> == Detailed description == > Currently, the i686 RPM packages are built in such a way that they are > compatible with very old i686 systems, such as the Pentium III. The > only addition over the i686/Pentium Pro baseline is a requirement to > support long NOPs, for Intel CET. However, the majority of > installations of i686 packages is for use on x86_64 systems, as > multi-lib RPMs. Based on what data?

> This proposal suggests to accept this reality and build the i686 > packages in such a way that they require the ISA level of (early) > x86-64 CPUs. On which x86 CPU families will Fedora continue to work?

On Mon, 2018-06-04 at 16:04 +0200, Florian Weimer wrote: >>> This proposal suggests to accept this reality and build the i686 >>> packages in such a way that they require the ISA level of (early) >>> x86-64 CPUs. >> >> On which x86 CPU families will Fedora continue to work? > > Based on this proposed change, you will need an x86-64-capable CPU. That's not how I read the proposal. No reason a P4 shouldn't work if all you're requiring is SSE2+FXSR, right?

Certainly, I'm no netburst fan either. But the last[*] 32-bit-only Intel chip was Yonah (Core 1), which went out of production in 2008- ish, so there's about seven years worth of CPUs between the introductions of SSE2 and AMD64.

On Mon, 2018-06-04 at 17:21 +0200, Florian Weimer wrote: > On 06/04/2018 05:07 PM, Adam Jackson wrote: > > On Mon, 2018-06-04 at 16:04 +0200, Florian Weimer wrote: > > > > > > > This proposal suggests to accept this reality and build the i686 > > > > > packages in such a way that they require the ISA level of (early) > > > > > x86-64 CPUs. > > > > > > > > On which x86 CPU families will Fedora continue to work? > > > > > > Based on this proposed change, you will need an x86-64-capable CPU. > > > > That's not how I read the proposal. No reason a P4 shouldn't work if > > all you're requiring is SSE2+FXSR, right? > > It will probably work, but it's more of a happy accident than a > deliberate decision. I would still suggest it if it were incompatible > with Pentium 4 CPUs without x86-64 support. (The fewer of these > Netburst power guzzlers are running, the better, to be honest.) Certainly, I'm no netburst fan either. But the last[*] 32-bit-only Intel chip was Yonah (Core 1), which went out of production in 2008- ish, so there's about seven years worth of CPUs between the introductions of SSE2 and AMD64. So this isn't eliminating the possibility of extant i686 kernels, which "will need an x86-64-capable CPU" might otherwise imply.

> >>support long NOPs, for Intel CET. However, the majority of > >>installations of i686 packages is for use on x86_64 systems, as > >>multi-lib RPMs. > >Based on what data? > The mirror data I've seen, but it's really outdated at this point. There are currently about 275,000 IP address checking in with x86_64 systems every day. and 25,000 x86_32. So it's about 11:1.

On 4 June 2018 at 14:18, Matthew Miller <mattdm(a)fedoraproject.org&gt; wrote: > On Mon, Jun 04, 2018 at 02:01:13PM -0400, Matthew Miller wrote: >> > >>support long NOPs, for Intel CET. However, the majority of >> > >>installations of i686 packages is for use on x86_64 systems, as >> > >>multi-lib RPMs. >> > >Based on what data? >> > The mirror data I've seen, but it's really outdated at this point. >> There are currently about 275,000 IP address checking in with x86_64 >> systems every day. and 25,000 x86_32. So it's about 11:1. > > Here's a breakdown as arch percentage over time: > > https://twitter.com/mattdm/status/1003701941724172295 > > Looks like the 50:50 point was about five years ago. > For F26,F27,F28 for the first 150 days of the year: Days: 150 X86_32: 3769.4 avg/day X86_64: 159540 avg/day Ratio (32/64) 0.0236267 (1:42) The reason is that the majority of i386 users are not moving off of dead/old releases. For the month of May Avg/day release 28623.7 epel6 16009.2 epel5 4043.37 f25 3318.53 f20 2759.2 f08 2127.8 f23 1589.63 f27 1401.7 f22 1371 f26 1277.53 f21 1056.33 f28 874.2 f24 733.8 f14 562.533 f19 509.067 f11 In comparison, x86_64 is mostly living on the latest release: [smooge@data-analysis01 mirrors]$ grep '^2018-05-.* x86_64' out-2018 | awk '{print $3}' | sort | uniq -c | sort -bnr | head -n 15 | awk '{print $1/30, $2}' 671455 epel7 605436 epel6 87367.4 f27 60159.7 f28 37857.6 epel5 34421.4 f26 27011 f25 19115.4 f23 14292.2 f24 8627.47 f22 6603.3 f20 5863.47 f21 1413.37 modular_f28 1091.87 f19 868.833 f18 The 30 is because only 30 days of May have been analyzed. [And yes there are still many Fedora 08 systems showing up for x86_32] Personally I think the number of those systems which are running Pentium III hardware with the latest OS are probably already on this mailing list.. I expect that they would also be looking at only needing to support a small subset of the software since running GNOME/KDE with usually 128->512 MB of RAM is not possible. It also would probably want a very stripped down installer since anaconda is aimed at 'current' hardware... probably something like the images that ARM-32 makes.

On Mon, Jun 04, 2018 at 02:01:13PM -0400, Matthew Miller wrote: > > >>support long NOPs, for Intel CET. However, the majority of > > >>installations of i686 packages is for use on x86_64 systems, as > > >>multi-lib RPMs. > > >Based on what data? > > The mirror data I've seen, but it's really outdated at this point. > There are currently about 275,000 IP address checking in with x86_64 > systems every day. and 25,000 x86_32. So it's about 11:1. Here's a breakdown as arch percentage over time: https://twitter.com/mattdm/status/1003701941724172295 Looks like the 50:50 point was about five years ago.

Thanks for the data. 25k is still a pretty healthy number. :) I realize that there are a lot of unknowns in the data, so it is difficult to draw any hard conclusions, but 25k is still much larger than 0. Splitting into i686 into i586 and i686 would give more insight into who still needs non-SSE2... Probably hurts my argument, though. :)

Thanks for the data. 25k is still a pretty healthy number. :) I realize

that there are a lot of unknowns in the data, so it is difficult to draw any hard conclusions, but 25k is still much larger than 0. Splitting into i686 into i586 and i686 would give more insight into who still needs non-SSE2... Probably hurts my argument, though. :)

On Mon, Jun 04, 2018 at 03:50:34PM -0400, Jeff Backus wrote: > Thanks for the data. 25k is still a pretty healthy number. :) I realize Yeah, absolutely. And it's likely that those mirror numbers undercount, because not every system checks in daily, and then there's also NAT. But, my gut feeling is that about half of those are not using a current release _anyway_. Honest question: do you think that 12k would still count as a healthy number? I mean, it's not peanuts. But maybe it'd be better served by a Fedora remix (or similar) specifically targetting older and low-powered systems?

> that there are a lot of unknowns in the data, so it is difficult to draw > any hard conclusions, but 25k is still much larger than 0. Splitting into > i686 into i586 and i686 would give more insight into who still needs > non-SSE2... Probably hurts my argument, though. :) Soooooo.... this is the kind of thing that more a detailed hardware census could really help us with!

On 5 June 2018 at 12:49, Jeff Backus <jeff.backus(a)gmail.com&gt; wrote: > > > On Mon, Jun 4, 2018 at 10:07 PM, Matthew Miller <mattdm(a)fedoraproject.org&gt; > wrote: >> >> On Mon, Jun 04, 2018 at 03:50:34PM -0400, Jeff Backus wrote: >> > Thanks for the data. 25k is still a pretty healthy number. :) I realize >> >> Yeah, absolutely. And it's likely that those mirror numbers undercount, >> because not every system checks in daily, and then there's also NAT. >> >> But, my gut feeling is that about half of those are not using a current >> release _anyway_. Honest question: do you think that 12k would still >> count as a healthy number? I mean, it's not peanuts. But maybe it'd be >> better served by a Fedora remix (or similar) specifically targetting >> older and low-powered systems? > > > Good question. I think it would be more productive to think in percentages > instead of raw numbers, in this case. There are a lot of FOSS projects out > there that would love to have 12k users. :) > > Certainly, I would consider 10% a healthy number when talking about portion > of user base. I would even argue that 1% is still a healthy number, > particularly with regard to decisions that have a reasonable chance of > disenfranchising those affected. While I hate seeing people leave a > community, I wouldn't be able to defend 0.1%. So, somewhere in there is my > general boundary. > > Now cost changes all of that, of course. Obviously if 75% of our effort is > going to please 10%, then 10% isn't a healthy number. > > Clearly effort is going into enabling Fedora to work on non-SSE2 systems by > teams invested in the success of Fedora in general and not the success of > non-SSE2 systems in particular. I just don't know how to quantify it. > > Based on Smooge's awesome numbers, it looks like x86_32 is in the 2.3% > range. It would be interesting to see how this stacks up to AArch64 and > other secondary arches. Unfortunately, what complicates things is how x86_32 > is so intertwined with x86_64. > It is also complicated in that most of the large sites using aarch64 and arm32 do so in ways which make them uncountable. They will have 'thousands' of nodes but all of them use an internal mirror so we see them as only 1.. The other issue is that the arm/aarch64 have active upstream help from people who are building the boards. There isn't any such support on the x86_32 side with the manufacturers getting to the point of saying "here is $20.00 and an ebay link.. buy at least a pentium iv or v please. " The question that the x86 group needs to figure out is how many of the 3800 active systems are going to not have SSE2. > To your point re: a remix, that is an option we've discussed within the SIG > and is one we are open to exploring. A remix wouldn't resolve issues > introduced by enabling SSE2 by default, unless we maintained a parallel set > of packages e.g. i586 (which I've already been warned about. :) ) > >> >> > that there are a lot of unknowns in the data, so it is difficult to draw >> > any hard conclusions, but 25k is still much larger than 0. Splitting >> > into >> > i686 into i586 and i686 would give more insight into who still needs >> > non-SSE2... Probably hurts my argument, though. :) >> >> Soooooo.... this is the kind of thing that more a detailed hardware >> census could really help us with! > > > Yes, I would agree :) > So it would probably be a lot more detailed than other sites are running. I looked at the popcorn data and they seem to count whether an OS is i386 or amd64 not if the CPU is pentium iii. Neither did any of the other OS census programs.