Here's something that's been bugging me for a while now...
I have several Linux servers that have been given enough RAM that they
rarely ever use any swap space. For example, here's the typical output
of uptime and free:
root@veronica$ uptime; free
03:55:33 up 225 days, 17:34, 0 users, load average: 0.09, 0.13, 0.19
total used free shared buffers cached
Mem: 2076136 1931288 144848 0 150220 964108
-/+ buffers/cache: 816960 1259176
Swap: 524280 156 524124
As you can see, it's been up for the better part of a year and is only
using 156k of swap. The swappiness is set to the default of 60, so
there's no reason why the server *shouldn't* be using swap -- it just
never does.
On some of my older servers with a little less RAM and the 2.4 kernel
I'll occasionally see the swap go up into the double digits, but even
this is relatively infrequent. On all the machines with a 2.6 kernel
and 2GB+ RAM, I've almost never seen more than 1-2MB swap used.
However, there is one exception: When a runaway process eats up all
the RAM and causes a server meltdown.
Fortunately runaway processes don't happen often, but when they do, a
large swap space seems to have been more of a hindrance than a help.
With out-of-control swapping going on, the other server processes
ground to a halt, and it is all but impossible to execute any commands
in an attempt to find and kill the runaway process. In fact, what has
ended up happening to me in these situations is that eventually --
after perhaps 10-20 minutes -- the swap space gets maxed out and
processes start getting killed automatically and/or the server
reboots.
The trouble is... these servers need to be available at all times, and
those 10-20 minutes where the server is completely unresponsive is
pretty bad for business.
So, what I'm getting at is this: If the only time swap is ever
actually used on these servers in quantities above a few megs is when
they're in the process of melting down, wouldn't it seem wiser to use
little or no swap and let them fail fast, rather than having a 1/4
hour of downtime?
In my research on this matter I've come across various comments by
people saying that getting rid of swap entirely will hurt performance
somehow, but on these servers I'm seeing very little swap used in the
first place, so I'm not seeing how that's really possible. But to
avoid that being an issue, let's suppose I give each server 32MB of
swap to cover the occasional usage. Is there any reason *not* to go
the route of minimizing swap in order to also minimize downtime?
The consensus these days seems to be that since hard drives are so big
now, go with a gig or more of swap even if you have plenty of RAM.
However, the way I'm seeing it is this: What's the point of having a
gig of swap if it only gets used during the worst possible time?
On Jun 5 2007 00:35, Mike Richards wrote:
>
> In my research on this matter I've come across various comments by
> people saying that getting rid of swap entirely will hurt performance
> somehow, but on these servers I'm seeing very little swap used in the
> first place, so I'm not seeing how that's really possible. But to
> avoid that being an issue, let's suppose I give each server 32MB of
> swap to cover the occasional usage. Is there any reason *not* to go
> the route of minimizing swap in order to also minimize downtime?
Generally yes. You want as much swap as the base daemons/programs take up in
RAM. Any [user] program should probably be allowed to max out your physical RAM
- the daemons go to swap in the meanwhile (mostly works with idle servers). A
program that requires more memory than physically present is anyway going to
swap in and out sometime later, degrading performance to a halt, as you
describe.
> The consensus these days seems to be that since hard drives are so big
> now, go with a gig or more of swap even if you have plenty of RAM.
> However, the way I'm seeing it is this: What's the point of having a
> gig of swap if it only gets used during the worst possible time?
> -
> To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
> the body of a message to [email protected]
> More majordomo info at http://vger.kernel.org/majordomo-info.html
> Please read the FAQ at http://www.tux.org/lkml/
>
Jan
--
On 5 Jun, 05:40, "Mike Richards" <[email protected]> wrote:
> Here's something that's been bugging me for a while now...
>
> I have several Linux servers that have been given enough RAM that they
> rarely ever use any swap space. For example, here's the typical output
> of uptime and free:
>
> root@veronica$ uptime; free
> 03:55:33 up 225 days, 17:34, 0 users, load average: 0.09, 0.13, 0.19
> total used free shared buffers cached
> Mem: 2076136 1931288 144848 0 150220 964108
> -/+ buffers/cache: 816960 1259176
> Swap: 524280 156 524124
>
> As you can see, it's been up for the better part of a year and is only
> using 156k of swap. The swappiness is set to the default of 60, so
> there's no reason why the server *shouldn't* be using swap -- it just
> never does.
>
> On some of my older servers with a little less RAM and the 2.4 kernel
> I'll occasionally see the swap go up into the double digits, but even
> this is relatively infrequent. On all the machines with a 2.6 kernel
> and 2GB+ RAM, I've almost never seen more than 1-2MB swap used.
[snip]
> The consensus these days seems to be that since hard drives are so big
> now, go with a gig or more of swap even if you have plenty of RAM.
> However, the way I'm seeing it is this: What's the point of having a
> gig of swap if it only gets used during the worst possible time?
Provided you won't be running an application which will require all
the memory, I've found 256MB swap on small (1GB mem) and larger (8GB)
servers to be adequate, but this clearly depends on what you're
running.
The idea would be to allow linux to swap out cold pages which are
(almost) never accessed while runaway processes get selected by the
OOM killer sooner. You may find 512MB to be more useful on large
untuned servers with lots of processes perhaps.
Daniel
--
Daniel J Blueman