3 ms·

I wish he would have gone more into the design of the kernel and significant changes that are going to be needed to take advantage of the NVRAM based systems th

I wish he would have gone more into the design of the kernel and significant changes that are going to be needed to take advantage of the NVRAM based systems that are coming very soon. The I/O design needs to be completely overhauled. It's probably going to be a great time to rewrite and resign portions of Linux, but I honestly don't think the community will be capable of doing a major architectural change. Kernel modules have been a great step towards modular design, and this needs to be pushed everywhere so that more changes can be isolated in the development process. I've been a huge Linux (and GNU, most of what this article is about really is GNU, not Linux) user since being introduced to it in 1996, but as much as I love it, I do wonder if there are new options that will reveal themselves in the next few years that will better answer some of the modern hardware advancements. Linux is a beast of a system now, with a lot of technical debt and a hard to penetrate C code base, I hope it can evolve where it needs to, but I think it will require huge commitments from the community.

sievebrain 10y ago
Could you elaborate on the changes for the NVRAM devices that you'd like to see? My understanding is that the 3D XPoint type devices are just exposed as a regular mmapped device and provide storage which is still a fair bit slower than DRAM, but a lot faster than SSDs.
- bluejekyll 10y ago
  It's mainly around IO scheduling in the kernel against other processes. This is a good intro paper on some issues: https://homes.cs.washington.edu/~luisceze/publications/novos-hotos2011.pdf https://homes.cs.washington.edu/~luisceze/publications/novos... This discussion is relevant too: https://lwn.net/Articles/547903/ https://lwn.net/Articles/547903/ Those are both a few years, old, but I haven't seen discussion as to the progress made around scheduling when the disks start performing as fast as the CPU load operations. Again, I was just pointing out that I wish that article got into more details around that, b/c I don't know the state of the world, but have some concerns in this particular area and the ability for the kernel to be able to accommodate new scheduling techniques to better utilize the much faster disk access.

sievebrain 10y ago

Could you elaborate on the changes for the NVRAM devices that you'd like to see? My understanding is that the 3D XPoint type devices are just exposed as a regular mmapped device and provide storage which is still a fair bit slower than DRAM, but a lot faster than SSDs.

bluejekyll 10y ago
It's mainly around IO scheduling in the kernel against other processes. This is a good intro paper on some issues: https://homes.cs.washington.edu/~luisceze/publications/novos-hotos2011.pdf https://homes.cs.washington.edu/~luisceze/publications/novos... This discussion is relevant too: https://lwn.net/Articles/547903/ https://lwn.net/Articles/547903/ Those are both a few years, old, but I haven't seen discussion as to the progress made around scheduling when the disks start performing as fast as the CPU load operations. Again, I was just pointing out that I wish that article got into more details around that, b/c I don't know the state of the world, but have some concerns in this particular area and the ability for the kernel to be able to accommodate new scheduling techniques to better utilize the much faster disk access.