From mboxrd@z Thu Jan 1 00:00:00 1970 From: lm@mcvoy.com (Larry McVoy) Date: Mon, 20 Nov 2017 11:59:40 -0800 Subject: [TUHS] UNIX on S/370 In-Reply-To: <51d71c07-ca95-b6b4-2907-a21814c6da1e@kilonet.net> References: <20171120160504.3C46B18C091@mercury.lcs.mit.edu> <20171120191055.GF9146@mcvoy.com> <51d71c07-ca95-b6b4-2907-a21814c6da1e@kilonet.net> Message-ID: <20171120195940.GL9146@mcvoy.com> Who knows if it will go anywhere. I got dragged out of retirement with hints of piles of money, so far, they loaned me a box. I believe the likely target for this would be AMD's Epyc. They have already pushed one box to serve up about 100Gbit/sec of movies and that's with them doing TLS in the kernel; be faster if they could get the NIC to do that. https://news.ycombinator.com/item?id=15367421 The limiting factor, so far as I can tell, is memory BW and PCIe lanes. Epyc seems to deliver more of that. On Mon, Nov 20, 2017 at 02:56:04PM -0500, Arthur Krewat wrote: > I would love to see the results of that, including more information about > the architecture in question. > > > On 11/20/2017 2:10 PM, Larry McVoy wrote: > >On Mon, Nov 20, 2017 at 02:07:27PM -0500, Paul Winalski wrote: > >>It would mean that you wouldn't have to implement machine check > >>support and other hardware error handling. The VM hypervisor would do > >>that for you. It would also let you run multiple versions of UNIX > >>simultaneously. Very convenient if you're doing kernel or driver > >>development. > >Indeed. I'm currently trying to convince Netflix that the way to get the > >most performance out of a NUMA machine is to boot a different kernel on > >each NUMA domain. One way we might demo that is on a 4 domain system > >lock down 3 hypervisors and their guest OS to 3/4 of the NUMA domains > >and give the host kernel the 4th. > > -- --- Larry McVoy lm at mcvoy.com http://www.mcvoy.com/lm