[cod] Semi off topic: COD rentals

Sat Sep 25 15:51:07 EDT 2004

But we execute the same binary in the same path for each server.? Why
the symlink?

------------------------------------
Chris Adams
Fragzzhost

T (07005) 964 855
F (07005) 964 857
www.fragzzhost.com

-----Original Message-----
From: Mark J. DeFilippis [mailto:defilm at acm.org] 
Sent: 25 September 2004 20:33
To: cod at icculus.org
Subject: Re: [cod] Semi off topic: COD rentals

You are thinking application level.  This is Kernel level.
I spent lots of time in my career developing real time embedded coding
in Unix/Linux kernels.

Look at two linuxded files that are not sym linked. But use this:

ls -li

Note the left side has a value called an inode.  An inode is a unique
identifier for that file.  Notice that your two files have different
inodes?

When the loader goes to load code for your two servers using those
files, each is copied in to swap (so your swap space is n servers *
sizeof executable)
The same is true of code segments. 

If you sym link the two executables (Note they must be on the same
filesystem to do so. If you symlink two files on different file systems,
obviously Linux will have to copy the file to the new file system, and
it
will have a new inode number.  (This is because each filesystem has
it's own superblock, which maps Inodes to file blocks, and file names.

The filename if only for human consumption.  All Linux cares about is
that
inode number.

anyway... If you symlink the two server files, and now do a "ls -li",
notice the inode numbers are the same!

When you execute the server (with sticky bit on) it is copied to swap, 
then pages copied in to ram and executed.  When the second server
is executed, even though the file has a different path, it has the same
Inode.  The linux kernel looks up the inode, and notes this inode has
the sticky bit set, and in fact that it already exists in Swap. It
copies
the pointers to the code segments as I previously mentioned, and
begins execution.

You ask why?

Because 10 servers your way will load 10 copies in to swap, 10
copies in to memory, etc.  Using the method of symlinks, which
is really nothing more than a link in the super block.  The super block
maintains all information about which disk blocks belong tio which
files.  For your simlink the new file entry simply gets the same
inode id copied to the table.  Same inode, same code... With
shared libraries in memory,  for the 10 servers only 1 copy
exists in swap. Only 1 copy exists in memory.

I hope this is more clear.

Dr. D

At 02:18 AM 9/25/2004, you wrote:

Guess im a little confuused here, why would you symlink when cod (all
quake3 base games) support multiple users. using fs_basepath and
fs_homepath accomplishes the same thing as symlinking doesnt it?
Jase
NateDog wrote:

Woah......now there's a guy who knows his stuff!  Awesome tips man!  You
really explain things well.  Much appreciated.

--
NateDog

----- Original Message ----- From: Mark J. DeFilippis
To: cod at icculus.org
Sent: Friday, September 24, 2004 9:59 PM
Subject: RE: [cod] Semi off topic: COD rentals

I had written much about this a while back. I will repeat a bit of
it here for the sake of those who wish to do this. Want to know why
you should do this, why it works and a bit on how it works...

Linking the binaries allows the CPU to share the same Code segment
pages.
Servers
will be allocated their own data segments for both Heap and Stack
(Which grow towards one another)... One of the reasons Ryan was able
to so quickly find that original prob back in 1.1.)

If there is a write attempt to the code segment, that server/user is
given
their own copy. In the case of most of the shared libs in Linux, the
code
is reentrant, and hence these writes don't happen.

One other recommendation, I am not certain if I made...

You can reduce the spikes you get when a server is restarted by
setting the "Sticky" bit on the executable.  (Do a man on "mode"
command)
What this does is the first time the executable is loaded, the entire
executable is copies to SWAP space. Once copied to swap, executable
pages are copied in to ram to be executed.

The best way to keep a server at optimum is to never have to page.
However, under certain conditions, this does happen.  It the executable
is sticky, it remains in swap, and the page segment need only be
brought back in to memory from swap.

Also note, when a second and subsequent  user of the Sym Linked
executable
starts his/her server, the executable IS NOT copied in to swap again, it
uses
the one already in swap (hence the concept "sticky")... it sticks there.

Thus on new startup, A call is made to load the executable, however the
Kernel immediately updates the CS and ES code pointers to the shared
memory mbufs where the executable code exists, allocates a DS data
segment, and moves your process back to the scheduler for CPU as
your I/O is complete.

You skip the copy of the executable to SWAP.
You skip the copy of pages to Real RAM.
You execute off shared pages in memory already with your own set of
executable
  registers CS, ES.  Get your data segment, and your server starts up.

Not only do you save ram, but start impact on the other servers due to
I/O
DMA transfer setup, and context switching between system and user space,
but you spare the CPU spike as well.

Regards

Dr. D

At 05:08 PM 9/24/2004, you wrote:

I just had that question recently also.  I did some research on the
internet
and a lot of peeps are doing symlinks.  I tried it with MOH:AA and it
works
beautifully, not sure if that's the "right" way to do it but it's pretty
cool cause' you have one base install and symlinks in the other client
folders.

--
NateDog

-----Original Message-----
From: John Kennington [mailto:john.kennington at buzzcard.gatech.edu]
Sent: Friday, September 24, 2004 3:04 PM
To: cod at icculus.org
Subject: RE: [cod] Semi off topic: COD rentals

Depending on the number of cpus in the box, you can run 10 to 15 CoD
servers per
box.  So it is quite cost effective.

John Kennington

-----Original Message-----
From: Jafo [mailto:jafo at nowhere.ca]
Sent: Friday, September 24, 2004 3:58 PM
To: cod at icculus.org
Subject: [cod] Semi off topic: COD rentals

Hello,

If this isn't the forum for this question, please forgive me for asking
here.

There seems like a lot of people on this list that run "server rental"
operations. Just curious how people are doing that cost effectively?
Obviously one can't run each customer's game server on seperate
hardware.
Are people using some sort of "virtual linux" installs to run multiple
servers on one box with seperate IP addresses? If that is the case how
many servers would one dual 2.4 Xeon w/2gig RAM run?

Thanks,
Jafo

------------------------------------------------------------------------
----
S2----------------------------------------------------------------------
---------
Mark J. DeFilippis, Ph. D EE          defilm at acm.org
                                      defilm at ieee.org

-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://icculus.org/pipermail/cod/attachments/20040925/b595dd1b/attachment.htm>