[opensuse] Have 4G, any reason to use x86_64 over i386?
I have an Intel cpu / 64-bit extensions on a machine with 4G of memory. Just want some confirmation (or not)...it seems like i386 would currently be my best choice: all memory is accessible, no program I am running needs >2G addr space, 32-bit programs execute faster and will likely be smaller than the equivalent 64-bit programs (and libraries). As long as I have under 4G mem (and my apps need <2-3G), I see only "downsides" to moving to the x86-64 version. Am I missing something or is this pretty much the "consensus view" (assuming there is one :-)). Thanks, Linda -- To unsubscribe, e-mail: opensuse+unsubscribe@opensuse.org For additional commands, e-mail: opensuse+help@opensuse.org
Linda Walsh wrote:
I have an Intel cpu / 64-bit extensions on a machine with 4G of memory. Just want some confirmation (or not)...it seems like i386 would currently be my best choice: all memory is accessible, no program I am running needs >2G addr space, 32-bit programs execute faster and will likely be smaller than the equivalent 64-bit programs (and libraries).
As long as I have under 4G mem (and my apps need <2-3G), I see only "downsides" to moving to the x86-64 version. Am I missing something or is this pretty much the "consensus view" (assuming there is one :-)).
What downsides are you referring to? The only one I'm aware of is 32 bit only plugins for browsers and that can be fixed by installing a 32 bit browser. -- To unsubscribe, e-mail: opensuse+unsubscribe@opensuse.org For additional commands, e-mail: opensuse+help@opensuse.org
James Knott wrote:
Linda Walsh wrote:
I have an Intel cpu / 64-bit extensions on a machine with 4G of memory. Just want some confirmation (or not)...it seems like i386 would currently be my best choice: all memory is accessible, no program I am running needs >2G addr space, 32-bit programs execute faster and will likely be smaller than the equivalent 64-bit programs (and libraries).
As long as I have under 4G mem (and my apps need <2-3G), I see only "downsides" to moving to the x86-64 version. Am I missing something or is this pretty much the "consensus view" (assuming there is one :-)). What downsides are you referring to? The only one I'm aware of is 32 bit only plugins for browsers and that can be fixed by installing a 32 bit browser.
Cool! Lack of consensus. :-) Hmmm... How difficult is it to run 32bit programs? Don't you have to install separate directories to store the libraries? I thought 64-bit programs were a bit larger? I guess the speed differences between 32 & 64 are alot more mixed than I thought. It's hard to run good benchmarks. Maybe I should have been more neutral in my comment and ask what would be benefits of running 64bit vs. 32? :-) This is assuming 4GB memory, not alot of floating point, no programs need more than 2GB...maybe I should install x86_64 -- I've been happy with the 32bit linux and its performance, but I suppose it could be even faster....:-) -linda -- To unsubscribe, e-mail: opensuse+unsubscribe@opensuse.org For additional commands, e-mail: opensuse+help@opensuse.org
On Thursday 25 January 2007 00:33, Linda Walsh wrote:
32-bit programs execute faster
Have you actually tried this? In many cases it is true, but it is heavily dependent on the application, and how well it has been optimised for 64 bit. I would suggest that you do your own benchmarks with the applications you intend to use -- To unsubscribe, e-mail: opensuse+unsubscribe@opensuse.org For additional commands, e-mail: opensuse+help@opensuse.org
On Wednesday 24 January 2007 14:44, Anders Johansson wrote:
On Thursday 25 January 2007 00:33, Linda Walsh wrote:
32-bit programs execute faster
Have you actually tried this? In many cases it is true, but it is heavily dependent on the application, and how well it has been optimised for 64 bit.
I would suggest that you do your own benchmarks with the applications you intend to use
Its also dependent on the processor. Core 2 Duo timeings are such that the most frequently used 64bit instructions execute in the same number of clocks as their 32bit version in spite of having to fetch twice the data from memory. The memory fetches are also just as fast as long as the processor correctly predicted any branching. I don't know about the AMD 64bit processors. -- _____________________________________ John Andersen
On Wednesday 24 January 2007 14:33, Linda Walsh wrote:
As long as I have under 4G mem (and my apps need <2-3G), I see only "downsides" to moving to the x86-64 version. Am I missing something or is this pretty much the "consensus view" (assuming there is one :-)).
Refer to the archives. Many users report up to a 30% performance improvement using 64bit installs. Mine are not that great but it is still quite a perceptible improvement using 64bit. I first installed the 32bit version 10.2 because that is what I had at hand, but then when the boxed set arrived I replaced that and it was noticeably improved. -- _____________________________________ John Andersen
John Andersen wrote:
On Wednesday 24 January 2007 14:33, Linda Walsh wrote:
As long as I have under 4G mem (and my apps need <2-3G), I see only "downsides" to moving to the x86-64 version. Am I missing something or is this pretty much the "consensus view" (assuming there is one :-)).
Refer to the archives. Many users report up to a 30% performance improvement using 64bit installs. Mine are not that great but it is still quite a perceptible improvement using 64bit. I first installed the 32bit version 10.2 because that is what I had at hand, but then when the boxed set arrived I replaced that and it was noticeably improved.
The best way to find out, is to run benchmarks on the same hardware. -- To unsubscribe, e-mail: opensuse+unsubscribe@opensuse.org For additional commands, e-mail: opensuse+help@opensuse.org
On Thursday 25 January 2007 03:08, James Knott wrote:
Refer to the archives. Many users report up to a 30% performance improvement using 64bit installs. Mine are not that great but it is still quite a perceptible improvement using 64bit. I first installed the 32bit version 10.2 because that is what I had at hand, but then when the boxed set arrived I replaced that and it was noticeably improved.
The best way to find out, is to run benchmarks on the same hardware.
Exactly, but we needed the server in production, and I couldn't justify the extra time to examine an issue I didn't even think of as important. I just noticed improved performance. Others ran glxgears and reported on another thread. See archives. -- _____________________________________ John Andersen -- To unsubscribe, e-mail: opensuse+unsubscribe@opensuse.org For additional commands, e-mail: opensuse+help@opensuse.org
On Jan 24, 07 15:33:42 -0800, Linda Walsh wrote:
I have an Intel cpu / 64-bit extensions on a machine with 4G of memory.
Just want some confirmation (or not)...it seems like i386 would currently be my best choice: all memory is accessible, no program I am running needs >2G addr space, 32-bit programs execute faster and will likely be smaller than the equivalent 64-bit programs (and libraries).
With i386 you will only be able to access 3 or 3.5GB of your memory. At least half a megabyte will be wasted. Background: The PCI cards have to be mapped into the available 32bit address space. 64-bit programs actually execute faster than 32-bit, but need more memory (all pointers have double size). This is counterintuitive, but can be explained: the x86_64 architecture has more free registers available than i586. Only programs that trash memory a lot (inlcuding lots of pointers) might run slower due to higher memory throughput. You can run i586 programs on a x86_64 kernel, in fact this is very much advised for firefox (stabiliy, plugins) and video players (due to win32codecs). AFAIK openoffice isn't even ported to x86_64 yet. Other ports are not really clean yet, so the i586 binaries are often more stable. So: Downside for x86_64: stability, complexity (mixed architecture setup), more memory needed Upside for x86_64: faster, more memory available HTH Matthias -- Matthias Hopf <mhopf@suse.de> __ __ __ Maxfeldstr. 5 / 90409 Nuernberg (_ | | (_ |__ mat@mshopf.de Phone +49-911-74053-715 __) |_| __) |__ R & D www.mshopf.de -- To unsubscribe, e-mail: opensuse+unsubscribe@opensuse.org For additional commands, e-mail: opensuse+help@opensuse.org
Matthias Hopf wrote:
You can run i586 programs on a x86_64 kernel, in fact this is very much advised for firefox (stabiliy, plugins) and video players (due to win32codecs). AFAIK openoffice isn't even ported to x86_64 yet.
ftp://ftp.suse.com/pub/projects/OpenOffice.org/10.2-x86_64 -- To unsubscribe, e-mail: opensuse+unsubscribe@opensuse.org For additional commands, e-mail: opensuse+help@opensuse.org
On Jan 25, 07 14:45:18 -0500, James Knott wrote:
Matthias Hopf wrote:
You can run i586 programs on a x86_64 kernel, in fact this is very much advised for firefox (stabiliy, plugins) and video players (due to win32codecs). AFAIK openoffice isn't even ported to x86_64 yet.
ftp://ftp.suse.com/pub/projects/OpenOffice.org/10.2-x86_64
I should have said: the port isn't completed yet. I knew there were efforts in getting the x86_64 port as stable as the i586 version. Matthias -- Matthias Hopf <mhopf@suse.de> __ __ __ Maxfeldstr. 5 / 90409 Nuernberg (_ | | (_ |__ mat@mshopf.de Phone +49-911-74053-715 __) |_| __) |__ R & D www.mshopf.de -- To unsubscribe, e-mail: opensuse+unsubscribe@opensuse.org For additional commands, e-mail: opensuse+help@opensuse.org
Matthias Hopf wrote:
With i386 you will only be able to access 3 or 3.5GB of your memory. At least half a megabyte will be wasted. Background: The PCI cards have to be mapped into the available 32bit address space.
Has anyone actually seen this first hand with any Intel processor with PAE? I've never seen this and it doesn't sound right. I have seen evidence to support the contrary: from /proc/meminfo on my (i386) system: MemTotal: 4147236 kB MemFree: 1321928 kB Buffers: 220 kB Cached: 2424472 kB SwapCached: 0 kB Active: 967832 kB Inactive: 1488312 kB HighTotal: 3275304 kB HighFree: 1117840 kB LowTotal: 871932 kB LowFree: 204088 kB SwapTotal: 4200956 kB SwapFree: 4200956 kB Dirty: 0 kB Writeback: 0 kB AnonPages: 31292 kB Mapped: 20620 kB Slab: 350396 kB PageTables: 788 kB NFS_Unstable: 0 kB Bounce: 0 kB CommitLimit: 6274572 kB Committed_AS: 98708 kB VmallocTotal: 112632 kB VmallocUsed: 14668 kB VmallocChunk: 97292 kB HugePages_Total: 0 HugePages_Free: 0 HugePages_Rsvd: 0 Hugepagesize: 2048 kB -------------------------------- Perhaps its a problem on AMD based systems? I know it's a problem on WindowsXP -- Microsoft crippled XP in SP2 to limit it to 3G. It supported 4G in SP1 and earlier, though it still required a patch to work properly, but MS forced XP to 3G, supposedly in the name of "security", but more likely to create some need to upgrade XP to the Server edition or Vista (which, I assume, they'd claim was more secure, and could handle 4G).
64-bit programs actually execute faster than 32-bit, but need more memory (all pointers have double size). This is counterintuitive, but can be explained: the x86_64 architecture has more free registers available than i586. Only programs that trash memory a lot (inlcuding lots of pointers) might run slower due to higher memory throughput.
Yeah -- wondered about that -- seems "longs" and 'long double's (floats) take twice the space on 64-bit, but it only seems to add ~10% (have seen 25 to over 40% on some 32-64 bit comparisons for other arch's (not "ia" arch). Performance is mixed, but under 10% in all benchmarks I've seen, but some figures show integer performance might have a slight edge on ia32, but floats are faster, memory moves/loads seem uniformly faster.
You can run i586 programs on a x86_64 kernel, in fact this is very much advised for firefox (stabiliy, plugins) and video players (due to win32codecs). AFAIK openoffice isn't even ported to x86_64 yet. Other ports are not really clean yet, so the i586 binaries are often more stable.
I noticed the 64-bit libraries are under a different path. To switch to x86_64, shouldn't I be able to just install an x86_64 kernel initially then install/switch over apps at leisure?
So: Downside for x86_64: stability, complexity (mixed architecture setup), more memory needed Upside for x86_64: faster, more memory available
Well, I don't get more memory until I buy some -- since x86_64 takes slightly more memory, I'd have slightly less, "free", memory in the short term..., no? Maybe if I can install the x64 kernel and side-by-side 32/64-bit apps, I can see where I benefit with the 64-bit compile? thanks, -linda -- To unsubscribe, e-mail: opensuse+unsubscribe@opensuse.org For additional commands, e-mail: opensuse+help@opensuse.org
On Jan 25, 07 22:25:39 -0800, Linda Walsh wrote:
Matthias Hopf wrote:
With i386 you will only be able to access 3 or 3.5GB of your memory. At least half a megabyte will be wasted. Background: The PCI cards have to be mapped into the available 32bit address space.
Has anyone actually seen this first hand with any Intel processor with PAE? I've never seen this and it doesn't sound right. I have seen evidence to support the contrary: from /proc/meminfo on my (i386) system:
With PAE, this is a different story. Never had any Intel chip with hardware PAE support (only Xenon processors had that), and the software version of the P4 was too slow for me.
I noticed the 64-bit libraries are under a different path. To switch to x86_64, shouldn't I be able to just install an x86_64 kernel initially then install/switch over apps at leisure?
This should work. But you will need several packages from the very begining (module loading etc.) and the initial ramdisk has to be adapted as well.
Well, I don't get more memory until I buy some -- since x86_64 takes slightly more memory, I'd have slightly less, "free", memory in the short term..., no?
No for AMD/P4, because you have 4GB instead of 3.5GB available ;) Matthias -- Matthias Hopf <mhopf@suse.de> __ __ __ Maxfeldstr. 5 / 90409 Nuernberg (_ | | (_ |__ mat@mshopf.de Phone +49-911-74053-715 __) |_| __) |__ R & D www.mshopf.de -- To unsubscribe, e-mail: opensuse+unsubscribe@opensuse.org For additional commands, e-mail: opensuse+help@opensuse.org
Matthias Hopf wrote:
On Jan 25, 07 22:25:39 -0800, Linda Walsh wrote:
Matthias Hopf wrote:
With i386 you will only be able to access 3 or 3.5GB of your memory. At least half a megabyte will be wasted. Background: The PCI cards have to be mapped into the available 32bit address space.
Has anyone actually seen this first hand with any Intel processor with PAE? I've never seen this and it doesn't sound right. I have seen evidence to support the contrary: from /proc/meminfo on my (i386) system:
With PAE, this is a different story. Never had any Intel chip with hardware PAE support (only Xenon processors had that), and the software version of the P4 was too slow for me.
Software version of P4? Haven't heard of that. From the documentation I've found (googling), PAE was/is in any machine with an Intel Pentium Pro or later, though reading about it, it also seemed to involve 2M page sizes. I don't think 2M page sizes are supported or used on ia32 linux. I think the kernel may not be able to see a full 4GB at at the same time, but can use all the memory by mapping memory that would overlap the PCI memory into "high" memory and make use of it that way.
I noticed the 64-bit libraries are under a different path. To switch to x86_64, shouldn't I be able to just install an x86_64 kernel initially then install/switch over apps at leisure?
This should work. But you will need several packages from the very begining (module loading etc.) and the initial ramdisk has to be adapted as well.
Hmm....maybe I should just try making a kernel that doesn't need a ramdisk and has all needed modules compiled in. Will keep that in mind though...thanks for the heads up! Right now, I have so many modules (lsmod) that are loaded but I don't have hardware for -- why does it do that?! It may be they load during some hardware probing, but I wish they'd unload the unneeded part when they were done. Seems like that's another area for saving memory. Guess I need to go experiment^Wplay^Wresearch some more ... :-)... Thanks! Linda -- To unsubscribe, e-mail: opensuse+unsubscribe@opensuse.org For additional commands, e-mail: opensuse+help@opensuse.org
On Friday 26 January 2007 14:54, Linda Walsh wrote:
Hmm....maybe I should just try making a kernel that doesn't need a ramdisk and has all needed modules compiled in. Will keep that in mind though...thanks for the heads up!
Why bother? How much improvement do you thing you will gain? I stopped compiling kernels when Hubert Mantel started publishing his KOD, and only used the to gain access to stuff I could not get without it. Since 8.1 I have never found any compelling reason to compile a kernel on a norma workstation or server.
Right now, I have so many modules (lsmod) that are loaded but I don't have hardware for -- why does it do that?! It may be they load during some hardware probing, but I wish they'd unload the unneeded part when they were done. Seems like that's another area for saving memory.
Is that what this whole thread is about, a small memory machine? The title suggest you have 4gig, so why are you worrying about memory? I can't believe you come anywhere near using 4gig unless you carrying a huge database/server load, in which case there is probably a corresponding budget to by more memory. -- _____________________________________ John Andersen -- To unsubscribe, e-mail: opensuse+unsubscribe@opensuse.org For additional commands, e-mail: opensuse+help@opensuse.org
On Jan 26, 07 15:54:27 -0800, Linda Walsh wrote:
With PAE, this is a different story. Never had any Intel chip with hardware PAE support (only Xenon processors had that), and the software version of the P4 was too slow for me.
Software version of P4? Haven't heard of that.
No, PAE support is "software" (actually micro code) on P4, and only supported in hardware on Xeon. To the best of my knowledge. I think on Xeons it was with the P3 core already.
This should work. But you will need several packages from the very begining (module loading etc.) and the initial ramdisk has to be adapted as well.
Hmm....maybe I should just try making a kernel that doesn't need a ramdisk and has all needed modules compiled in. Will keep that in mind though...thanks for the heads up!
That's definitely an option. Still you might run into - err - surprising issues :-]
Guess I need to go experiment^Wplay^Wresearch some more ... :-)...
Good luck! Matthias -- Matthias Hopf <mhopf@suse.de> __ __ __ Maxfeldstr. 5 / 90409 Nuernberg (_ | | (_ |__ mat@mshopf.de Phone +49-911-74053-715 __) |_| __) |__ R & D www.mshopf.de -- To unsubscribe, e-mail: opensuse+unsubscribe@opensuse.org For additional commands, e-mail: opensuse+help@opensuse.org
On Wednesday 24 January 2007 15:33, Linda Walsh wrote:
I have an Intel cpu / 64-bit extensions on a machine with 4G of memory.
Just want some confirmation (or not)...it seems like i386 would currently be my best choice: all memory is accessible, no program I am running needs >2G addr space, 32-bit programs execute faster and will likely be smaller than the equivalent 64-bit programs (and libraries).
As long as I have under 4G mem (and my apps need <2-3G), I see only "downsides" to moving to the x86-64 version. Am I missing something or is this pretty much the "consensus view" (assuming there is one :-)).
Linda: From my experience, I have seen DRASTIC imporovements in large-scale database applications running on 64-bit machines vs. the 32 bit machines. Adding quad-core processors helps, too. :) This is like processing where the same 5-10 TB databases will be on the machines. However, that is on servers. I have only anecdotal information about workstations, which have been that 64 bit gives you a better "seat of the pants" feeling but nothing more, unless you're a hard-core researcher doing massive calculations. -- kai www.perfectreign.com || www.4thedadz.com www.filesite.org || www.donutmonster.com closing the doors that surround me so no one will ever penetrate complete my retreat just to wait for the day that never comes so i will laugh alone -- To unsubscribe, e-mail: opensuse+unsubscribe@opensuse.org For additional commands, e-mail: opensuse+help@opensuse.org
Linda Walsh wrote:
I have an Intel cpu / 64-bit extensions on a machine with 4G of memory. Just want some confirmation (or not)...it seems like i386 would currently be my best choice: all memory is accessible, no program I am running needs >2G addr space, 32-bit programs execute faster and will likely be smaller than the equivalent 64-bit programs (and libraries).
As long as I have under 4G mem (and my apps need <2-3G), I see only "downsides" to moving to the x86-64 version. Am I missing something or is this pretty much the "consensus view" (assuming there is one :-)).
Thanks, Linda
4GB is considered the break even point for 64bit. Realize the 32bit was only designed to access 2GB of ram and with some pointer magic OSs have been able to access more. It also depends on the program. Things with lots of math will use native 64bit math to commutate vs thunking 64bit numbers into two 32bit pointers. So anything graphic or index intensive (databases) will make a difference even with only 2GB of memory. In general, desktops will see mixed pluses and minuses of using 64bit. So far, over all, my system feels faster then when on 32bit. This is a windows system. I do install all new systems with 64bit, just because I know in the future 8GB and 16GB memory systems will be common place and I might as well get use to it now. The difference in speed is not going to kill me. -- To unsubscribe, e-mail: opensuse+unsubscribe@opensuse.org For additional commands, e-mail: opensuse+help@opensuse.org
I have an AMD Athlon 64 and found that for most apps I do not find a lot of difference in performance. However, when I start my video editing and encoding there is a noticeable difference. Jobs that took 3-4 hours on and Athlon XP 1800 now finish is 1-2 hours. There are certain places where you will notice a difference. -- John Registered Linux User 263680, get counted at http://counter.li.org -- To unsubscribe, e-mail: opensuse+unsubscribe@opensuse.org For additional commands, e-mail: opensuse+help@opensuse.org
participants (8)
-
Anders Johansson -
James Knott -
John Andersen -
John Pierce -
Kai Ponte -
Linda Walsh -
Matthias Hopf -
Tom Miller