Date: Tue, 28 Oct 2008 08:29:43 -0700 From: Scott Silva ssilva@sgvwater.com
Wrote:
Even though the recommended swap is 2 times system memory, I have never made a swap partition over 2 GB. Maybe I am also flirting with disaster, but haven't been bit yet in years.
I believe that the current recommendation is 2 x physical memory up to 2 GB and then 1 x physical memory thereafter.
See: http://www.redhat.com/docs/en-US/Red_Hat_Enterprise_Linux/5.2/html/Deploymen...
" Swap should equal 2x physical RAM for up to 2 GB of physical RAM, and then an additional 1x physical RAM for any amount above 2 GB, but never less than 32 MB.
So, if:
M = Amount of RAM in GB, and S = Amount of swap in GB, then
If M < 2 S = M *2 Else S = M + 2
Using this formula, a system with 2 GB of physical RAM would have 4 GB of swap, while one with 3 GB of physical RAM would have 5 GB of swap. Creating a large swap space partition can be especially helpful if you plan to upgrade your RAM at a later time.
For systems with really large amounts of RAM (more than 32 GB) you can likely get away with a smaller swap partition (around 1x, or less, of physical RAM). "
Regards,
On Wed, Oct 29, 2008 at 09:56:34AM -0400, James B. Byrne wrote:
I believe that the current recommendation is 2 x physical memory up to 2 GB and then 1 x physical memory thereafter.
See: http://www.redhat.com/docs/en-US/Red_Hat_Enterprise_Linux/5.2/html/Deploymen...
" Swap should equal 2x physical RAM for up to 2 GB of physical RAM, and then an additional 1x physical RAM for any amount above 2 GB, but never less than 32 MB.
That's a silly recommendation, and never been true for Linux. RedHat don't always know what they're talking about.
In older BSD systems (eg around SunOS 4 times or before) swap space was utilised oddly; all memory was allocated from swap, so you needed _at least_ <physmem> of swap just to use all your real memory! So if you added <physmem> of swap then your total virtual memory size was still only <physmem>. No help! So the rule of thumb came along that said "swap = 2*<physmem>" and that gave you a VM of 2*<physmem>. Linux never did this (and I don't think modern BSDs do, either) so adding <physmem> of swap will automatically give you twice <physmem> of VM.
So the old BSD 'swap=2*<physmem>' rule of thumb is now a linux 'swap=<physmem>' guideline instead.
But it's still just a rule of thumb, or a guideline. How much swap you need depends on your circumstances and load. If you have 4Gb of RAM and find the system is using most of it for cache then you may not even need any swap!
Stephen Harris wrote:
On Wed, Oct 29, 2008 at 09:56:34AM -0400, James B. Byrne wrote:
I believe that the current recommendation is 2 x physical memory up to 2 GB and then 1 x physical memory thereafter.
See: http://www.redhat.com/docs/en-US/Red_Hat_Enterprise_Linux/5.2/html/Deploymen...
" Swap should equal 2x physical RAM for up to 2 GB of physical RAM, and then an additional 1x physical RAM for any amount above 2 GB, but never less than 32 MB.
That's a silly recommendation, and never been true for Linux. RedHat don't always know what they're talking about.
In older BSD systems (eg around SunOS 4 times or before) swap space was utilised oddly; all memory was allocated from swap, so you needed _at least_ <physmem> of swap just to use all your real memory! So if you added <physmem> of swap then your total virtual memory size was still only <physmem>. No help! So the rule of thumb came along that said "swap = 2*<physmem>" and that gave you a VM of 2*<physmem>. Linux never did this (and I don't think modern BSDs do, either) so adding <physmem> of swap will automatically give you twice <physmem> of VM.
Ah, yes, not an older release of Solaris...SunOS4
A certain release of SunOS4 required swap be twice the amount of RAM or it would crash according to an article that I read but I cannot find right now. Here is a FAQ article related to SunOS that tells us how SunOS4 handled virtual memory (swap must be at least the same amount of RAM): http://www.bjnet.edu.cn/sun-admin/FAQ/F-comp-sys-sun/Q30-0.html
The source of this 'recommendation' that swap be twice that of RAM is affects more than just Linux
http://blogs.sun.com/jimlaurent/entry/solaris_faq_myths_and_facts
So please, stop spreading this swap should be twice the amount of RAM installed nonsense.
/me gets off soapbox.
Stephen Harris wrote:
In older BSD systems (eg around SunOS 4 times or before) swap space was utilised oddly; all memory was allocated from swap, so you needed _at least_ <physmem> of swap just to use all your real memory! So if you added <physmem> of swap then your total virtual memory size was still only <physmem>. No help! So the rule of thumb came along that said "swap = 2*<physmem>" and that gave you a VM of 2*<physmem>.
<old codger hat> this was also how early IBM System/370 DOS/VS worked... the swap space (which was contiguous cylinders of the disk, btw) was mapped 1:1 to virtual memory address space </old codger hat>
now, on late model Solaris, 'swap' is also used as tmpfs, which /tmp and /var/run utilizes by default, so you definitely want to allocate sufficient swap space for this at least, regardless of how large your physical memory is.
On Wed, Oct 29, 2008 at 07:07:03PM -0700, John R Pierce wrote:
Stephen Harris wrote:
In older BSD systems (eg around SunOS 4 times or before) swap space was utilised oddly; all memory was allocated from swap, so you needed _at least_ <physmem> of swap just to use all your real memory! So if you added <physmem> of swap then your total virtual memory size was still only <physmem>. No help! So the rule of thumb came along that said "swap = 2*<physmem>" and that gave you a VM of 2*<physmem>.
now, on late model Solaris, 'swap' is also used as tmpfs, which /tmp and
You also had this option on SunOS4 but it wasn't used by default and had a lot of limitations and didn't really work too well for /tmp (IIRC, didn't support FIFOs and some other file types). The Solaris version (first appeared in Solaris 2.1, so it's not _new_) is a LOT better :-)
/var/run utilizes by default, so you definitely want to allocate sufficient swap space for this at least, regardless of how large your physical memory is.
And, no matter what, ensure you limit your /tmp size with the size=####m option in vfstab, otherwise someone will clobber your machine!
To bring this back to Linux... Linux _also_ does have tmpfs which works similarly to Solaris. If you use that then also remember the size option :-)
On Wed, 2008-10-29 at 19:07 -0700, John R Pierce wrote:
Stephen Harris wrote:
In older BSD systems (eg around SunOS 4 times or before) swap space was utilised oddly; all memory was allocated from swap, so you needed _at least_ <physmem> of swap just to use all your real memory! So if you added <physmem> of swap then your total virtual memory size was still only <physmem>. No help! So the rule of thumb came along that said "swap = 2*<physmem>" and that gave you a VM of 2*<physmem>.
<old codger hat> this was also how early IBM System/370 DOS/VS worked... the swap space (which was contiguous cylinders of the disk, btw) was mapped 1:1 to virtual memory address space </old codger hat>
<snip>
<another old codger hat> I believe the differing "recommendations" is a result of the transitions over the life of *IX systems. When I first began running real UNIX (TM) on Dec and PC equipment (1978-1985), memory was slow, expensive and limited. Ditto for HD space. In order to run even fairly basic systems with any utility involving multi-tasking or multi-users, swap was need to page memory. Tunables were provided for high/low water marks and preferences for high/low priority processes. IIRC, initially there was no oom kill capability.
Recommended swap at that time was at least two times real memory and 4 times for "heavily loaded" systems. If you were doing more, bump that number appreciably. But it had severe costs when you maximum HD size might be 40MB, so you didn't want to over-allocate.
Over time, memory/HD attributes and "flavors" of *IX changed. Applications changed. Shared memory, libraries came into vogue. Relocation became possible. ELF appeared.
All this stuff changed the real needs for swap. But old habits die hard. So we still have conflicting recommendations floating around. Several on this list have been successfully configuring 0 swap for their usage. For other usage that may not be appropriate. </another old codger hat>
Summary: as it always has been, your application and load profile determine the need. Since HD is so "cheap" now, I have a couple gig of swap, but haven't seen it used since sometime in CentOS 5.0. Even when I login in as multiple users and fire up several X/Gnome desktops for each, no swap usage. There's probably been some swap used, but it wasn't seen by me.
Stephen Harris wrote:
" Swap should equal 2x physical RAM for up to 2 GB of physical RAM, and then an additional 1x physical RAM for any amount above 2 GB, but never less than 32 MB.
That's a silly recommendation, and never been true for Linux. RedHat don't always know what they're talking about.
I disagree with that. It has been true for linux for a long time, and is true even now. When was the last time you looked at the linux swap handling code ?
Recently, with large amounts of ram available on physical machines, eg > 2 GB, you need to start reconsidering a part of that strategy based around the fact that if your machine is really using 2 GB+ of swap, there might be other problems that need fixing and the absurd levels of latency introduced by that would mean those other issues really need looking at.
I'd still go with 2xRAM size for swap size upto a swap size of 2 GB, then make it the same as RAM size, till you get to 32 GB, at which point, half of real RAM size till 128 GB. Beyond that point, you really should be doing some benchmarks and monitoring to work out how much swap you need, and where it needs to be.
The other school of thought that seems to be doing the rounds is, 2xRAM till 4 GB, then just leave swap at 4GB irrespective of what your real machine RAM is, as long as its a multiple of 4.
-
Christopher Chan -
James B. Byrne -
John R Pierce -
Karanbir Singh -
Stephen Harris -
William L. Maltby