Re: [gentoo-user] LTSP vs. Diskless Nodes

[Bob Sanders]

On Sat, 29 Oct 2005 22:55:24 +0200
Uwe Thiem <uwix@××××.na> wrote:


> > One of the big problems with Linux diskless is it really doesn't scale
> > well, it doesn't allow for clients to run multiple versions of the os,
>
> Why would you want to do that?
>

Ah!  Not everyone would.  But there are some who run realtime flight simulators
where the main Gfx system use 3 to 7 Gfx pipes to provide a 180 degree to 270 degree
view, puling in 1 TB or so of texture data during the sim.  This Gfx system has the problem
of needing proprietary drivers for both the SAN and the Gfx cards, so it's selection of
OS may be limited to a certain range, while the PCs that drive the instruments don't need
access to the SAN, and the 32P realtime server that runs the ssimulation and controls the
simulator reacts to the pilots  inputs, weather setup, etc., also doesn't need data from the
san, but has needs as to what's the best kernel to run for realtime simulation vs. realtime
Gfx.  And all this is booted off a 2P diskless server where the limits of what's seen by pilots
and perhaps maintainence crew is determined by whether they are running commerical,
military, or private aircraft that day. 

The diskless server could be any 64-bit capable 2P unit, wile the Gfx system would be a 
multi-pipe ia64 system, the 32P realtime system could be an ia64 or an x86-64 system and
the PCs would be standard x86, probably running WinXX and Linux.
 

> A typical LTSP server doesn't export /usr at all. There is no need for it. The
> client runs a kernel and an X server. If you want local devices to work, it
> also needs to run some other small daemons. All *applications* run on the
> server.
>

And this is a critical difference between LTSP - thin client serving, vs. a full diskless
client where the applications run on the client.  Sometimes one works fine (LTSP) for
the needs.  But other needs requires a different approach.

> My experiences with LTSP so far show: With a server like mentioned at the
> begin and fast ethernet, up to 20 clients are working well if you don't allow
> too graphics-intensive apps like movie players or that type of games. For
> more clients (up to 40), you need more ram on the server and a Gb connection
> between the server and the switch (clients can remain on 100Mb ethernet, of
> course).
>

A typical setup I run for testing has a 2P 600 MHz MIPS system with 512MB ram
as the server, serving 6 1P and 2P Gfx system, with the Gfx systems running 6 different
OpenGL apps, along with floating point work, local disk DMA and Xwindow DMA on 
all the clients.  One customer of ours runs 11 CAD systems off a single 2P diskless server.
  
> For small businesses, I prefer a different solution that involved solid state
> clients that boot from non-volatile ram. In that case, the client is
> completely independent of the server. All they talk to each other is X.
>

Yep, a great solution! 

> Cheers from the beginning southern African summer

 it's getting cold up here.  Shorter days and silly time changes.

Cheers,

Bob
-  
-- 
gentoo-user@g.o mailing list