2002-01-15 20:46:20

by Peter Zaitsev

[permalink] [raw]
Subject: 3.5G user space speed

Hello Linux,

2.4.xaa Series as well as SuSE kernels have 3.5G userspace option,
which seems to be quite useful, therefore I see it's not included
is stock kernel for some reasons. Also I've heard this
configuration may have some performance problems.

Can anyone comment on this topic ?

I need large amount of address space for my application but I also
need to get as much I/O performance as it's possible, so I can switch
to 3.0/1.0 memory distribution if it will benefit here.

--
Best regards,
Peter mailto:[email protected]


2002-01-15 21:06:20

by Brian Gerst

[permalink] [raw]
Subject: Re: 3.5G user space speed

Peter Zaitsev wrote:
>
> Hello Linux,
>
> 2.4.xaa Series as well as SuSE kernels have 3.5G userspace option,
> which seems to be quite useful, therefore I see it's not included
> is stock kernel for some reasons. Also I've heard this
> configuration may have some performance problems.
>
> Can anyone comment on this topic ?
>
> I need large amount of address space for my application but I also
> need to get as much I/O performance as it's possible, so I can switch
> to 3.0/1.0 memory distribution if it will benefit here.

You can't have it both ways with the x86 (speed vs. large userspace).
Kernel 2.5 may help a bit here because changes were made to allow DMA
from all memory (subject to card limitations), lessening the burden for
direct-mapped memory. Otherwise you'll need to move to a 64-bit arch.

--

Brian Gerst

2002-01-15 21:27:30

by Peter Zaitsev

[permalink] [raw]
Subject: Re[2]: 3.5G user space speed

Hello Brian,

Wednesday, January 16, 2002, 12:05:39 AM, you wrote:


Well. May be you can tell about the numbers a bit ? I can chose 3.0G
for user instead of 3.5G for user with not really huge loss, but I'd
like to know how much it will increase speed and in which cases, also
about standard 2/2 mode.

>>
>> 2.4.xaa Series as well as SuSE kernels have 3.5G userspace option,
>> which seems to be quite useful, therefore I see it's not included
>> is stock kernel for some reasons. Also I've heard this
>> configuration may have some performance problems.
>>
>> Can anyone comment on this topic ?
>>
>> I need large amount of address space for my application but I also
>> need to get as much I/O performance as it's possible, so I can switch
>> to 3.0/1.0 memory distribution if it will benefit here.

BG> You can't have it both ways with the x86 (speed vs. large userspace).
BG> Kernel 2.5 may help a bit here because changes were made to allow DMA
BG> from all memory (subject to card limitations), lessening the burden for
BG> direct-mapped memory. Otherwise you'll need to move to a 64-bit arch.




--
Best regards,
Peter mailto:[email protected]

2002-01-15 21:42:20

by Brian Gerst

[permalink] [raw]
Subject: Re: 3.5G user space speed

Peter Zaitsev wrote:
>
> Hello Brian,
>
> Wednesday, January 16, 2002, 12:05:39 AM, you wrote:
>
> BG> You can't have it both ways with the x86 (speed vs. large userspace).
> BG> Kernel 2.5 may help a bit here because changes were made to allow DMA
> BG> from all memory (subject to card limitations), lessening the burden for
> BG> direct-mapped memory. Otherwise you'll need to move to a 64-bit arch.
>
> Well. May be you can tell about the numbers a bit ? I can chose 3.0G
> for user instead of 3.5G for user with not really huge loss, but I'd
> like to know how much it will increase speed and in which cases, also
> about standard 2/2 mode.

I don't have quantifiable numbers available, but the speed issue will be
a result of the kernel running out of direct-mapped memory and having to
start swapping even though there is free memory in the system (in the
highmem zone). The best thing you can do is try both and find what
works best for you.

--

Brian Gerst

2002-01-15 21:47:50

by Peter Zaitsev

[permalink] [raw]
Subject: Re[2]: 3.5G user space speed

Hello Brian,

Wednesday, January 16, 2002, 12:41:56 AM, you wrote:
>>
>> Well. May be you can tell about the numbers a bit ? I can chose 3.0G
>> for user instead of 3.5G for user with not really huge loss, but I'd
>> like to know how much it will increase speed and in which cases, also
>> about standard 2/2 mode.

BG> I don't have quantifiable numbers available, but the speed issue will be
BG> a result of the kernel running out of direct-mapped memory and having to
BG> start swapping even though there is free memory in the system (in the
BG> highmem zone). The best thing you can do is try both and find what
BG> works best for you.

So how this should appear ? Like real swapping or like increased
system CPU usage because of moving pages to/from zone.


--
Best regards,
Peter mailto:[email protected]

2002-01-15 22:10:04

by H. Peter Anvin

[permalink] [raw]
Subject: Re: 3.5G user space speed

Followup to: <[email protected]>
By author: Peter Zaitsev <[email protected]>
In newsgroup: linux.dev.kernel
>
> 2.4.xaa Series as well as SuSE kernels have 3.5G userspace option,
> which seems to be quite useful, therefore I see it's not included
> is stock kernel for some reasons. Also I've heard this
> configuration may have some performance problems.
>

Be careful... 3.5G breaks the initrd protocol (assuming you have
enough RAM for it to matter anyway), unless you have a 2.03 boot
protocol capable (a) kernel and (b) bootloader. At this point I don't
know of any bootloader other than SYSLINUX 1.65 or later that is 2.03
compatible.

-hpa

--
<[email protected]> at work, <[email protected]> in private!
"Unix gives you enough rope to shoot yourself in the foot."
http://www.zytor.com/~hpa/puzzle.txt <[email protected]>