercomputing vs. cluster computing is a spectrum. You could imagine starting with cluster computing at zero hardware support, and slowly adding more hardware interconnects and putting more work into topology, as you climb towards the platonic ideal supercomputer which, fitting its name, would simply be one super computer.
- remram 5y agoI don't know if manufacturers are going up the ladder though. Even desktop computers are looking more and more like independent interconnected cores (eg NUMA and GPUs). Even in software, compute-heavy workloads tend to be written that way regardless, to sidestep locking issues. I just don't think people would want to address random parts of memory even if a hypothetical exabyte memory hardware allowed it.
- whatshisface 5y agoGPU cores address a unified memory with many ALUs. That's supercomputer-y architecture.
- remram 5y agoWhat I meant is that CPU and GPU are separate, and the GPU cannot access all the memory.
- fanf2 5y agoSome years ago, the COSMOS supercomputer in Cambridge (used by Stephen Hawking’s group) was the largest shared-memory computer in Europe, with 14.8 terabytes of RAM, rather below the exabyte scale and well within the 48 bit virtual address space http://www.ctc.cam.ac.uk/news/140623_newsitem.php http://www.ctc.cam.ac.uk/news/140623_newsitem.php
- remram 5y agoThanks! That is bigger than I thought!
- jbverschoor 5y agoIf we let some people have their way, we'll have an Electron app for each widget on your screen.
- na85 5y agoI'm not sure even 18 exabytes would be enough memory for that.
- whatshisface 5y agoI'm tired of people complaining about this. If you don't see the need to process-isolate each DOM element, you don't understand frontend. What if a <div> crashes, or stalls? It used to be that when one DOM element's event listeners failed the whole application would fail. Now we simply restart the process and repaint the screen. The messaging and coordination is a nightmare but the reliability and scalability is worth it. With modern support for post-async language constructs it does not even increase development time by more than a large constant factor.