This probably comes up once in a while, thought I'd raise it again.
I'd like to suggest switching the default kernel to -PAE on machines that support it, for the following reasons:
- many machines have 4GB+ these days, even desktops - NX is only available with -PAE, improves security - kvm is significantly faster on AMD when PAE is selected (since we don't support NPT on non-PAE)
Not subscribed - please copy.
On Mon, Jan 19, 2009 at 03:49:20PM +0200, Avi Kivity wrote:
This probably comes up once in a while, thought I'd raise it again.
I'd like to suggest switching the default kernel to -PAE on machines that support it, for the following reasons:
- many machines have 4GB+ these days, even desktops
- NX is only available with -PAE, improves security
- kvm is significantly faster on AMD when PAE is selected (since we
don't support NPT on non-PAE)
What's needed to set this by default is changes in anaconda. They have their own list at anaconda-list@redhat.com
Dave
Dave Jones (davej@redhat.com) said:
On Mon, Jan 19, 2009 at 03:49:20PM +0200, Avi Kivity wrote:
This probably comes up once in a while, thought I'd raise it again.
I'd like to suggest switching the default kernel to -PAE on machines that support it, for the following reasons:
- many machines have 4GB+ these days, even desktops
- NX is only available with -PAE, improves security
- kvm is significantly faster on AMD when PAE is selected (since we
don't support NPT on non-PAE)
What's needed to set this by default is changes in anaconda. They have their own list at anaconda-list@redhat.com
anaconda-devel-list, actually. But I could have sworn we already did this.
Bill
On Mon, Jan 19, 2009 at 11:23:26AM -0500, Bill Nottingham wrote:
Dave Jones (davej@redhat.com) said:
On Mon, Jan 19, 2009 at 03:49:20PM +0200, Avi Kivity wrote:
This probably comes up once in a while, thought I'd raise it again.
I'd like to suggest switching the default kernel to -PAE on machines that support it, for the following reasons:
- many machines have 4GB+ these days, even desktops
- NX is only available with -PAE, improves security
- kvm is significantly faster on AMD when PAE is selected (since we
don't support NPT on non-PAE)
What's needed to set this by default is changes in anaconda. They have their own list at anaconda-list@redhat.com
anaconda-devel-list, actually. But I could have sworn we already did this.
Unless we keep the non-PAE i586 kernel around as a fallback, we're not going to be able to boot on a whole raft of crappy i386 chips (original Pentium M most notably...)
regards, Kyle
Kyle McMartin wrote:
On Mon, Jan 19, 2009 at 11:23:26AM -0500, Bill Nottingham wrote:
Dave Jones (davej@redhat.com) said:
On Mon, Jan 19, 2009 at 03:49:20PM +0200, Avi Kivity wrote:
This probably comes up once in a while, thought I'd raise it again.
I'd like to suggest switching the default kernel to -PAE on machines that support it, for the following reasons:
- many machines have 4GB+ these days, even desktops
- NX is only available with -PAE, improves security
- kvm is significantly faster on AMD when PAE is selected (since we
don't support NPT on non-PAE)
What's needed to set this by default is changes in anaconda. They have their own list at anaconda-list@redhat.com
Unless we keep the non-PAE i586 kernel around as a fallback, we're not going to be able to boot on a whole raft of crappy i386 chips (original Pentium M most notably...)
I'm not suggesting dropping non-PAE. Simply defaulting to PAE where possible.
Are Pentium Ms (really the memory that comes with them) actually capable of running recent Fedoras? I'm talking desktop, not I'm-using-my-laptop-as-a-firewall-just-because-I-can.
On Mon, 2009-01-19 at 19:31 +0200, Avi Kivity wrote:
Kyle McMartin wrote:
On Mon, Jan 19, 2009 at 11:23:26AM -0500, Bill Nottingham wrote:
Dave Jones (davej@redhat.com) said:
On Mon, Jan 19, 2009 at 03:49:20PM +0200, Avi Kivity wrote:
This probably comes up once in a while, thought I'd raise it again.
I'd like to suggest switching the default kernel to -PAE on machines that support it, for the following reasons:
- many machines have 4GB+ these days, even desktops
- NX is only available with -PAE, improves security
- kvm is significantly faster on AMD when PAE is selected (since we
don't support NPT on non-PAE)
What's needed to set this by default is changes in anaconda. They have their own list at anaconda-list@redhat.com
Unless we keep the non-PAE i586 kernel around as a fallback, we're not going to be able to boot on a whole raft of crappy i386 chips (original Pentium M most notably...)
I'm not suggesting dropping non-PAE. Simply defaulting to PAE where possible.
Are Pentium Ms (really the memory that comes with them) actually capable of running recent Fedoras? I'm talking desktop, not I'm-using-my-laptop-as-a-firewall-just-because-I-can.
I've got a P3 (Coppermine) with 256M memory running F10. My significant other took it with her to Antarctica (Well F9 has been to Antarctica but it'll be F10 in Antarctica next month). You can only run one app at a time and have to be patient, but it's perfectly usable (and noone cares if this laptop is lost, stolen or destroyed [aside from her being pissed she lost all her research data]). I wouldn't/couldn't to use it as a daily machine, so while I'm in favor of -PAE default, F10 is "usable" on such small machines. I don't care if old machines need some bit twiddling to get to work, but we aren't dead yet :)
-Eric
Eric Paris wrote:
I've got a P3 (Coppermine) with 256M memory running F10. My significant other took it with her to Antarctica (Well F9 has been to Antarctica but it'll be F10 in Antarctica next month). You can only run one app at a time and have to be patient, but it's perfectly usable (and noone cares if this laptop is lost, stolen or destroyed [aside from her being pissed she lost all her research data]). I wouldn't/couldn't to use it as a daily machine, so while I'm in favor of -PAE default, F10 is "usable" on such small machines. I don't care if old machines need some bit twiddling to get to work, but we aren't dead yet :)
By F12 you'll be down to zero apps at the same time, and slow...
We can keep the non-PAE kernel, but as non-default in recognition that technology has moved on.
On Tuesday 20 January 2009 01:06:17 am Avi Kivity wrote:
Eric Paris wrote:
I've got a P3 (Coppermine) with 256M memory running F10. My significant other took it with her to Antarctica (Well F9 has been to Antarctica but it'll be F10 in Antarctica next month). You can only run one app at a time and have to be patient, but it's perfectly usable (and noone cares if this laptop is lost, stolen or destroyed [aside from her being pissed she lost all her research data]). I wouldn't/couldn't to use it as a daily machine, so while I'm in favor of -PAE default, F10 is "usable" on such small machines. I don't care if old machines need some bit twiddling to get to work, but we aren't dead yet :)
By F12 you'll be down to zero apps at the same time, and slow...
We can keep the non-PAE kernel, but as non-default in recognition that technology has moved on.
I have a P1 with 128M of memory running F9 just fine, I just keep it at runlevel 3. It works great.
On Tue, Jan 20, 2009 at 11:06:17AM +0200, Avi Kivity wrote:
Eric Paris wrote:
I've got a P3 (Coppermine) with 256M memory running F10. My significant other took it with her to Antarctica (Well F9 has been to Antarctica but it'll be F10 in Antarctica next month). You can only run one app at a time and have to be patient, but it's perfectly usable (and noone cares if this laptop is lost, stolen or destroyed [aside from her being pissed she lost all her research data]). I wouldn't/couldn't to use it as a daily machine, so while I'm in favor of -PAE default, F10 is "usable" on such small machines. I don't care if old machines need some bit twiddling to get to work, but we aren't dead yet :)
By F12 you'll be down to zero apps at the same time, and slow...
We can keep the non-PAE kernel, but as non-default in recognition that technology has moved on.
Look, I'm sorry if I'm just not thinking big picture enough here, but what exactly is the use case for a PAE kernel these days? The compat code in x86_64 should be more than good enough for the apps that require an i686 chroot.
I just don't see the status quo as doing any real harm, as the only generations of CPU that benefit are really P4 (which aren't worth the electricity used to power them) or Core (One) Duo (which didn't exist for a particularly long time...) Neither of which actually supported more than 3GB of RAM on their northbridges except for the Xeon chipsets anyway.
I have no idea what the installer and livecd do, but to me, it would seem to be a waste of space to carry two sets of installable kernels on the discs, when one would do. That said again, I'm suprised we aren't installing i586 kernels by default... Odd.
I think the ideal solution here is to support x86_64 kernel, i686 userspace more actively.
What, honestly, are the odds of someone with a bunch of P4 Xeons these days with 32GB of ram running Fedora? Are there really enough of them that it's worth caring? ;-)
Of course, take what I say with a grain of salt. I don't particularly care at all, I'm just trying to play the pragmatist.
Another question is what's the perf penalty of going to PAE on a 2GB of ram machine versus the vanilla HIGHMEM4G config?
The only argument I really buy into is the NX one, honestly...
What about a yum plugin that recommends a kernel that the user could override? I'll poke at it this afternoon (hey, I've always wanted to learn python...)
cheers, Kyle
Kyle McMartin wrote:
Look, I'm sorry if I'm just not thinking big picture enough here, but what exactly is the use case for a PAE kernel these days? The compat code in x86_64 should be more than good enough for the apps that require an i686 chroot.
It's certainly very good. I converted my i386 install to an x86_64 one, and the intermediate step of running the i386 userspace on x86_64 kernel worked well.
I just don't see the status quo as doing any real harm, as the only generations of CPU that benefit are really P4 (which aren't worth the electricity used to power them) or Core (One) Duo (which didn't exist for a particularly long time...) Neither of which actually supported more than 3GB of RAM on their northbridges except for the Xeon chipsets anyway.
I have no idea what the installer and livecd do, but to me, it would seem to be a waste of space to carry two sets of installable kernels on the discs, when one would do. That said again, I'm suprised we aren't installing i586 kernels by default... Odd.
I think the ideal solution here is to support x86_64 kernel, i686 userspace more actively.
I'm all in favor of pushing x86_64. But I think currently most installs are i386.
What, honestly, are the odds of someone with a bunch of P4 Xeons these days with 32GB of ram running Fedora? Are there really enough of them that it's worth caring? ;-)
Of course, take what I say with a grain of salt. I don't particularly care at all, I'm just trying to play the pragmatist.
Another question is what's the perf penalty of going to PAE on a 2GB of ram machine versus the vanilla HIGHMEM4G config?
I'm guessing, pretty low.
The only argument I really buy into is the NX one, honestly...
What about a yum plugin that recommends a kernel that the user could override? I'll poke at it this afternoon (hey, I've always wanted to learn python...)
Users won't be running yum. They're running that applet thing.
Kyle McMartin wrote:
On Tue, Jan 20, 2009 at 07:28:10PM +0200, Avi Kivity wrote:
Users won't be running yum. They're running that applet thing.
Which just shells out to yum... ;-)
Great, but any 'suggestions' need to find their way to the user.
I'd prefer to do the right thing in the first place rather than offer suggestions about the kernel. What's a kernel anyway?
2009/1/20 Kyle McMartin kyle@infradead.org:
On Tue, Jan 20, 2009 at 11:06:17AM +0200, Avi Kivity wrote:
Eric Paris wrote:
I've got a P3 (Coppermine) with 256M memory running F10. My significant other took it with her to Antarctica (Well F9 has been to Antarctica but it'll be F10 in Antarctica next month). You can only run one app at a time and have to be patient, but it's perfectly usable (and noone cares if this laptop is lost, stolen or destroyed [aside from her being pissed she lost all her research data]). I wouldn't/couldn't to use it as a daily machine, so while I'm in favor of -PAE default, F10 is "usable" on such small machines. I don't care if old machines need some bit twiddling to get to work, but we aren't dead yet :)
By F12 you'll be down to zero apps at the same time, and slow...
We can keep the non-PAE kernel, but as non-default in recognition that technology has moved on.
Look, I'm sorry if I'm just not thinking big picture enough here, but what exactly is the use case for a PAE kernel these days? The compat code in x86_64 should be more than good enough for the apps that require an i686 chroot.
I just don't see the status quo as doing any real harm, as the only generations of CPU that benefit are really P4 (which aren't worth the electricity used to power them) or Core (One) Duo (which didn't exist for a particularly long time...) Neither of which actually supported more than 3GB of RAM on their northbridges except for the Xeon chipsets anyway.
I have no idea what the installer and livecd do, but to me, it would seem to be a waste of space to carry two sets of installable kernels on the discs, when one would do. That said again, I'm suprised we aren't installing i586 kernels by default... Odd.
I think the ideal solution here is to support x86_64 kernel, i686 userspace more actively.
What, honestly, are the odds of someone with a bunch of P4 Xeons these days with 32GB of ram running Fedora? Are there really enough of them that it's worth caring? ;-)
Of course, take what I say with a grain of salt. I don't particularly care at all, I'm just trying to play the pragmatist.
Another question is what's the perf penalty of going to PAE on a 2GB of ram machine versus the vanilla HIGHMEM4G config?
The only argument I really buy into is the NX one, honestly...
What about a yum plugin that recommends a kernel that the user could override? I'll poke at it this afternoon (hey, I've always wanted to learn python...)
May I point out that those that care enough to want PAE usually know how to go about getting it enabled whereas those that have install failure because they're running non-PAE hardware probably wont know how to go about getting it disabled.
The fall-out from this going onto the livecd makes me shudder.
The original argument that many machines have 4GB of memory is simply false. Manufacturers aren't shipping anything more than 2GB on desktops at most unless you have oodles of money to throw at a Alienware box or something. Sure, servers come with more but Fedora is not really a reality for a long term server O.S.
Christopher Brown wrote:
May I point out that those that care enough to want PAE usually know how to go about getting it enabled whereas those that have install failure because they're running non-PAE hardware probably wont know how to go about getting it disabled.
You mean, ordinary users don't care about security? Because that's one of the advantages that PAE brings.
You're right, they don't care, we have to care for them.
The fall-out from this going onto the livecd makes me shudder.
You're pushing out a development problem to the users.
The original argument that many machines have 4GB of memory is simply false.
My ~3yo home box has 4GB. I'm not an ordinary user (or it would be a computer, not a "box"), but I don't think you can claim 4GB is rare.
Manufacturers aren't shipping anything more than 2GB on desktops at most unless you have oodles of money to throw at a Alienware box or something. Sure, servers come with more but Fedora is not really a reality for a long term server O.S
Servers should use x86_64 anyway. But I strongly disagree about penalizing the future to cater for the past.
2009/1/21 Avi Kivity avi@redhat.com:
Christopher Brown wrote:
May I point out that those that care enough to want PAE usually know how to go about getting it enabled whereas those that have install failure because they're running non-PAE hardware probably wont know how to go about getting it disabled.
You mean, ordinary users don't care about security? Because that's one of the advantages that PAE brings.
No, I meant ordinary users don't care about anything over 2GB.
You're right, they don't care, we have to care for them.
They do care about security but want it to be easy. Or simply don't want to have to care.
The fall-out from this going onto the livecd makes me shudder.
You're pushing out a development problem to the users.
Um, no. I'm simply against cutting out a tranche of people because of the needs of the few. We have x86_64 Live anyway.
The original argument that many machines have 4GB of memory is simply false.
My ~3yo home box has 4GB. I'm not an ordinary user (or it would be a computer, not a "box"), but I don't think you can claim 4GB is rare.
Rare for this use case, yes.
Manufacturers aren't shipping anything more than 2GB on desktops at most unless you have oodles of money to throw at a Alienware box or something. Sure, servers come with more but Fedora is not really a reality for a long term server O.S
Servers should use x86_64 anyway. But I strongly disagree about penalizing the future to cater for the past.
I'm not sure you're penalizing the past, I think this is penalizing the present.
On Wed, Jan 21, 2009 at 09:15:24AM +0000, Christopher Brown wrote:
The original argument that many machines have 4GB of memory is simply false. Manufacturers aren't shipping anything more than 2GB on desktops at most unless you have oodles of money to throw at a Alienware box or something. Sure, servers come with more but Fedora is not really a reality for a long term server O.S.
Actually, it works fine. I have a server that has only ever run Fedora.
josh
On Wed, Jan 21, 2009 at 09:15:24AM +0000, Christopher Brown wrote:
The fall-out from this going onto the livecd makes me shudder.
Jesse explained this to me,
To clarify: the anaconda installer is i586, and all 3 kernel flavours are shipped on the disc. I would imagine the livecd is i586 and will remain that way.
So it's really a moot point, and the only difference is installing PAE by default on machines which support it versus the i686 kernel.
Since we're not giving anything up on the disc, then there's no big deal.
The NX argument is alright, but we still ship execshield anyway, so it's (somewhat) mitigated.
Anyway, cheers. Kyle
On Wednesday, January 21 2009, Kyle McMartin said:
On Wed, Jan 21, 2009 at 09:15:24AM +0000, Christopher Brown wrote:
The fall-out from this going onto the livecd makes me shudder.
Jesse explained this to me,
To clarify: the anaconda installer is i586, and all 3 kernel flavours are shipped on the disc. I would imagine the livecd is i586 and will remain that way.
No, the livecd is i686. Because due to other things about the livecd, you have to have more ram than most i586s would support.
The number of i586 users we have is minimal. The number of non-pae i686 users we have is much more significant.
Jeremy
On Wed, 2009-01-21 at 09:27 -0500, Kyle McMartin wrote:
On Wed, Jan 21, 2009 at 09:15:24AM +0000, Christopher Brown wrote:
The fall-out from this going onto the livecd makes me shudder.
Jesse explained this to me,
To clarify: the anaconda installer is i586, and all 3 kernel flavours are shipped on the disc. I would imagine the livecd is i586 and will remain that way.
Sorry to disappoint. The live image is a different beast than the installer. The live image is explicitly i686. It will not run anywhere else as there is only one kernel image there to boot. Since live install is essentially a dd of the image to your new filesystem, there is only one kernel there as well.
So without some more context, I can't tell you what kind of doom the live image is in for... (:
So it's really a moot point, and the only difference is installing PAE by default on machines which support it versus the i686 kernel.
Since we're not giving anything up on the disc, then there's no big deal.
The NX argument is alright, but we still ship execshield anyway, so it's (somewhat) mitigated.
Anyway, cheers. Kyle
Avi Kivity (avi@redhat.com) said:
Are Pentium Ms (really the memory that comes with them) actually capable of running recent Fedoras? I'm talking desktop, not I'm-using-my-laptop-as-a-firewall-just-because-I-can.
Sure, I had a T40 that had 1.5GB of memory in it, and it could have taken more.
Bill
On Mon, 19 Jan 2009 19:31:12 +0200, Avi Kivity avi@redhat.com wrote:
Are Pentium Ms (really the memory that comes with them) actually capable of running recent Fedoras? I'm talking desktop, not I'm-using-my-laptop-as-a-firewall-just-because-I-can.
If only it were laptops. I think they still sell systems like this:
[zaitcev@mallorn ~]$ more /proc/cpuinfo processor : 0 vendor_id : CentaurHauls cpu family : 6 model : 7 model name : VIA Samuel 2 stepping : 3 cpu MHz : 800.065 cache size : 64 KB fdiv_bug : no hlt_bug : no f00f_bug : no coma_bug : no fpu : yes fpu_exception : yes cpuid level : 1 wp : yes flags : fpu de tsc msr cx8 mtrr pge mmx 3dnow bogomips : 1601.64 clflush size : 32
[zaitcev@mallorn ~]$
You point about non-default kernel is well taken though. I'm not against making PAE default as long as some king of alternative is provided.
BTW, does anyone know if the Geode in OLPC XO has PAE?
-- Pete
On Mon, Jan 19, 2009 at 11:24:14AM -0700, Pete Zaitcev wrote:
On Mon, 19 Jan 2009 19:31:12 +0200, Avi Kivity avi@redhat.com wrote:
Are Pentium Ms (really the memory that comes with them) actually capable of running recent Fedoras? I'm talking desktop, not I'm-using-my-laptop-as-a-firewall-just-because-I-can.
If only it were laptops. I think they still sell systems like this:
[zaitcev@mallorn ~]$ more /proc/cpuinfo processor : 0 vendor_id : CentaurHauls cpu family : 6 model : 7 model name : VIA Samuel 2
The samuel 2 stopped being made ~5 years ago. VIA have been PAE capable since the Nehemiah generation.
Dave
On Mon, 2009-01-19 at 11:24 -0700, Pete Zaitcev wrote:
BTW, does anyone know if the Geode in OLPC XO has PAE?
This is off a C2:
[olpc@xo-10-DD-46 ~]$ cat /proc/cpuinfo processor : 0 vendor_id : AuthenticAMD cpu family : 5 model : 10 model name : Geode(TM) Integrated Processor by AMD PCS stepping : 2 cpu MHz : 430.931 cache size : 128 KB fdiv_bug : no hlt_bug : no f00f_bug : no coma_bug : no fpu : yes fpu_exception : yes cpuid level : 1 wp : yes flags : fpu de pse tsc msr cx8 sep pge cmov clflush mmx mmxext 3dnowext 3dnow bogomips : 863.19 clflush size : 32
On Mon, Jan 19, 2009 at 07:31:12PM +0200, Avi Kivity wrote:
Kyle McMartin wrote:
On Mon, Jan 19, 2009 at 11:23:26AM -0500, Bill Nottingham wrote:
Dave Jones (davej@redhat.com) said:
On Mon, Jan 19, 2009 at 03:49:20PM +0200, Avi Kivity wrote:
This probably comes up once in a while, thought I'd raise it again.
I'd like to suggest switching the default kernel to -PAE on
machines > that support it, for the following reasons:
- many machines have 4GB+ these days, even desktops
- NX is only available with -PAE, improves security
- kvm is significantly faster on AMD when PAE is selected (since
we > don't support NPT on non-PAE)
What's needed to set this by default is changes in anaconda. They have their own list at anaconda-list@redhat.com
Unless we keep the non-PAE i586 kernel around as a fallback, we're not going to be able to boot on a whole raft of crappy i386 chips (original Pentium M most notably...)
I'm not suggesting dropping non-PAE. Simply defaulting to PAE where possible.
Are Pentium Ms (really the memory that comes with them) actually capable of running recent Fedoras? I'm talking desktop, not I'm-using-my-laptop-as-a-firewall-just-because-I-can.
Absolutely, I've got a Thinkpad T42 here that does just fine on fedora 10. Unless of course, I try to load a PAE enabled kernel on it.
I've not looked into it at all, but this thread got me thinking, is there any particular reason that we can't merge the pae and non-pae kernels using the same alternatives approach we used to merge smp & up?
Neil
/***************************************** *Neil Horman nhorman@redhat.com *919-754-4323 ****************************************/
On Mon, Jan 19, 2009 at 01:36:08PM -0500, Neil Horman wrote:
Absolutely, I've got a Thinkpad T42 here that does just fine on fedora 10. Unless of course, I try to load a PAE enabled kernel on it.
I've not looked into it at all, but this thread got me thinking, is there any particular reason that we can't merge the pae and non-pae kernels using the same alternatives approach we used to merge smp & up?
I looked into this several years ago, it's actually fairly gnarly since depending on PAE we set up the swapper page tables differently, and other ugly differences.
I should resurrect the patch set, though I think rationalizing it with the Xen merge might be more pain than it's worth... last time I looked at it there was a ridiculous amount of rejects.
It was pretty close to booting (though I had to hack the bootloader to get a PAE enable bit so people without the bit could turn it off.)
regards, Kyle
Kyle McMartin wrote:
On Mon, Jan 19, 2009 at 01:36:08PM -0500, Neil Horman wrote:
Absolutely, I've got a Thinkpad T42 here that does just fine on fedora 10. Unless of course, I try to load a PAE enabled kernel on it.
I've not looked into it at all, but this thread got me thinking, is there any particular reason that we can't merge the pae and non-pae kernels using the same alternatives approach we used to merge smp & up?
I looked into this several years ago, it's actually fairly gnarly since depending on PAE we set up the swapper page tables differently, and other ugly differences.
I should resurrect the patch set, though I think rationalizing it with the Xen merge might be more pain than it's worth... last time I looked at it there was a ridiculous amount of rejects.
Well, it depends on if you are talking about the non-upstream Xen patch, or the current upstream pv_ops. The former is all but dead, so you don't have to worry about that, and the latter *should* (famous last words) mostly stay out of your way through pv_ops. Though I haven't gone in and implemented it myself, so it's easy for me to say :).
I think choosing PAE at runtime would be the ideal situation, if we can get there.
Avi Kivity wrote:
Kyle McMartin wrote:
Unless we keep the non-PAE i586 kernel around as a fallback, we're not going to be able to boot on a whole raft of crappy i386 chips (original Pentium M most notably...)
I'm not suggesting dropping non-PAE. Simply defaulting to PAE where possible.
Are Pentium Ms (really the memory that comes with them) actually capable of running recent Fedoras?
Well, we can switch the default for F11, then watch the smolt statistics to figure ;)
cheers, Gerd
On Monday, January 19 2009, Bill Nottingham said:
Dave Jones (davej@redhat.com) said:
On Mon, Jan 19, 2009 at 03:49:20PM +0200, Avi Kivity wrote:
This probably comes up once in a while, thought I'd raise it again.
I'd like to suggest switching the default kernel to -PAE on machines that support it, for the following reasons:
- many machines have 4GB+ these days, even desktops
- NX is only available with -PAE, improves security
- kvm is significantly faster on AMD when PAE is selected (since we
don't support NPT on non-PAE)
What's needed to set this by default is changes in anaconda. They have their own list at anaconda-list@redhat.com
anaconda-devel-list, actually. But I could have sworn we already did this.
Well, we're not going to make the change unless kernel guys want us to.
That said, we currently do install the PAE kernel if you have 4 GB+ of RAM[1]. Switching to it by default is problematic because then we're back to using different kernels for different cases and it also makes the 'what do you with the live image' case a lot more complex.
The _real_ fix here is to get PAE runtime much like was finally done with SMP :-)
Jeremy
[1] Well, any memory above 4 GB actually -- so if you have a machine with a large hole and only 3 GB of RAM, you could end up with the PAE kernel
Jeremy Katz wrote:
That said, we currently do install the PAE kernel if you have 4 GB+ of RAM[1]. Switching to it by default is problematic because then we're back to using different kernels for different cases
You have that now, don't you? One case for <4GB and one for >=4GB. Worse, if you install more memory, the kernel doesn't see it.
Downgrading your CPU to one which does not support PAE should be rare.
and it also makes the 'what do you with the live image' case a lot more complex.
I'd just go with PAE here.
The _real_ fix here is to get PAE runtime much like was finally done with SMP :-)
Patches, as they say, are welcome. But you could install both kernels and have the bootloaded choose (sticks wax balls into ears).
On Monday, January 19 2009, Avi Kivity said:
Jeremy Katz wrote:
That said, we currently do install the PAE kernel if you have 4 GB+ of RAM[1]. Switching to it by default is problematic because then we're back to using different kernels for different cases
You have that now, don't you? One case for <4GB and one for >=4GB. Worse, if you install more memory, the kernel doesn't see it.
Downgrading your CPU to one which does not support PAE should be rare.
Yes, but at least the "running a different kernel" case is currently the relatively rare one.
and it also makes the 'what do you with the live image' case a lot more complex.
I'd just go with PAE here.
Can't do so -- the live image is definitely used on a lot of hardware that isn't PAE capable. Many/most Pentium M's didn't support it, the OLPC doesn't[1]. And those are common hardware targets for the live image
The _real_ fix here is to get PAE runtime much like was finally done with SMP :-)
Patches, as they say, are welcome.
Low-level x86 setup code isn't quite my forte... Hence I go for goading others into doing it ;-)
But you could install both kernels and have the bootloaded choose (sticks wax balls into ears).
Want to write code for syslinux and grub to do the auto-choosing? Then we also have to figure out a way to shoe-horn another 50 MB of stuff into the already full live image
Jeremy
Jeremy Katz wrote:
On Monday, January 19 2009, Avi Kivity said:
Jeremy Katz wrote:
That said, we currently do install the PAE kernel if you have 4 GB+ of RAM[1]. Switching to it by default is problematic because then we're back to using different kernels for different cases
You have that now, don't you? One case for <4GB and one for >=4GB. Worse, if you install more memory, the kernel doesn't see it.
Downgrading your CPU to one which does not support PAE should be rare.
Yes, but at least the "running a different kernel" case is currently the relatively rare one.
It will become more and more frequent as machines get beefier. If we switch to PAE, it will be the non-PAE case which is rare.
btw, what's wrong with running a different kernel? So long as the configs match, there shouldn't be a difference in reliability.
and it also makes the 'what do you with the live image' case a lot more complex.
I'd just go with PAE here.
Can't do so -- the live image is definitely used on a lot of hardware that isn't PAE capable. Many/most Pentium M's didn't support it, the OLPC doesn't[1]. And those are common hardware targets for the live image
So either put both kernels there, or the non-PAE kernel.
The _real_ fix here is to get PAE runtime much like was finally done with SMP :-)
Patches, as they say, are welcome.
Low-level x86 setup code isn't quite my forte... Hence I go for goading others into doing it ;-)
I meant, "this requires a monumental amount of work".
But you could install both kernels and have the bootloaded choose (sticks wax balls into ears).
Want to write code for syslinux and grub to do the auto-choosing?
Not really. I don't see it as particularly difficult.
With grub, you could use multiboot to add a module to select the correct kernel, maybe.
I think Windows works this way, with the /PAE switch on the kernel command line causing a different kernel to be loaded.
Then we also have to figure out a way to shoe-horn another 50 MB of stuff into the already full live image
Cut away the drivers for Infiniband and similarly useless hardware (for livecds, that is). Put them in a separate subpackage or use steam-powered rm -rf.
Bill Nottingham wrote:
Dave Jones (davej@redhat.com) said:
On Mon, Jan 19, 2009 at 03:49:20PM +0200, Avi Kivity wrote:
This probably comes up once in a while, thought I'd raise it again.
I'd like to suggest switching the default kernel to -PAE on machines that support it, for the following reasons:
- many machines have 4GB+ these days, even desktops
- NX is only available with -PAE, improves security
- kvm is significantly faster on AMD when PAE is selected (since we
don't support NPT on non-PAE)
What's needed to set this by default is changes in anaconda. They have their own list at anaconda-list@redhat.com
anaconda-devel-list, actually. But I could have sworn we already did this.
Just tested; F10 i386 still installs a non-PAE kernel.
kernel@lists.fedoraproject.org
-
Avi Kivity -
Bill Nottingham -
Chris Lalancette -
Christopher Brown -
Conrad Meyer -
Dave Jones -
Eric Paris -
Gerd Hoffmann -
Ignacio Vazquez-Abrams -
Jeremy Katz -
Jesse Keating -
Josh Boyer -
Kyle McMartin -
Neil Horman -
Pete Zaitcev