[Author Prev][Author Next][Thread Prev][Thread Next][Author Index][Thread Index]

Re: [school-discuss] Spreading LTSP over multiple servers?

On Mon, 2003-12-01 at 23:28, Tom Adelstein wrote:
> ipaterson@shaw.ca wrote:
> > Has anyone had any hands on experience using LTSP with multiple servers?
> > What with the power of modern desktop machines, it's becoming much more
> > plausible to have a pair of 1.2Ghz with oodles of RAM rather than a
> > Quad-Althon SpaceHeater, but I'm curious if there are issues that would
> > negate the cost benefit of doing this.
> > 
>  From my understanding, oodles of RAM may not help. For example, if your 
> CPU has 128 MB of L2 cache, I'm under the impression that it stops using 
> RAM at 128 MB. I may be completely inaccurate about that but my tests 
> indicate that you'll see diminishing returns.

Ok, let's correct some hardware misconceptions.

First, 128MB of L2 does not happen on consumer cpus, period.  128K is
only on the smallest of cpus today, and has very little to do with the
cacheable memory region of RAM.

Back in the early pentium days (430TX, 430FX chipsets, etc), the chipset
was limited to 64MB of cacheable memory.  This could cause a slow down,
due to the way Windows used memory (Windows 9x, not NT and not Linux),
as it would use uncached memory first.  Memory is ALWAYS faster than
hard disk, and thus if you need the ram, or can use the ram, it's best
to have the ram.

Today, our Terminal Server is running with 3GB of ram, here are the
stats on it right now:

JeryNelson:~# free -m
              total       used       free     shared    buffers
Mem:          3031       1613       1417          0        770
-/+ buffers/cache:        129       2901
Swap:          486          0        486

It was rebooted a few days ago, for NIS reasons, but it's normally only
at about 30-40 MB free, also I am not seeing anyone logged into

Modern chipsets/cpus (as the L2 has moved onto the CPU) have a cacheable
memory region that is as large as their memory address space.  Thus the
two cpus in this system, each with 256KB L2 and 128KB of L1 of cache,
are able to fully cache the entire 3GB of ram, as needed.


> I also understand that you can build a small GRID (Globus Tool Kit) 
> http://www.globus.org/ has provided significant power to Linux Terminal 
> Servers. But, that's hearsay. My experiments have been limited to 
> building the grid and running jobs from multiple work stations.

Harry McGregor, CEO, Co-Founder
Hmcgregor@osef.org, (520) 661-7875 (CELL)
Open Source Education Foundation, http://www.osef.org
A non-profit tax exempt charitable organization